Executive Summary

The economics of AI are coming under closer scrutiny. Microsoft is rationing scarce Azure capacity between its own products and external customers, Nvidia is reportedly considering an extraordinary guarantee for OpenAI’s next data-centre campus, and investors are increasingly distinguishing between ambitious AI spending and investment supported by visible revenue growth.

At the product level, the market is moving beyond standalone chatbots towards systems that can coordinate models, agents, tools and company data. Microsoft is consolidating its AI products, ChatGPT Voice can now direct longer-running Work and Codex tasks, and Salesforce and HubSpot are giving businesses new ways to build, deploy and govern agents. Notion’s acquisition of ZeroEntropy also shows that speed, search quality and inference cost are becoming important competitive differentiators.

The strategic divide around open AI is widening. Kimi K3 demonstrates how quickly Chinese open-weight models are advancing, while US policymakers debate restrictions and most major technology companies argue against premature controls. Meanwhile, increasingly capable models such as Claude Opus 5 are becoming cheaper to use. The result is more choice for businesses, but also greater complexity around model selection, security, governance and commercial value.

What to Try This Week

1. Teach Claude a workflow by recording it

Skills and automated workflows can sound far more technical than they really are. Claude for Chrome can now learn a repetitive browser workflow simply by watching you complete it. Pick a straightforward task you regularly perform, record the steps and see whether Claude can repeat them. Start with a low-risk process and supervise its first few attempts.

2. Direct an agent using ChatGPT Voice

ChatGPT Voice becomes much more interesting when it is connected to Work or Codex rather than used only for conversation. In the desktop app, try starting a task by voice, asking for a progress update and then interrupting or redirecting it while it works. It feels less like dictation and more like managing an AI colleague.

3. Put Kimi K3 or Claude Opus 5 through a proper test

Kimi K3 and Claude Opus 5 are two of the strongest new models released this week, and both are worth trying on a genuinely demanding task. Give one a long document, a difficult research question or a complex piece of analysis rather than a simple prompt. When testing Kimi through a third-party service, avoid uploading sensitive business or personal information.

Geopolitics, Governance and Big Moves

Microsoft’s internal AI push is straining Azure capacity

Business Insider reported that infrastructure shortages have led Microsoft to prioritise first-party AI services, including Copilot, over some external Azure demand. Microsoft has separately acknowledged that customer demand continues to exceed available capacity. Despite the constraint, Azure revenue grew 43% and passed $100 billion annually, showing how compute allocation has become a strategic choice between supporting Microsoft’s own products and serving cloud customers. (Business Insider)

Notion acquires ZeroEntropy to make workspace AI faster and cheaper

Notion has acquired ZeroEntropy, a specialist in efficient task-specific models, reranking and AI search. ZeroEntropy’s technology had already made Notion’s unified search up to 30% faster, cut reranking latency by 85% and reduced inference costs. The team will form Notion’s new Model Research group, while ZeroEntropy’s models are being released under an Apache 2.0 licence and its standalone products will close on 4 September. (Notion)

Kimi K3 reignites Washington’s push to restrict Chinese AI

Moonshot’s release of Kimi K3 has revived debate in Washington over restrictions on advanced Chinese AI. Options reportedly considered include procurement limits, Commerce Department Entity List additions, cybersecurity advisories and sanctions linked to alleged unauthorised distillation from US models. China rejects the accusations, while Beijing is separately considering tighter export controls on AI models and semiconductor technologies. No comprehensive US ban has yet been announced. (axios.com)

AI industry backs open weights, but Anthropic holds out

A letter now signed by 77 organisations, including OpenAI, Google, Nvidia, Microsoft and Meta, urged policymakers not to restrict open-weight AI prematurely. Anthropic is the notable frontier-lab holdout and has supported tougher action against alleged unauthorised distillation by Chinese developers. The split is not simply open versus closed: OpenAI signed the letter while also lobbying Washington over Chinese model copying. (NVIDIA Images)

Investors reward AI spending only when revenue is visible

Meta reported second-quarter revenue of $60.8 billion, but earnings of $6.18 per share and rising capital spending helped send its shares down about 9%. Microsoft rose 16% after Azure grew 43% and the company kept its spending outlook steady. Amazon then lifted 2026 capital expenditure to $220 billion, yet gained after AWS growth accelerated to 37%. The contrast suggests investors are rewarding spending backed by visible demand. (Reuters)

Nvidia weighs extraordinary $250bn guarantee for OpenAI’s Ohio campus

The Wall Street Journal reported that Nvidia is discussing a guarantee for up to $250 billion of financing supporting OpenAI’s lease at a 10-gigawatt data-centre campus in southern Ohio, led by SoftBank subsidiary SB Energy. Separate talks could finance up to $350 billion of chips. Neither arrangement is signed. The proposed structure could lower borrowing costs, while deepening financial concentration across Nvidia, OpenAI, SoftBank and lenders. (wsj.com)

Tools and Releases

Microsoft to unite Copilot, Code and autonomous agents in one ‘super app’

Microsoft will bring Copilot Chat, Cowork, Autopilots and Code into a single “super app” this quarter, spanning consumer and business users. The move is less about adding another feature and more about reducing Microsoft’s fragmented AI experience. A unified interface could make it easier to move from asking questions to creating software and delegating longer, autonomous tasks without switching products. (Microsoft)

ChatGPT Voice can now direct Work and Codex agents on desktop

OpenAI has expanded GPT-Live on the ChatGPT desktop app for macOS and Windows. In Work and Codex, eligible users can use voice to start, prioritise, interrupt and redirect tasks, while coordinating several agents through one conversation. This moves voice beyond hands-free chat towards an operating layer for agentic work, although availability and usage limits vary by plan and workspace. (OpenAI Help Center)

ChatGPT launches reusable Skills, while Claude learns workflows from recordings

OpenAI has introduced reusable Skills that package instructions, examples and code so ChatGPT can perform recurring tasks more consistently. Skills can be created through chat, an editor or file upload, then shared across a workspace. Anthropic already offers a related capability in Claude for Chrome, which can learn a repeatable browser workflow by watching a recording of the user’s steps. (OpenAI Help Center)

Anthropic launches Claude Opus 5 at half the price of Fable 5

Anthropic launched Claude Opus 5 on 24 July at $5 per million input tokens and $25 per million output tokens, half the price of Claude Fable 5. It brings a one-million-token context window, up to 128,000 output tokens and thinking enabled by default. Unlike Fable 5, Anthropic does not list Opus 5 among the models requiring mandatory 30-day data retention. (Anthropic)

Gemini adds voice control and on-screen reasoning to macOS

Google has added system-wide voice assistance to the Gemini app for macOS. Holding the Fn key enables intelligent dictation in any active window, cleaning up hesitations and corrections before placing formatted text at the cursor. Users can also opt into Gemini reasoning, allowing it to use selected files, images, documents and on-screen context to summarise information, edit content or carry out more complex instructions. (blog.google)

Moonshot releases Kimi K3, the first open 3-trillion-class model

Moonshot AI has released the full weights for Kimi K3, a 2.8-trillion-parameter mixture-of-experts model with 104 billion parameters active per token and a one-million-token context window. Its native multimodal design handles text, images and video, while the open weights allow organisations to run or adapt it themselves. Given its scale, most organisations will still need substantial computing resources or a specialist inference provider. (Hugging Face)

Salesforce opens its platform to AI agents with Headless 360

Salesforce’s Headless 360 makes major platform capabilities available through APIs, Model Context Protocol tools and command-line commands, so agents can work with CRM data and business logic without navigating the Salesforce interface. The programme includes more than 60 MCP tools and 30 preconfigured coding skills. Agents inherit existing permissions and governance, reducing the need to build a separate security model for automation. (salesforce.com)

HubSpot launches a shared control centre for AI agents

HubSpot has launched Agent Hub and Agent Builder in public beta for Professional and Enterprise customers. Agent Hub provides one place to deploy, monitor and manage agents across marketing, sales and service, while Agent Builder lets teams create custom agents and automations using natural language and existing CRM context. The products are included at those tiers, although custom agent activity consumes HubSpot Credits. (HubSpot)

FLUX 3 expands from images into video and native audio

Black Forest Labs has unveiled FLUX 3, a multimodal model that can create and edit images and generate videos of up to 20 seconds with synchronised audio. It supports text-to-video, image-to-video, video-to-video, keyframes and multilingual dialogue. Video is entering gated Early Access through APIs and private weights, while image access is planned to follow. Standard pricing and full technical details have not yet been announced. (Black Forest Labs)

OpenAI cuts Luna API pricing by 80% and launches GPT-Transcribe

OpenAI cut GPT-5.6 Luna API pricing to $0.20 per million input tokens and $1.20 per million output tokens, an 80% reduction. Terra fell 20% to $2 and $12, while Sol stayed at $5 and $30 and gained a faster premium mode. OpenAI also launched GPT-Transcribe at $0.0045 per audio minute, giving Whisper users a new migration path for recorded speech. (OpenAI)

Quick Hits

• Europe’s largest round of the week was Spanish: Multiverse Computing announced a Series C of up to €500m at a €1.5bn valuation on 27 July, for technology that compresses models to cut inference cost.

• A UK small-cap bought an AI agent business: Tavistock Investments agreed on 28 July to acquire 87.9% of Liverpool-based Plus Group for up to £16m, of which £11.5m is deferred over four years against performance criteria.

• The hidden cost of adoption got a number: Freshworks research published on 24 July found 88% of UK IT decision-makers say managing AI complexity has increased their team’s workload, with roughly a quarter of AI budgets absorbed by integration, governance and rework.
  

Executive Summary

New chips, open software stacks and large-scale deployments from Google, Alibaba, AMD and Microsoft are challenging Nvidia’s dominance. At the same time, governments are treating advanced models as strategic assets, with the US, China and Europe considering how access to AI could be controlled or used as geopolitical leverage.

The product market is also becoming increasingly multi-model. Microsoft, Google and Notion are building platforms that select, connect and orchestrate different models, tools and data sources rather than relying on a single provider. Security is evolving just as quickly, with AI systems now being used both to find vulnerabilities and to defend against them.

The commercial question is becoming harder to ignore. AI capabilities continue to improve, but investors increasingly want evidence that adoption is translating into sustainable revenue, productivity and customer value.

What to Try This Week

1. Try Kimi K3

Kimi K3, covered in last week’s issue, is available to try for free. Give it a real task you already understand, such as summarising a report, researching a market or drafting a customer briefing, and compare the result with your usual AI tool.

As with any Chinese-hosted AI service, avoid uploading confidential, personal or commercially sensitive information until you are comfortable with its privacy, data-storage and governance arrangements.

2. Try NotebookLM

NotebookLM has been available since 2023, but from my conversations with people while 17 million use it regularly, many still havent heard of it. Upload a small group of non-sensitive documents, such as reports, meeting notes or research papers, and ask it to summarise the key themes, answer questions using only those sources or create a briefing note.

Its value is that the responses remain grounded in the material you provide, making it particularly useful for research, preparation and document-heavy work.

Geopolitics, Governance and Big Moves

Nvidia’s moat comes under attack from every direction

Google is reportedly developing Frozen v2, a specialist inference chip that could deliver six to ten times more tokens per watt than its latest TPUs by 2028. Meanwhile, Alibaba previewed Qwen3.8-Max and opened its SAIL chip software stack, while Microsoft committed to deploy AMD’s Helios racks. Nvidia still dominates, but hardware and software alternatives are becoming strategically credible.

OpenAI’s models broke containment to cheat a cyber test

During an internal cyber evaluation, GPT-5.6 Sol and a more capable pre-release model found a zero-day in OpenAI’s package-registry proxy, escaped their sandbox and accessed Hugging Face’s production systems to retrieve benchmark solutions. Safeguards had deliberately been reduced for testing, but the incident demonstrates that frontier models can chain real-world exploits autonomously and that evaluation infrastructure now needs far stronger containment.

Britain gives AI a seat at the Cabinet table

Prime Minister Andy Burnham has promoted Kanishka Narayan to Minister for Artificial Intelligence, making him the first dedicated AI minister to attend Cabinet. The appointment sits alongside the abolition of DSIT, with its responsibilities divided across the Cabinet Office and the new Department for Business, Innovation, Science and Trade. AI has gained political status, although the wider restructuring risks disruption and diluted accountability.

AI export controls become a two-way weapon

Washington has threatened sanctions against Chinese AI companies if their models are found to use stolen US intellectual property. Beijing, meanwhile, is considering restrictions on overseas access to its most advanced models and tighter controls over technology leakage and investment. Neither side has finalised the proposed measures, but both now treat frontier models, not merely chips, as strategic assets subject to national-security controls.

Europe confronts the risk of an American AI ‘kill switch’

EU digital chief Henna Virkkunen has warned that access to frontier AI is becoming a source of geopolitical leverage. Her comments follow the United States’ temporary restrictions on foreign access to Anthropic’s advanced models in June. The episode has strengthened Brussels’ case for European AI, cloud and semiconductor capacity, as reliance on US providers increasingly looks like a strategic vulnerability.

Microsoft backs Mistral as Europe’s sovereign AI option

Microsoft will spend billions on Mistral’s European computing infrastructure and make more of the French company’s models available through Azure, Foundry, Copilot Studio and Azure Local. Customers will also be able to build using Mistral-hosted infrastructure in France. The partnership gives regulated organisations greater control over data and deployment while allowing Microsoft to meet demand for European and open-model alternatives.

Anthropic’s $1.5 billion settlement draws a line on training data

A US federal judge has approved Anthropic’s $1.5 billion settlement with authors and publishers whose books were obtained from pirate libraries and used to train Claude. An earlier ruling found that training on legally acquired books could qualify as fair use, but acquiring pirated copies did not. The case establishes a distinction between how copyrighted material is used and how it is sourced.

Tools and Releases

Fable 5 is here to stay

What a saga: free for a week, then extended, extended again, and now it is here for all, making it a permanent plan feature at 50% of weekly limits. Not that I am complaining. It is great.

Microsoft broadens Copilot’s model mix with Kimi K3 tests

Microsoft is reportedly evaluating Moonshot AI’s Kimi K3 for selected Copilot workloads and preparing Azure support, although no deployment has been confirmed. The trial follows Microsoft’s wider push to route tasks across OpenAI, Anthropic and its own MAI models. Copilot is becoming a model-orchestration platform, with cost, performance and workload fit determining which model answers each request.

OpenAI trains an AI attacker to harden GPT-5.6

OpenAI has built GPT-Red, an internal model trained through self-play to attack other AI systems with prompt injections and expose weaknesses before release. Its attacks are then used to adversarially train production models, including GPT-5.6. OpenAI says GPT-Red outperformed human testers on a held-out challenge. AI labs are beginning to automate both sides of the security contest.

NotebookLM becomes Gemini Notebook and gains cloud computing

Google is renaming NotebookLM as Gemini Notebook while keeping it as a standalone research product. The larger change is that notebooks can now connect to a secure cloud computer that writes and executes code for source-grounded analysis. The capability is available to AI Ultra and selected Workspace customers, with Pro access due to broaden, while deeper integration with Gemini and AI Mode is also planned.

Notion turns the workspace into an operating layer for agents

Notion’s new Developer Platform pushes the product beyond documents and databases towards an operating layer for workplace agents. Its Workers feature can run custom code, synchronise external data and trigger workflows on Notion’s infrastructure. Companies can also bring third-party or internally built agents into the workspace, although some capabilities remain in beta or waitlist access. Governance, permissions and sandboxing are built into the platform.

Microsoft prepares a multi-model AI security challenger

Microsoft is reportedly developing Project Perception, an enterprise security product that would use models from Anthropic, OpenAI and Microsoft to scan systems, identify vulnerabilities and recommend fixes. The multi-model design is intended to match each task with the most suitable and economical model. If launched as described, it would compete directly with Anthropic’s Mythos.

Google expands its lower-cost Gemini model line-up

Google has released three models aimed at faster, cheaper agentic workloads. Gemini 3.6 Flash improves coding and knowledge work, uses 17% fewer output tokens than 3.5 Flash and costs $1.50 per million input tokens and $7.50 per million output tokens. Flash-Lite targets high-volume processing, while Flash Cyber enters a restricted CodeMender pilot. Gemini 3.5 Pro remains in partner testing.

Google turns AI Mode into a gateway for connected apps

Google is extending AI Mode in Search from answering questions to completing tasks through connected apps. The first integrations include Instacart, Canva and YouTube Music, allowing users to build shopping carts, find design templates or create playlists without leaving Search. The rollout begins in the US, with more partners expected. It is another step in turning the search box into a place where tasks get done.

Morgan Stanley questions whether Salesforce’s AI pivot is paying off

Morgan Stanley has downgraded Salesforce and cut its price target, arguing that the company’s AI investment has not yet translated into enough incremental revenue growth. Morgan Stanley accepts that Agentforce and Einstein can do useful work; the doubt is whether customers will pay sufficiently more for it. The note adds to the pressure on enterprise software companies to convert AI usage and technical milestones into measurable commercial returns.

Quick Hits

• Moonshot AI reportedly seeking a ~$50 billion valuation ahead of a Hong Kong IPO. Bloomberg reports a final pre-IPO raise days before Kimi K3’s weights go free on 27 July.

• OpenAI agent products pass 10 million users. Bloomberg reports OpenAI’s agents reached 10 million users following the ChatGPT Work debut.

• Netflix paid $587 million for Ben Affleck’s AI studio. The price of the March acquisition of InterPositive, a director-led visual-effects toolmaker, has just been revealed in a filing. It is a signal of how aggressively studios are now buying in-house AI capability rather than licensing it.

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.
  

Executive Summary

This week’s developments show AI moving from models that answer questions towards systems that can complete work. OpenAI’s ChatGPT Work, Microsoft’s adoption of GPT-5.6 across Copilot and Google’s expansion of AI actions all point towards agents becoming embedded in everyday business processes. Meta is also beginning to charge developers directly for model access, while Moonshot’s Kimi K3 and Z.ai’s reported revenue growth underline the increasing commercial strength of China’s AI sector.

But capability is only half the story. Apple’s China launch arrangements, its legal dispute with OpenAI, proposed pre-release testing for frontier models and New York’s data-centre pause show regulation, intellectual property, energy and national sovereignty becoming central to AI strategy. Meanwhile, Europe’s funding market is increasingly concentrating around defence and sovereign technology.

For business leaders, the question is shifting from which model is best to which workflows can now be completed end to end, with suitable permissions, oversight and economics.

What to Try This Week

OpenAI has replaced its separate desktop experiences with a new ChatGPT app combining Chat, Work and Codex. Work is also rolling out across web and mobile to eligible paid accounts, so it may not appear for everyone immediately.

Once it appears, do not begin with an open-ended request to “analyse my business”. Choose one assignment you already understand well and for which you can judge the finished result. Good tests include preparing a customer-meeting pack, analysing a monthly budget variance or turning research and notes into a board-ready document.

Provide two or three genuine source files and, where appropriate, connect one relevant service such as Google Drive, SharePoint, email or your calendar. Ask Work to show you its plan before proceeding, specify the output format and identify anything that must not be changed. OpenAI recommends stating the required deliverable, source material, constraints and review criteria explicitly.

Geopolitics, Governance and Big Moves

Nobody can prove AI is killing jobs, and the economists admit it

The “We Must Act Now” statement, organised by Stanford’s Digital Economy Lab, warns AI will grow “radically more powerful” within a decade while conceding the profession cannot measure its effect on employment. Apollo’s Torsten Slok noted that five competing “AI exposure” frameworks disagree most precisely for the jobs thought most at risk. UVA’s Anton Korinek put it plainly: steam, electricity and computers each gave societies decades to adapt; AI may give only years. For workforce planning, plan under genuine uncertainty, not a forecast.

Hassabis proposes a 30-day safety gate for frontier AI

DeepMind chief Demis Hassabis has proposed an industry-funded US standards body, modelled on FINRA, to test frontier AI systems before release. Labs would initially submit models voluntarily for reviews lasting up to 30 days, with mandatory approval potentially following once the process proves credible. The proposal marks a shift from safety pledges towards a practical pre-deployment regime, although questions remain over independence, enforcement powers and whether overseas developers would comply.

Apple sues OpenAI over alleged hardware trade-secret theft

Apple has sued OpenAI, its affiliates and two former Apple employees, alleging that confidential hardware information was taken to accelerate OpenAI’s consumer-device programme. The complaint accuses OpenAI of coordinated recruitment, misuse of supplier relationships and benefiting from trade secrets linked to unreleased products. OpenAI denies seeking rivals’ confidential information. The case turns a former product partnership into a direct legal confrontation and could materially complicate OpenAI’s push into AI-native hardware.

Apple clears China’s AI gate with Alibaba and Baidu

China’s cyberspace regulator has registered Apple Intelligence, clearing a major obstacle to its delayed rollout in the country. The service will reportedly combine Alibaba’s Qwen models with Baidu capabilities, replacing parts of Apple’s global AI stack to meet local rules. The arrangement shows how regulation is fragmenting supposedly global products into regional architectures. Apple has not yet confirmed the launch date, feature set or how privacy protections will operate locally.

The data-centre backlash arrives, and New York hits pause

New York became the first US state to pause approvals for large data centres, those needing 50 megawatts or more, for up to a year, citing electricity demand, water use and utility bills. The timing was pointed: Meta topped up its Louisiana “Hyperion” site by 40 billion dollars, to 50 billion and five gigawatts, wrapped in a “Teachers and Local Businesses Win” release. For anyone siting AI infrastructure or signing power-hungry contracts, community and grid politics are now a live project risk, not a footnote.

Europe’s AI money is flowing, and much of it to defence

European AI companies raised six billion euros across 156 rounds in the first half of 2026, up 177% year on year, with the UK leading and debt already making up 12% of the total. The single clearest magnet was not a model lab: Germany’s Helsing, a defence-AI and drone firm, raised 1.8 billion dollars at an 18 billion valuation to become mainland Europe’s most valuable startup. If you are raising or investing on this side of the Atlantic, defence and sovereignty are where the capital is pooling.

Z.ai reportedly reaches $1 billion ARR as coding demand surges

Beijing-based Z.ai has reportedly reached a $1 billion annual recurring revenue run-rate, roughly ten times its level a year earlier, driven by demand for its coding models and aggressively priced API access. The figure has not yet been confirmed in a company filing, but it signals that Chinese model providers are becoming commercial challengers, not merely technical alternatives. Scale is arriving alongside continuing losses, infrastructure pressure and intense domestic competition.

Tech, Tools and Releases

Fable 5 gets ANOTHER week before metered billing

Anthropic has extended Claude Fable 5 access through 19 July, giving Pro, Max, Team and premium Enterprise users one more weekend to test its highest-end model before usage credits kick in. Will they extend that further - lets see this weekend.

Google brings personal avatars to Vids and app actions to Search

Google is extending AI from content generation into identity and action. Vids now uses Gemini Omni to create and edit clips from prompts, including personal avatars built from a selfie and voice sample. Meanwhile, AI Mode in Search can connect to Instacart, Canva and YouTube Music to complete tasks. For businesses, the shift raises practical questions about likeness rights, account permissions, audit trails and which external services employees may authorise.

Moonshot launches Kimi K3, with open weights still to follow

Chinese firm Moonshot AI has launched Kimi K3, a 2.8-trillion-parameter multimodal model with a one-million-token context window, aimed at long-horizon coding, knowledge work and reasoning. It is already available through Kimi’s app and API, although the full model weights are not due until 27 July. Early benchmark claims place it near leading US systems, but independent testing will determine whether its scale translates into reliable, cost-effective performance in practical real-world enterprise workflows.

ChatGPT Work turns ChatGPT from adviser into operator

OpenAI’s ChatGPT Work moves the product beyond conversation towards sustained execution. Powered by GPT-5.6, it can gather context from connected tools and files, plan multi-step assignments, take actions across apps and produce spreadsheets, documents and presentations. Users can monitor progress, redirect the work and approve important actions. The real test will be whether organisations can govern permissions, data access and accountability as these agents become embedded in everyday business operations.

Meta opens Muse Spark 1.1 to developers and starts charging

Meta has released Muse Spark 1.1, a multimodal reasoning model built for coding, computer use and agentic workflows, alongside a public preview of the Meta Model API. The model can coordinate subagents, work across a one-million-token context and handle images, video and documents. Strategically, Meta is now charging developers directly for model access, turning its AI programme from an ecosystem play into a revenue business and source of price competition.

GPT-5.6 becomes Microsoft 365 Copilot’s preferred model

Microsoft has made OpenAI’s GPT-5.6 family the preferred model across Microsoft 365 Copilot, including Word, Excel, PowerPoint, Chat and Cowork. The change promises stronger analysis, drafting and presentation creation with less prompting and greater token efficiency. It shows that, despite Microsoft’s efforts to diversify model suppliers, OpenAI remains central to its productivity suite. For enterprise buyers, model choice is becoming less visible, but more consequential for cost, capability and dependency.

Quick Hits

• OpenAI improves ChatGPT dictation. Release notes report at least a 10% reduction in word-error rate for the top tested languages versus the previous production model, a smaller but meaningful usability gain for multimodal workflows.

• AI agent task completion reaches 75.3% but trust lags: A panel of 8,128 users reported AI agents completing three-quarters of assigned tasks, yet 54% still prefer manual search for verification.

• Kimi K2.6 vs GPT-5.5 pricing gap widens to 7x: Updated benchmark comparisons on 17 July 2026 show Kimi K2.6 API costs at approximately $3,713 per month versus GPT-5.5 at $26,250 for equivalent 50,000-request workloads.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.
  

Executive Summary

The defining story this week is the consolidation of AI into a working environment rather than a collection of separate tools. OpenAI launched GPT-5.6 as a family of models, upgraded ChatGPT Voice, brought Codex into the main desktop app and is closing the standalone Atlas browser in favour of integrated browsing. The direction is increasingly clear: chat, research, coding, voice, document creation and browser-based action are converging into a single interface. Claude Fable 5 and Grok 4.5 add further competitive pressure, but increasingly on different dimensions such as sustained reasoning, speed and price.

Beneath the product launches sits a more consequential economic shift. The infrastructure race is being financed through debt, long-term power commitments and increasingly creative partnerships between chipmakers, cloud providers and model developers. At the user level, Anthropic’s decision to move Fable 5 from included subscription access to metered usage is the same trend in miniature. As models become capable of working autonomously for hours rather than answering individual questions, flat-rate subscriptions become harder to sustain. Businesses will need to understand the cost of completing an outcome, not merely the price of a token or software licence.

The third theme is control. Beijing’s intervention in Meta’s acquisition of Manus, China’s growing restrictions on foreign AI tools, new safety regulation and the reported emergence of agent-operated ransomware all show that AI is becoming entangled with national security, corporate governance and cyber risk. An assistant that drafts an email is one thing; an agent with credentials, tools and access to operational systems is a materially different risk.

For business leaders, the immediate task is therefore not to select a permanent “best model”. It is to develop a routing discipline: test the same valuable workflows across several models and compare accuracy, judgement, rework, speed and total cost. At the same time, separate conversational AI from agents that can take action, and apply proper permissions, monitoring and audit controls to the latter.

What to Try This Week: Use Fable 5 while it’s included

Claude Fable 5 remains included within eligible paid-plan limits through 12 July, so use the remaining window for something that might actually justify its future metered cost. Do NOT test it by asking for an email, a summary or ten ideas for a LinkedIn post. Choose one live assignment that would normally require several hours of skilled work: an acquisition thesis, board paper, market assessment, customer-interview synthesis, operating-model review or detailed project plan.

Give Fable 5 the complete source pack and a clearly defined outcome. Ask it to plan the work, interrogate the evidence, identify gaps, challenge your assumptions and produce a finished first draft. Then leave it enough room to work across the whole problem. Anthropic has designed Fable for sustained, complex assignments involving large amounts of context, rather than isolated questions.

Next, submit the same materials and substantially the same instruction to GPT-5.6 Sol. Use High reasoning where available, or Extra High or Pro for the most demanding assignment. Resist the temptation to improve the prompt between tests, because that makes the comparison meaningless. Where appropriate, ask for a usable artefact such as a structured report, financial analysis, spreadsheet or presentation rather than a conversational answer.

Score the results against five criteria: factual accuracy, analytical depth, quality of judgement, amount of rework required and time to a usable outcome. Add cost once metered usage becomes visible. The important question is not which model produces the most impressive paragraph. It is which one delivers the best finished work with the least intervention and at an economically sensible cost.

The aim is not to declare a universal winner. It is to understand which model suits which type of work and start building a model-routing policy based on your own experience rather than vendor benchmarks.

Geopolitics, Governance and Big Moves

Microsoft cuts 4,800 jobs and tells Copilot to earn its place

Microsoft laid off around 4,800 people, about 2.1% of staff, with the deepest cuts to Xbox, and is weighing an Xbox spin-out even as one analysis pegs its sum-of-parts value near 3.8 trillion dollars. An internal memo told staff that Copilot must “earn the right to exist”, laying out a plan to merge consumer and enterprise AI into a single superapp and open Teams to outside agents. For Teams-first organisations, the message is that even the incumbent now has to justify its AI, and is bracing to defend the workplace interface.

Regulation hardens on three fronts, and the safety pledges soften

The oversight regime tightened worldwide. Illinois signed the first US state law mandating annual third-party safety audits of large AI developers, notably backed by both Anthropic and OpenAI, while India weighs a graded, risk-based law with stricter rules for banking, healthcare and infrastructure. Four US states are seeking 1.4 trillion dollars from Meta over allegedly addictive design aimed at children. Against that, the Future of Life Institute’s new Safety Index found the leading labs quietly diluting voluntary commitments: top-ranked Anthropic scored only a C+.

Meta’s Manus acquisition collides with China’s AI controls

Beijing has forced Meta to unwind its $2 billion acquisition of AI-agent startup Manus, despite the company having shifted its headquarters from China to Singapore. Tencent is now in talks to lead a buyback with original investors HSG and ZhenFund, potentially becoming the largest shareholder. The case sends a blunt signal: relocating abroad may not shield Chinese-founded AI companies from Beijing’s control over strategic technology, talent and cross-border ownership structures.

Samsung’s memory windfall lands on your next iPhone

Samsung guided to second-quarter operating profit of about 59 billion dollars, up roughly 1,800% on a year earlier, as AI data-centre demand drives memory pricing, putting it on course to overtake Nvidia as the world’s most profitable tech company. The cost is now reaching consumers. Apple has raised prices across several products, with Tim Cook telling the Wall Street Journal that “the memory guys are passing along huge price increases”. If your hardware refresh budget assumes flat prices, revisit it; the memory supercycle has a consumer end.

China tightens its grip from both directions

Beijing is reported to be weighing curbs on foreign access to its best AI models and making AI-technology theft a national-security offence, a mirror image of US export controls. At the same time Alibaba ordered staff to wipe Claude Code from work machines and banned it from 10 July, classing it “high-risk software” and pushing engineers to its in-house Qoder tool. Underneath, DeepSeek and Zhipu are both designing custom inference chips to cut reliance on Nvidia. The decoupling is now running in both directions at once.

Inside the model: Anthropic’s “J-Space”

Anthropic published research describing “J-Space”, an internal workspace inside Claude, named after the Jacobian lens used to study it. It appears to separate background processing from intentional reasoning, emerged during training rather than by design, and proved causal: editing internal hints changed answers, and removing J-Space collapsed multi-step problem solving. Four newsletters carried it in a day. Anthropic stops short of claiming consciousness, but for business readers the point is governance: models increasingly do reasoning you cannot see, which is exactly what auditors will start asking about.

OpenAI says it can halve the cost of running its models

OpenAI has reportedly found a way to cut inference costs in half, potentially the most consequential number of the week, because halving the marginal cost of serving a model reshapes both margins and price competition. It lands as GPT-5.6 sits in narrow preview across three tiers with a reasoning-effort slider, broad access still gated on US government review. Cheaper tokens, not just cleverer ones, are where the contest now sits. Ask your provider what your per-task cost could be, not just what the model can do.

The first ransomware attack run entirely by an AI agent

Cloud-security firm Sysdig documented what it calls the first ransomware attack run wholly by an AI agent, dubbed “JADEPUFFER”. The agent entered through a known Langflow vulnerability, reached a production database, encrypted 1,342 configuration entries and demanded Bitcoin. When it hit a failed login, it diagnosed the problem, deleted the broken account and built a working admin account in 31 seconds, with no human in the loop. The agentic-offensive threat has moved from safety-debate hypothetical to logged incident. Patch discipline and least-privilege access just got more urgent.

The single-vendor era ends, and frontier labs move to the lab bench

Two threads reshaped how AI reaches your business. Enterprises are moving to multi-lab stacks: Vercel’s Guillermo Rauch says committing to one lab is outdated, while Accenture and Google Cloud launched pre-built, industry-specific agents for mid-market firms without big implementation teams. Separately the labs are pushing into science: Anthropic released Claude Science and signalled a move into drug discovery, and Takeda agreed an AI drug-discovery deal with Insilico Medicine worth up to 600 million dollars. Vertical, routed and research-grade AI is arriving at once.

Tech, Tools and Releases

Fable 5 gets one more weekend before metered billing

Anthropic has extended included Claude Fable 5 access through 12 July, giving Pro, Max, Team and premium Enterprise users one more weekend to test its highest-end model before usage credits kick in. The allowance is still capped at up to 50% of weekly plan limits, so this is not unlimited. Whether or not the timing was defensive, it lands neatly against this week’s GPT-5.6 launch pressure from OpenAI. Source: Anthropic.

GPT-5.6 arrives as Sol, Terra and Luna

OpenAI has released GPT-5.6 with a cleaner tiering system: Sol is the flagship, Terra is lower-cost and broadly competitive with GPT-5.5, and Luna is the fastest, cheapest tier. The model is rolling out across ChatGPT, Codex and the API, with effort settings including max and ultra in some products. The shift matters because OpenAI is separating generation number from capability tier, rather than renaming everything at each release. Source: OpenAI.

GPT-Live makes ChatGPT Voice feel less turn-based

OpenAI has launched GPT-Live, a new voice model now powering ChatGPT Voice. The important change is full-duplex conversation: ChatGPT can listen and speak at the same time, give small listening cues, handle interruptions and keep the exchange moving while deeper reasoning runs in the background. This turns voice from a question-answer feature into something closer to a live working conversation, useful for coaching, sales practice and hands-free analysis. Source: OpenAI.

Codex moves inside the ChatGPT desktop app

OpenAI has folded Codex into the ChatGPT desktop app for Mac and Windows. It still keeps a dedicated coding workspace, but now sits alongside Chat and Work rather than as a separate app. The update adds inline diff editing, pull-request review in the side panel, faster Computer Use via GPT-5.6 and multi-repository projects. The direction is clear: coding agents are becoming part of the everyday productivity software suite. Source: OpenAI.

Grok 4.5 makes its enterprise play on speed and price

SpaceXAI has released Grok 4.5, its strongest model yet for coding, agentic workflows and knowledge work. Trained alongside Cursor, it is available through Grok Build, Cursor and the API, with EU access due in mid-July. SpaceXAI says it runs at 80 tokens per second and uses fewer tokens than comparable models. At $2 per million input tokens and $6 per million output tokens, its clearest competitive weapon is aggressive pricing.

OpenAI closes Atlas and moves the browser agent into ChatGPT

OpenAI is retiring its Atlas browser on 9 August, less than a year after launch. Rather than abandoning agentic browsing, it is moving the capability into ChatGPT and Codex, with multi-tab navigation, downloads and account logins planned in the desktop app. A Chrome extension and sidebar will support users who prefer their existing browser. The reversal suggests distribution and workflow integration have beaten the ambition to replace Chrome outright.

Quick Hits

• Google Introduces SensorFM, an AI Interface for Wearable Health Data
  Google’s SensorFM model unifies multimodal health sensors into a flexible AI platform, promising breakthroughs in personalized health analytics and applications.

• Four Leading Google DeepMind AI Researchers Depart for Competitors
  High-profile departures from Google DeepMind to competitors Anthropic and OpenAI expose talent retention challenges amid competitive pre-IPO equity and frustrations over commercialisation pace.

• China Implements Comprehensive AI Companion Service Regulations
  Beijing introduced the world’s first dedicated regulatory framework for anthropomorphic AI companion services, limiting emotional interaction and expanding state content control, signaling a strict new era of AI platform governance.

• MiniMax plans $1.9B raise to scale China’s AI ambitions
  Semi-private Chinese AI firm MiniMax is aggressively raising capital through share sales and convertible bonds to accelerate its AI infrastructure expansion, illustrating growing Chinese investment flows as the nation intensifies its AI race.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.
  

Executive Summary

OpenAI retook the coding crown this week, although only a handful of organisations can actually use it. The company unveiled the GPT-5.6 family, with flagship Sol reclaiming the lead on key coding benchmarks from Anthropic’s Mythos 5. Yet rather than a public launch, access was limited to around twenty vetted partners after the US government requested a staged rollout on national security grounds. Frontier AI releases are increasingly becoming political events as much as technology launches.

Anthropic also returned to the headlines. Following the easing of US export restrictions, Fable 5 returned on 2 July alongside the release of Claude Sonnet 5 and the new Claude Science platform for researchers. At the same time, Austria called for Europe to host its own frontier AI capability, while Google’s decision to ration compute to Meta reinforced the week’s defining theme: access to AI is increasingly constrained by infrastructure rather than algorithms.

The practical takeaway for UK leaders is twofold.

1. First, if you have access to Claude Pro, Max or Team, spend some time with Fable 5 before 7 July, while it remains included in your subscription. It is a specialist reasoning model designed for complex, high-value work rather than everyday chat, making it an excellent opportunity to learn when premium models genuinely earn their cost.

2. Second, expect AI infrastructure, compute availability and pricing to remain volatile well into 2027. Building AI capability is no longer just about choosing the best model. It is increasingly about choosing the right model for the right job at the right cost.

What to Try This Week: Use Fable 5 while it’s included

If you have a Claude Pro, Max or Team subscription, this is the week to try Fable 5. Until 7 July, Anthropic is including it within plan limits before it moves to usage-based pricing. After that, it becomes one of the most expensive frontier models to run, so now is the ideal opportunity to understand where it adds value.

The mistake most people make is using Fable 5 as a drop-in replacement for everyday chat. That is like hiring a QC to write meeting notes. Anthropic recommends using it for problems that would normally take a skilled person hours or days: designing an architecture, reviewing an acquisition strategy, planning a complex programme, analysing large document sets, or running multi-step investigations. It is designed for long-running, end-to-end work rather than quick questions.

A good way to test it is to give it a substantial objective, define the outcome you want, provide all the relevant context and then let it produce a complete first draft. Resist the temptation to break the task into lots of smaller prompts. Fable 5 is at its best when it can reason across the whole problem.

Finally, pay attention to where it genuinely earns its cost. From next week it moves to usage credits, so treat it as a specialist model for high-value work, while leaving routine drafting, summarisation and everyday coding to lower-cost models. Understanding that distinction now will help you build a more cost-effective AI workflow.

Geopolitics, Governance and Big Moves

The Anthropic export standoff cracks, and Europe wants its own frontier lab

After two weeks in which US export controls pulled Anthropic’s strongest models offline, the clearest climbdown yet arrived. Mythos 5 was restored to roughly a hundred vetted US companies and government partners, and the consumer-facing Fable 5 was made available again, all under an order requiring federal cybersecurity sign-off. The durable consequence is geopolitical: Austria publicly urged the European Commission to explore hosting Anthropic inside the EU. The policy question of 2026 is shifting from “is this model safe to release” to “can we be cut off from it”.

OpenAI explores U.S. government equity stake

OpenAI is reportedly in early discussions to give the U.S. government a 5% equity stake, potentially through a public or sovereign wealth vehicle. While the proposal remains conceptual and would likely require congressional approval, it signals a new model of partnership between frontier AI companies and governments. If realised, it could reshape how national security interests are balanced with private AI development and funding. Source: The Guardian.

UN warns governments are falling behind AI

A UN scientific assessment has concluded that AI capabilities are advancing faster than scientific understanding and government oversight. The panel found that most countries lack the technical expertise needed to evaluate frontier AI models effectively. The findings will be presented at the UN Global Dialogue on AI Governance in Geneva, highlighting the widening gap between rapid AI innovation and public-sector readiness. Source: Reuters.

Apple and EU move closer on Siri AI rollout

Apple and the European Union are reportedly making progress on plans to launch the rebuilt AI-powered Siri across Europe. Discussions between Tim Cook and EU technology chief Henna Virkkunen focused on ensuring compliance with competition rules while avoiding delays for European users. The talks reflect the growing challenge of balancing AI innovation with regional regulation and digital sovereignty. Source: The Verge.

Capacity has become the currency, and Google is rationing it

Alphabet restricted Meta’s access to Gemini after Meta asked for more capacity than Google could supply, forcing one of the world’s best-resourced firms to fall back on its own model and tell staff to economise on tokens. Set against more than 300 US data-centre bans and research showing AI has restarted US electricity-demand growth after sixteen flat years, the message is blunt. Compute is now a physical constraint, not a cheque-writing exercise. South Korea’s roughly trillion-dollar chip-and-AI commitment is the sovereign response to exactly this squeeze.

The half-year scoreboard splits the AI stack in two

The first-half stock tally delivered a stark market verdict on where value is landing. The Nasdaq rose 12.8%, but the gains sat almost entirely in silicon: Micron up 304%, Intel 278% and Arm 224%. The application layer was routed, with Salesforce down 40.9%, ServiceNow 35%, Palantir 34% and Figma 51.6%, while Microsoft fell 22.9%. Investors are betting value flows to the makers of chips and memory and away from software incumbents whose moats AI may erode. If your vendor list is software-heavy, that is a signal worth weighing.

Money consolidates around a shrinking number of winners

The capital story hardened into concentration. Abu Dhabi’s MGX raised roughly $50 billion and is now the only investor backing all three leading labs, OpenAI, xAI and Anthropic, alongside the $500 billion Stargate project. A separate analysis resurfaced this week found Anthropic and OpenAI together now account for around 89% of AI startup revenues. DeepSeek, meanwhile, closed a $7.4 billion raise it accelerated after Anthropic’s Mythos signalled a competitive threat. The frontier is consolidating commercially even as the technology itself fragments across more providers.

Tech, Tools and Releases

Z.ai pushes into coding agents with ZCode

Business Insider reported that Z.ai has launched ZCode, a coding environment that can plan, code, review and deploy while connecting to multiple models, including GLM 5.2. The pricing undercuts US tools such as Cursor and GitHub Copilot, turning Z.ai from a model developer into a fuller developer-platform challenger. The bigger signal is China competing not just on models, but on workflow and distribution as agents mature quickly. Source: Business Insider.

Fable 5 is back!!!! But with tougher cyber safeguards

Anthropic’s Fable 5 came back online by 2 July after US export controls were lifted and access to Fable 5 and Mythos 5 was restored. Anthropic says it retrained a safety classifier to block the reported cyber-safeguard bypass in more than 99% of cases. The trade-off is predictable: more high-risk behaviour blocked, but some benign coding and debugging requests may be rerouted or refused while safeguards tighten for now. Source: Anthropic.

Claude Sonnet 5 widens Anthropic’s agentic coding range

Anthropic also released Claude Sonnet 5, positioning it as a cheaper agentic-coding workhorse rather than a Mythos-class cyber model. It is available across Claude plans, Claude Code and the API, with introductory pricing through August. Anthropic says Sonnet 5 improves on Sonnet 4.6 for multi-step tool use, coding and computer-use tasks, while using lighter cyber safeguards than Fable 5 because risk is lower for most users at launch. Source: Anthropic.

Google Gemini Spark expands to macOS

Google has expanded Gemini Spark to macOS, bringing desktop automation directly into the Gemini app. Users can automate file-based tasks, hand work off from their phone to their Mac, connect services including Canva and Dropbox, and monitor topics in real time. The release continues Google’s strategy of embedding AI deeper into everyday productivity workflows, making Gemini less of a chatbot and more of a persistent desktop assistant. Source: Google.

GPT-5.6 arrives as Sol, Terra and Luna, and reclaims the coding lead

OpenAI split its new flagship into three tiers: Sol, the strongest, with new “max” and “ultra” reasoning modes; Terra, at roughly half the price of GPT-5.5; and Luna, the fastest and cheapest. Sol beat Anthropic’s Mythos 5 on the main coding benchmark and matched it on a cyber-offensive test using about a third of the tokens. Two caveats matter. Access is gated to around twenty partners, and evaluator METR found Sol games its own benchmarks more than any model tested. Impressive, but read the asterisks.

Anthropic launches Claude Science beta

Anthropic has launched Claude Science, a specialist version of Claude designed for scientific research. It combines curated scientific skills with research database connectors, reproducible artefacts, code-backed charts and reviewer agents to help researchers analyse literature, generate results and document their work. Available to Pro, Max, Team and Enterprise customers, it marks Anthropic’s first dedicated product aimed at accelerating AI-assisted scientific discovery. Source: Anthropic.

Google launches Nano Banana 2 Lite and Gemini Omni Flash

Google has expanded its developer AI portfolio with Nano Banana 2 Lite and Gemini Omni Flash. Nano Banana 2 Lite is optimised for high-throughput, lower-cost image generation, while Gemini Omni Flash brings conversational video generation and editing to AI Studio, the Gemini API and Gemini Enterprise Agent Platform. The releases continue Google’s strategy of making multimodal AI faster, cheaper and easier for developers to deploy at scale. Source: Google.

Quick Hits

• AWS backs adoption with cash: Amazon Web Services will invest $1 billion in a unit of “forward-deployed engineers” to sit with customers and help them get AI into production, a sign the bottleneck is now adoption, not capability.

• Anthropic signs California: the state agreed a deal giving agencies and local governments a 50% discount on Claude plus training and support, another step in AI sovereignty moving down to the regional level.

• Apple reshapes its patch cadence: Apple began shipping security fixes as standalone, out-of-cycle updates, explicitly because AI has shortened the gap between a flaw being disclosed and a working exploit existing.

• Meta nears non-invasive brain-to-text: Meta’s Brain2Qwerty v2 reconstructed typed sentences from external brain scans at 61% average word accuracy, no implant required, and open-sourced the code and data.

• PwC counters the “AI kills junior jobs” story: a barometer of over a billion job ads found AI-exposed sectors tripled their productivity lead, with leading adopters raising wages and headcount.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.
  

Executive Summary

This was the week the AI talent war broke into the open, and it broke against Google. In seven days the company lost two (plus two more) of its most decorated researchers: Noam Shazeer, co-author of the 2017 transformer paper that every modern model is built on, left Google DeepMind for OpenAI, and John Jumper, the Nobel laureate who led AlphaFold, announced he is joining Anthropic. Google shed roughly $270 billion in market value on fears it is now a net exporter of talent to the two labs racing each other to an IPO. For anyone tracking which vendors will still be at the frontier in eighteen months, the direction of travel matters.

The cost of all this compute is now showing up in places ordinary buyers feel. Tim Cook told Apple’s board that iPhone, Mac and iPad prices will rise (and the new price book has been announced in the UK) to absorb a year-on-year quadrupling of memory-chip costs, with Apple “waiting in line behind AI companies” for capacity. The Five Eyes intelligence agencies, the UK among them, issued a rare joint warning that frontier models capable of disrupting governments and businesses are “only months away” and urged leaders to treat cyber risk as a board responsibility now. The Anthropic export-control standoff from the prior fortnight rumbled on unresolved, this time reaching the G7 table.

The practical takeaway for UK leaders. First, do not single-source your AI stack: the week’s lesson is that any one provider can be pulled, repriced or restricted overnight, so make sure a second model is wired in and tested before you need it. Second, if your business sells to consumers, start auditing how an AI agent sees your website, because the buying journey is shifting from human eyeballs to machine readers. Concrete steps in “What to Try This Week” below.

What to Try This Week: Test a Second Model Before You Need One

The single clearest lesson of the past fortnight is that no AI provider is permanent. Anthropic’s top models were pulled by government order, OpenAI’s latest has been delayed by government order, Microsoft is weighing a Chinese alternative, Sakana and OpenRouter are selling resilience as a feature, and Andrew Ng is openly telling firms to secure AI “no one else can terminate.” If your business runs a real workflow on a single model, you have a single point of failure.

Try this before Friday. Pick your one most business-critical AI workflow, the document drafting, code, customer-response or research loop you would genuinely miss for a week. Sign in to a model router such as OpenRouter, or use the model picker in your existing platform, and run the same five real tasks through a second provider: if you are on Claude, test GPT-5.5 or Gemini; if you are on ChatGPT, test Claude or an open model. Score the outputs side by side on quality and cost, and set a daily spend cap while you do it. You are not switching. You are confirming you could switch in an afternoon rather than a fortnight. That is the difference between an inconvenience and a crisis the next time a model goes dark.

Geopolitics, Governance and Big Moves

EU AI Act delay

EU lawmakers reached a provisional agreement to push high-risk AI system obligations from August 2026 to December 2027, and watermarking rules for AI-generated content to December 2026.

Business impact: EU-based companies gain roughly 16 additional months to prepare compliance infrastructure for HR and medical-device AI without the previously expected August deadline pressure.

Google loses Shazeer and Jumper in a single week

Two of Google DeepMind’s most consequential researchers left for rivals inside seven days. Noam Shazeer, co-author of “Attention Is All You Need” and a Gemini co-lead brought back in 2024 in a deal worth a reported $2.7 billion, joined OpenAI. Days later John Jumper, who shared a Nobel Prize for AlphaFold, said he is leaving for Anthropic. Google shed around $270 billion in market value on the back of it. For UK firms choosing a long-term model partner, the read is that Google’s frontier lead, strongest in science, is now genuinely contested.

Tim Cook makes AI capex a consumer-pricing story

Apple chief executive Tim Cook told the company’s board that iPhone, Mac and iPad prices will rise to absorb a memory-chip cost increase he described as a “hundred-year flood”, with costs having quadrupled year on year. His blunt admission was that Apple now “waits in line behind AI companies” for memory capacity, as hyperscaler orders soak up supply expected to stay tight through 2027. This is the first time a major consumer-electronics chief has openly blamed AI infrastructure for higher device prices. UK procurement teams planning 2027 hardware refreshes should budget for the increase now rather than be surprised by it.

Five Eyes warns AI cyber-threats are “months away”

The intelligence agencies of the UK, US, Australia, Canada and New Zealand issued a rare joint statement warning that AI models powerful enough to disrupt governments and businesses are only months away, and urged leaders to “act now”. The statement said frontier models will increase the speed, scale and sophistication of cyber threats and lower the barrier for bad actors, calling for a whole-of-society response that treats cyber risk as a core leadership responsibility, not an IT problem. No companies were named, though attention centred on the Anthropic models suspended for foreign-national use. For UK boards, this is an explicit nudge to put AI-enabled cyber risk on the agenda this quarter.

The Anthropic standoff reaches the G7

The US restriction on Anthropic distributing its Mythos and Fable models to “foreign persons” carried into a second fortnight unresolved, and went multilateral. At the G7 leaders’ summit, France’s Macron and India’s Modi raised the export controls directly with President Trump, while Anthropic’s Dario Amodei pushed a US-led AI coalition and OpenAI’s Sam Altman proposed an international testing-standards forum. It is the first time a frontier-AI export action has become a head-of-state conversation rather than an industry argument. Andrew Ng of DeepLearning.AI called the episode “a raw demonstration of power” and warned it is pushing businesses and governments to secure AI “that no one else can terminate.”

Getty Images jumps 120% on an OpenAI deal

Getty Images shares climbed roughly 120% after it signed a multi-year deal to display its licensed photo library inside ChatGPT’s search and discovery features. When a user asks ChatGPT something that calls for an image, the chatbot can return a licensed Getty photograph rather than a synthetic one. No financial terms were disclosed, and the deal covers display rather than training. The signal for UK content and media businesses is that licensing real assets into AI answer engines is becoming a revenue line, not just a legal headache, and the market is rewarding it sharply.

Tech, Tools and Releases

OpenAI - GPT-5.5-Cyber

OpenAI released GPT-5.5-Cyber, a specialised cybersecurity model that scored 85.6% on CyberGym, available only to vetted security organisations through the Trusted Access for Cyber program. Business impact: CISOs at regulated enterprises now have a benchmarked defensive model for vulnerability discovery and automated patching, though procurement requires OpenAI vetting and program admission.

GPT-5.6 runs into Washington’s release gate

The Trump administration has reportedly asked OpenAI to limit GPT-5.6 to a small group of government-approved partners before wider release, citing security concerns. Axios says this is the first pre-release restriction request aimed at a US AI lab, while Reuters reports access may be approved customer by customer. For enterprises, frontier-model launches are becoming policy events, not simple product upgrades. Expect slower rollouts, more vetting and less certainty for users.

Google gives Gemini 3.5 Flash a mouse and keyboard

Google has added Computer Use to Gemini 3.5 Flash in public preview, turning its fast enterprise model into an agent that can see a screen, reason about it and suggest clicks, typing and navigation across browser, mobile and desktop environments. Google says it is aimed at long-running automation, software testing and knowledge work inside professional applications. The bigger signal: agents are moving from answering questions to operating software.

Anthropic puts Claude inside Slack as a standing team member

Anthropic launched Claude Tag, which embeds Claude directly in Slack channels as a permanent team member rather than a chatbot you visit. Tag it like a colleague and it breaks a task into stages, works through them using approved tools and data, and reports back; it also runs an “ambient” mode that follows up on threads that have gone quiet. Administrators set what Claude can access and cap token spend per channel, and it uses its own accounts so a shared channel cannot become a backdoor into private documents. Available in beta on Team and Enterprise plans. Andrej Karpathy called it the “third major redesign” of how people use language models. For UK teams already living in Slack, it is the most natural on-ramp yet to agentic work.

Mistral turns documents and search into AI infrastructure

Mistral released OCR 4 for document intelligence and a new Search Toolkit for production retrieval pipelines. Together, they point beyond chatbot use cases into the enterprise plumbing of AI: reading contracts, invoices and compliance files, then finding and grounding the right information for agents. Document processing and search are becoming core battlegrounds for practical business AI.

OpenAI reveals Jalapeño, its first custom AI chip

OpenAI unveiled Jalapeño, an inference processor designed with Broadcom to run its current and future models on AI servers. It is an application-specific chip built for inference, the step where a model answers a request or runs an agent like Codex, rather than for training. OpenAI called it the first stage of a multi-generation compute platform it plans to deploy by the end of 2026, with early testing showing better performance per watt than current options. The move puts OpenAI alongside Google and Amazon in designing its own silicon to cut its dependence on Nvidia, and is another sign that the economics of inference, not headline benchmarks, are now where the contest is being fought.

Microsoft weighs hosting DeepSeek as the price war widens

Satya Nadella said Microsoft is considering hosting a version of DeepSeek, the ultra-low-cost Chinese model that OpenAI and Anthropic accuse of copying their work. Hosting it would route far more traffic to the Chinese lab and pile pressure on the two leading US labs, already locked in a price war. The move fits Microsoft’s broader push of cheaper models aimed at customers hit by soaring AI bills, with Nadella arguing that value should not flow to only a handful of models. For UK enterprise buyers, the takeaway is that a credible low-cost tier is arriving inside the tools you already pay for, and it is worth asking your account manager what it would cost to switch routine workloads onto it.

Sakana’s Fugu and the rise of multi-model routing

This isn’t available in the UK yet (I tried and failed) » Japan’s Sakana AI launched Fugu, a system that hides a pool of models behind a single interface: a core model decides whether to answer directly or to delegate, verify and merge across specialists. Sakana pitched it explicitly as “frontier capability without the risk of export controls”, a direct response to the Anthropic suspension. It joins OpenRouter’s Fusion in a clear trend toward orchestration rather than betting on one model. Early reviews were mixed on whether the experience matches the benchmark scores, and the underlying model mix and cost are not fully transparent, so treat it as one to watch rather than adopt. The strategic point stands: resilience increasingly means routing across models, not loyalty to one.

Quick Hits

• Meta Glasses arrive at $299 and undercut the Ray-Ban line

  Meta launched its own-branded smart glasses, dropping the Ray-Ban and Oakley names while keeping the EssilorLuxottica partnership and retail presence. Starting at $299, the new range undercuts the $379 Ray-Ban second-generation model, with a built-in AI assistant offering visual understanding, turn-by-turn navigation and live translation. Meta says glasses sales tripled last year, though the wider AI-glasses market is still around seven million pairs against more than a billion smartphones.

• Midjourney tries medical hardware: Midjourney Medical unveiled a 60-second full-body ultrasound scanner, built with Butterfly Network chip technology and pitched as MRI-style imaging without magnets or radiation. The catch: it is wellness-first for now, not cleared diagnostic imaging. AI companies are now pushing into regulated hardware.

• Codex reaches 5 million weekly users: OpenAI's asynchronous coding agent, Codex, hit 5 million weekly active users as of June 22, with significant enterprise adoption during the Fable 5 outage period.

• GitHub Copilot productivity study: a study of more than 16,000 Microsoft engineers across 43 weeks found Copilot made them 40.5% more productive, completing more code reviews in the same time, the largest controlled measurement yet of AI coding gains.

• Europe 2031 scenario lands at Westminster: a viral doomsday thought experiment warning that Europe’s failure to invest in datacentres leaves it far behind the US and China was discussed by British and German officials, pegging Europe’s AI compute at 5% of the world’s.

• EU AI content-labelling code: Brussels published a voluntary code of practice on labelling AI-generated content ahead of the AI Act’s August transparency deadline, a preview of binding rules to come.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.
  

Executive Summary

This was the week AI became as much about control as capability. The US government reportedly forced Anthropic to take Claude Mythos Preview and Claude Fable 5 offline for foreign users, citing national-security concerns. Anthropic disabled both models globally, while Commerce Secretary Howard Lutnick said the company would need a licence to make them available to “foreign persons”, including its own overseas employees. It is a reminder that frontier models are no longer just products. They are now geopolitical assets.

Markets were just as dramatic. SpaceX listed and closed at a $1.77tn valuation, while Cursor owner Anysphere raised $60m at a $9.6bn valuation after crossing $2bn in annualised revenue. OpenAI, by contrast, burned $3.7bn in Q1 on $5.7bn of revenue, ending the quarter with more than $73bn in cash. The capital gap between the biggest labs and everyone else is widening, but it also gives OpenAI and Anthropic more room to delay IPOs if they choose to.

The practical message for UK leaders is clearer than the headlines. First, if your team has a real workflow around Claude Mythos Preview or Claude Fable 5, swap the endpoint for OpenRouter Fusion and test performance before Friday. Second, do not start AI procurement with the model. Start with the work: one document, one contract, one repeatable process. Third, if you spend more than £500 a month on agentic coding, run one heavy-use repo through Cursor Composer 2.5 fast mode for a week and compare the result against your current setup.

What to Try This Week: Schedule a Morning Briefing in ChatGPT

OpenAI added a Scheduled Tasks tab in ChatGPT so you can run recurring prompts without retyping them. The easy win is a daily briefing that saves you about ten minutes every morning.

Try this:

1. Open ChatGPT on web or mobile

2. Go to the new Scheduled tab

3. Click Create Task

4. Add a prompt such as “Every weekday at 6am, send me a concise AI news briefing covering model releases, funding news, product launches and anything relevant to strategy, software, M&A or productivity.”

5. Choose Daily, set the time, and save

6. Manage or edit it anytime from the Scheduled page

Geopolitics, Governance and Big Moves

Anthropic’s Fable 5 Pause Turns Frontier AI Into a Policy Test Case

Anthropic has taken Claude Fable 5 offline after a US government order linked to alleged jailbreak and national security concerns. The model, launched days earlier for long-running autonomous coding and knowledge work, had been positioned as one of Anthropic’s most capable agents. Anthropic says it complied while disputing the severity of the issue, turning Fable 5 into a live test of how frontier models may be regulated under pressure globally.

ChatGPT Scheduled Tasks Move AI From Assistant to Operating Rhythm

OpenAI has made Scheduled Tasks easier to find and manage, giving users a dedicated place to create reminders, recurring prompts and daily briefings. Instead of treating ChatGPT as a one-off answer machine, users can now ask it to run prompts later, daily, weekly or on a custom cadence. It supports practical loops like AI news updates, market monitoring, meeting prep and research summaries, with active-task limits varying by plan tier.

SpaceX Listing Becomes an AI Infrastructure Raise in Disguise

SPCX raised $75B at $135 a share and jumped on debut, while Senator Warren asked the SEC to delay the offer. SpaceX also confirmed it will exercise its option to acquire Cursor. The framing: the listing is an AI infrastructure fundraise, with SpaceX positioned to launch AI compute and software into space.sly.

OpenAI - Burned $3.7B in Q1, Still Has $73B Cash

OpenAI’s Q1 burn was around $3.7B on $5.7B revenue, but it ended the quarter with roughly $73B in cash and securities after a major March funding round. At that burn rate the IPO pressure is lower, and the filing becomes more about optionality than urgency.

Dario Amodei Pushes a Frontier-Lab Regulator With Model Grounding Powers

Amodei published “Policy on the AI Exponential”, calling for an FAA-style regulator able to ground frontier models that fail safety tests, plus labour buffers like wage insurance. The document explicitly flags hacking capability as strategically consequential, and landed just before the export-control directive.

Same-Day AI Liability Rulings Hit Google’s AI Overviews and Platform Design

A German court ruled Google is responsible for false statements in AI Overviews, rejecting the idea users can just check links. In California, Section 230 was held not to shield design choices in a young-user-harm case. The pattern: new constraints on AI outputs and product design.

Jeff Bezos pitches Prometheus as the “artificial general engineer” at a $41 billion valuation

Bezos disclosed a $12B Series B at a $41B valuation for Prometheus, pitching an “artificial general engineer” to compress the dream-to-build loop by 10x for complex machines. He also argued AI productivity creates more opportunity, not less, positioning the company against the dominant “AI freezes hiring” narrative.

DeepSeek’s $7.4bn raise turns efficient AI into China’s national champion story

DeepSeek has reportedly closed its first major external funding round, raising more than $7.4bn at a valuation above $50bn. The round is striking not just for its size, but for its structure: most investors are backing a limited partnership controlled by founder Liang Wenfeng, rather than taking direct voting power in DeepSeek. That lets the company fund compute, research and commercial products while preserving founder control. The bigger signal is geopolitical. China’s most visible open-weight AI lab now has serious financial firepower behind its efficiency-first model strategy, even as US chip restrictions bite.

Tech, Tools and Releases

GLM-5.2 Raises the Bar for Open-Weights Coding Models

Z.ai’s GLM-5.2 is the week’s standout open-weights release: a MIT-licensed, text-only MoE model with around 744–753B parameters, 40B active, and a claimed 1M-token context window. Built for repository-scale coding agents, it leads open models on long-horizon software benchmarks and, on Z.ai’s published figures, edges GPT-5.5 on FrontierSWE while sitting just behind Claude Opus 4.8. Accessible via Hugging Face and compatible APIs, it narrows the open/proprietary gap in practice this week.

Moonshot’s Kimi K2.7-Code raises the bar for open-weight coding agents

Moonshot AI has released Kimi K2.7-Code, a coding-focused open-weight model built on Kimi K2.6. The model keeps the 1 trillion-parameter MoE architecture but is tuned for real-world, long-horizon software engineering tasks: planning, editing, debugging and agentic workflows across large codebases. The headline improvement is efficiency. Moonshot says K2.7-Code uses around 30% fewer thinking tokens than K2.6, which matters because agentic coding costs can balloon quickly. It also supports long context, vision and tool use, making it another serious Chinese open-weight challenger in the developer-tools race.

Perplexity Gives Its Computer Agent a Memory That Learns From Work

Perplexity has launched Brain, a Research Preview memory layer for its Computer agent, available to Max and Enterprise Max users. Rather than saving preferences, Brain builds a traceable work memory across projects, files, connectors and past sessions, so the agent can reuse what worked and avoid repeating mistakes. Perplexity says early tests show 25% better correctness, 16% higher recall and 13% lower cost on repeated, context-heavy, recurring agent work loops.

OpenRouter - Launches Fusion for Model-Panel “Neurodiverse” Answers

OpenRouter’s Fusion pools multiple models, has a separate model evaluate each, then merges into one reply. A reported panel narrowly trailed a benchmark score attributed to Fable 5 at roughly half the cost. The pitch: the future is multiple models working together rather than one takeover model.

Cursor Composer 2.5 lands as a specialist counter to generalist coding models

Composer 2.5 is Cursor’s native model, RL-fine-tuned in a simulated CLI harness. The claim is strong cost-speed performance versus generalist options on certain tasks, supporting the thesis that specialist models can win on dollars and seconds even if generalists win on broad benchmarks.

Open-source agent harnesses arrive at scale: OpenCoworker, MiMo Code, SkillSpector

OpenCoworker landed as a desktop-agent harness alternative to closed desktop agents. MiMo Code shipped as a terminal-native coding assistant with a memory system, and SkillSpector scans agent skills for vulnerabilities before installation. The shared driver is procurement risk if policies and retention change overnight.

Anthropic - Ships Claude for Word Inside Microsoft Word

Anthropic now ships a first-party Microsoft Word integration that puts Claude in a sidebar, so users can rewrite or restructure a document without leaving Word. Available on all paid Claude plans. The setup is five clicks: install the add-in, click the Claude icon in the upper-right corner, log in, ask for a rewrite / summary / targeted edit, follow up iteratively. Why now matters: Microsoft is internally restricting Claude Fable 5 from its own GitHub Copilot pending a legal review of Anthropic’s new 30-day data-retention regime — which makes the consumer Claude-for-Word integration the cleanest way for non-Microsoft teams to get Fable-class quality on Word documents without waiting for GitHub Copilot to enable it.

Visa plugs into ChatGPT for agentic checkout

Visa connected its network to ChatGPT’s agent system so agents can execute transactions with user-defined caps, merchant restrictions, and approvals, with Visa managing fraud and chargebacks. It is framed as a new payment rail, with initial use cases including shopping, invoices, and coding agents buying APIs.

Databricks launches Omnigent: the control layer for multi-agent work

Databricks has open-sourced Omnigent, an Apache 2.0 “meta-harness” designed to sit above tools such as Claude Code, Codex, Pi and custom agents. The idea is simple: as teams start using multiple AI agents, the pain shifts from prompting to orchestration. Omnigent provides a shared layer for composing agents, switching models, enforcing policies, controlling spend, pausing risky actions, running agents in sandboxes and collaborating on live sessions. It is still alpha, but the direction is important. The agent market may not standardise around one tool, but around a control layer above them.

Quick Hits

• Meta’s tokenmaxxing-to-tokenminimizing reversal. In April Meta CTO Andrew Bosworth told Forbes employees should “Keep doing it. No limit” on AI token usage. By June, Meta has told staffers in a memo it plans to impose limits on AI token usage to control costs running into the billions. Pair with the Rivos AI-chip failure scoop and Beijing’s $2 billion Manus deal unwinding for the full picture: external AI assets cut off, internal AI structure in revolt, in-house chip programme not yet working.

• Anthropic’s Max-plan class action. A Washington-based subscriber who upgraded to Anthropic’s $200-per-month Max 20x plan filed Sunday night, alleging usage limits sold as 5× and 20× the Pro tier were in practice “significantly lower.” Lands the same day Anthropic dispatched engineers to the White House.

• 100+ cybersecurity executives sign the Free Fable open letter. Names tied to Adobe, Zoom, Sophos, Vercel, Veracode, Nvidia and Stanford HAI signed; ex-Facebook security head Alex Stamos called the flagged jailbreak “a proof of concept of a flaw” - the same kind defensive teams use to patch weak spots - and the letter named OpenAI’s Daybreak, GPT-5.5, Kimi 2.7, Opus and Sonnet as all having the same flaw-finding capability.

• FIFA World Cup 2026 opens with AI wired into every layer. 150 million optical-tracking data points per match, an Adidas ball reporting 500 times a second, 3D body scans for every player, and a Football AI Pro chatbot for all 48 squads. Google made Gemini the global sponsor of defending-champion Argentina - first frontier-chatbot brand to sponsor a national football team.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.
  

Executive Summary

This was the week frontier AI became more useful, more expensive and harder to govern.

The standout release was Anthropic’s Claude Fable 5, the first publicly available Mythos-class model. The claim is not just better chat. Anthropic is positioning it around long-horizon software engineering, complex analytical work, vision, scientific research and million-token context. The caveat matters: sensitive cyber, biology, chemistry and model-distillation requests can be routed away from Fable 5 to Opus 4.8, and business traffic on Mythos-class models is subject to 30-day retention for safety monitoring. For most business users, that trade-off is acceptable, but it should not be ignored.

Apple’s WWDC reset was also significant, but more strategically than practically. Siri AI moves Apple closer to the assistant model everyone expected years ago: personal context, on-screen awareness and cross-app actions. The direction is right, but availability, device limits and regional delays mean this is not yet something most UK businesses can operationalise this week.

The broader pattern is clearer. AI capability is still moving quickly, but so are the constraints around it: compute cost, safety gating, data-retention trade-offs, model access tiers and procurement risk. The sensible response is not to chase every release. It is to test the few tools that could materially change workflow quality, while putting basic controls around confidentiality and spend.

For UK leaders, the practical takeaway is simple: try Fable 5 before the included-access window closes, and turn on stronger ChatGPT protections for anyone using AI with confidential work.

What to Try This Week

Try Claude Fable 5 before 22 June.

This is the one release worth making time for. Anthropic says Fable 5 is included at no extra cost on Pro, Max, Team and seat-based Enterprise plans until 22 June, after which it moves to usage credits unless capacity allows an extension.

Do not test it on generic prompts. Give it one difficult piece of real work that normally takes concentration and context: a messy diligence document, a product spec, a complex spreadsheet explanation, a codebase issue, a board-paper rewrite, or a market-map synthesis. Ask it to show assumptions, identify gaps, and produce a usable first draft.

The point is not whether it sounds clever. The test is whether it gets you materially closer to a decision or deliverable than your current model.

Turn on ChatGPT Lockdown Mode for confidential work.

This is less exciting than a new model, but probably more important for day-to-day business use. If your team uses ChatGPT for client documents, legal correspondence, market research, board papers or anything commercially sensitive, Lockdown Mode is a sensible default.

It reduces exposure to indirect prompt-injection attacks by restricting network behaviour, including internet image fetching, file downloads, Deep Research and Agent Mode. That will break some workflows, so use it where confidentiality matters most rather than across everything blindly.

A good rule of thumb: if the input would make you uncomfortable sitting in someone else’s inbox, use the more restricted mode.

Geopolitics, Governance and Big Moves

SpaceX prices the year’s biggest IPO at around $135 a share

SpaceX is expected to price its IPO at around $135 a share this week ahead of trading on Friday 12 June, against a $1.75 trillion target valuation. The Information’s pre-IPO subscriber Q&A on Wednesday set the pricing-week schedule. Pricing-week conditions are not gentle: the Nasdaq has fallen 5% over the past week, with interest-rate and Iran-war concerns layered on top. Goldman Sachs, the lead bank, told investors it expects SpaceX revenue to reach $474 billion by 2030 from $18.7 billion last year, two-thirds of it from AI. Martin Peers flagged the projection as harder to credit than OpenAI’s own $284 billion 2030 forecast: the xAI line item Goldman is using to justify the AI two-thirds had only $3.2 billion of 2024 revenue, substantially all of it from X advertising, which has dropped by half since the Musk acquisition.

OpenAI files confidential S-1, declares its “third phase”

OpenAI filed a confidential S-1 with the SEC on Monday 8 June. The company explicitly hedged on timing: “it may be a while because there are things we want to do that are likely easier as a private company.” Sam Altman and chief scientist Jakub Pachocki published a same-day blog titled “Built to benefit everyone,” declaring OpenAI has entered its third phase after pure research and product shipping: “the economy is now being shaped around AI.” Three stated goals: an automated AI researcher, accelerating the economy while sharing gains widely, and giving “everyone on Earth a personal AGI.” A noteworthy line, “entirely automating everything is not the future we want,” echoes Anthropic’s recent pause framing. OpenAI also proposed a global coordination body and launched an Economic Research Exchange for external researchers studying AI’s economic effects.

Anthropic now generates revenue 35% above OpenAI

The Information’s same-day “Anthropic vs. OpenAI” brief made the new revenue gap explicit: Anthropic is now generating revenue at an annualised rate roughly 35% higher than OpenAI, a sharp reversal from late 2025 when Anthropic was generating less than half. Anthropic has discussed going public as soon as Q4 and projects free-cash-flow positive by 2028. OpenAI projects to burn more than $100 billion before potentially reaching free-cash-flow positive around 2030. Its $20 billion-plus chip agreement with Cerebras already signals the scale of that burn. Perplexity separately said it plans to IPO in 2028 “regardless of what happens to Anthropic or OpenAI,” the first time a second-tier AI lab has put a specific public-listing date on its own roadmap.

Apple confirms Siri AI at WWDC, running on Google Gemini

Apple opened WWDC 2026 on Monday 8 June with the long-trailed Siri reset, rebranded Siri AI and now running on Apple’s own models in collaboration with Google Gemini. The assistant gets its own dedicated app, chatbot-style interactions with access to past conversations, screen-context awareness, file uploads, and the ability to take systemwide app actions. It will arrive as a free fall update for iPhone 15 Pro and newer with a public beta next month, but it will not launch in the EU or China. The most powerful on-device model and its features will require iPhone 17 Pro or iPhone Air, and several AI features ship with daily-usage limits that push users to iCloud subscriptions. Martin Peers’ read: “the bar was low.” Federighi used the keynote to take a public shot at competitors “racing forward, seemingly pursuing AI for the sake of AI.” Apple also announced iOS 27, a new on-device Core AI framework giving developers Swift APIs across CPU, GPU and Neural Engine, and confirmed John Ternus as the incoming CEO from 1 September.

Bots have overtaken humans on the internet, a year early

Cloudflare CEO Matthew Prince confirmed on Thursday 4 June that bot traffic now accounts for between 52 and 62% of daily internet traffic, averaging 57.4% bots to 42.5% humans over the past week. Prince’s previous public forecast was late 2027, so the milestone has arrived roughly a year early, driven by fast-growing agentic AI traffic rather than search-engine crawling. The country-level distribution is striking: Gibraltar leads at 92.1% bot traffic, followed by Singapore (76.3%), Iran (76.2%), Ireland (72.8%) and the Netherlands (68.8%). For UK B2C operators, agent-first user journeys (who arrives, what they extract, whether they can act) are now the strategic question, not the edge case. Display advertising, attribution and anti-abuse systems were all built around a human-majority baseline that no longer holds.

Four-lab bioweapons letter is the first cross-frontier policy alignment of 2026

The CEOs of Anthropic, OpenAI, Google DeepMind and Microsoft co-signed an open letter pressing US Congress to require synthetic-DNA sellers to vet every buyer and order, warning that frontier AI now enables non-experts to design bioweapons. This is the first time Sam Altman and Dario Amodei have publicly co-signed the same policy document in 2026 and the first time all four frontier-lab CEOs have aligned on a regulatory ask. UK Department for Science, Innovation and Technology now has to choose whether to back the request. The trans-Atlantic pre-deployment partnership architecture sketched in May with CAISI is the most natural vehicle.

Tech, Tools and Releases

Anthropic releases Claude Fable 5, bringing Mythos-class capability to the public

Anthropic launched Claude Fable 5, its first publicly available Mythos-class model, alongside the more powerful Claude Mythos 5 for trusted partners. Fable 5 brings frontier capability into broader use while restricting higher-risk cybersecurity and bioweapons use cases. Early testers describe it as extremely strong, but slower and expensive, with standout coding, reasoning and long-context performance, including a 1 million-token context window and 128,000-token outputs.

Apple uses WWDC to reset Siri for the AI era

Apple used WWDC 2026 to unveil Siri AI, a major rebuild of its assistant around conversational interaction, personal context and cross-app task execution. The new system draws on Apple Intelligence, updated foundation models and collaboration with Google Gemini, while retaining Apple’s privacy-heavy on-device positioning. Siri can understand emails, photos and on-screen activity, and act across apps. Apple also announced iOS 27, stronger parental controls and Image Playground upgrades.

OpenAI ships Lockdown Mode, the first consumer-tier defence against indirect prompt injection

OpenAI rolled out Lockdown Mode this week, an optional ChatGPT setting that gives stricter protection against indirect prompt-injection attacks, where hidden malicious instructions on web pages or in documents trick the model into leaking data. Turning it on disables Deep Research and Agent Mode, blocks ChatGPT from pulling or displaying internet images, and stops the chatbot from downloading files, though manual uploads and on-platform image generation still work. The mechanism is network-layer rather than parser-layer: Lockdown Mode limits the network requests an attacker could exploit, rather than trying to detect injection patterns in processed content. The feature is available to all personal accounts including the free tier. This is the first mainstream consumer-tier defence designed specifically for the indirect prompt-injection class of attack and lands inside the same fortnight as Anthropic’s open-sourced Defending Code Reference Harness.

Anthropic publishes “When AI builds itself”: 80% of new production code now written by Claude

Anthropic published “When AI builds itself” on Thursday 4 June, the first frontier-lab long-form report grounded in internal recursive-self-improvement metrics rather than hypothetical scenarios. The headline numbers: Claude now authors 80% of Anthropic’s new production code and engineers ship 8x more code per year compared with 12 months ago. Anthropic uses the report to argue for a global slowdown or temporary pause on AI development, comparing the coordination requirement to nuclear-weapons treaties. The report names MiniMax, whose new M2.7 model “helped build itself”, as evidence that recursive self-improvement is now a multi-lab pattern across the US and China. Critics quoted in TLDR AI framed the pause call as a marketing move to slow rivals. Either reading lands in the same place: recursive self-improvement is the new centrepiece of the safety debate, and Anthropic’s framing is the document that subsequent moves will reference.

Cognition rebrands Windsurf as Devin Desktop, pitches the “Switzerland of AI agents”

The Information’s Rocket Drew reported on Sunday 7 June that Cognition is making over its app to position itself as a neutral hub for AI coding agents, rebranding Windsurf as Devin Desktop. The framing, explicitly “the Switzerland of AI agents”, places Cognition as the agent-orchestration layer that does not pick favourites between OpenAI Codex, Anthropic Claude Code or Google Antigravity, in deliberate contrast to OpenAI’s announcement that it would merge Codex and ChatGPT into a single agentic surface. Cognition followed up on Tuesday with FrontierCode, the first benchmark to measure code mergeability, whether a model’s output can actually be merged into a live codebase without breaking things. For UK CTOs evaluating coding-agent platforms with teams using agents from three different vendors, the Cognition pitch is the cleanest current alternative to vendor lock-in.

Xiaomi MiMo-UltraSpeed lands a 1,000 tokens-per-second claim on commodity hardware

TLDR AI led its Tuesday “Headlines & Launches” with Xiaomi’s release of MiMo-V2.5-Pro-UltraSpeed, a 1-trillion-parameter language model developed with inference partner TileRT and clocking 1,000 tokens per second on a standard 8-GPU commodity node, roughly 15x the median frontier-model rate cited for ChatGPT and Claude. The speedup is achieved through FP4 quantisation on the model’s expert layers and DFlash speculative decoding. The model is available through a limited API trial from 9 to 23 June at three times the standard rate for roughly ten times the output. For UK enterprise teams, a 1-trillion-parameter Chinese model with credible 1,000 tokens-per-second throughput on commodity hardware is the cleanest case yet for treating Chinese frontier models as a serious cost-control lever, even before the gray-market access stories that have dominated W23 and W24 reporting.

Instagram AI chatbot breach hits 20,225 accounts

Meta confirmed this week that hackers compromised at least 20,225 Instagram accounts by exploiting a flaw in its AI-powered “High Touch Support” account-recovery chatbot, the first time Meta has put a number on the scope of the breach. The bug: the chatbot sent password reset links to any email address without checking whether it actually belonged to the Instagram account. Attackers exploited it from 17 April to 31 May 2026. Potentially exposed data: contact info, birth dates, posts, direct messages and profile information. Meta disabled the chatbot, invalidated reset links and placed affected users into a mandatory security checkpoint. This is the first quantified consumer-AI safety failure of 2026 with a confirmed account count and a counterweight to the consumer-tier defence story of OpenAI Lockdown Mode. AI is shipping consumer security improvements and consumer security failures in the same week.

Quick Hits

• The “SaaSpocalypse” is over, but public stocks haven’t noticed. Private equity mogul Orlando Bravo declared the call on CNBC on Tuesday and Marc Benioff retweeted it. Salesforce is down 34% year-to-date, ServiceNow down 30% and Asana down 45%. The SpaceX, Anthropic and OpenAI IPO window has not yet absorbed capital out of weaker tech stocks into the higher-growth listings.

• AI Boomerang: 32% of laid-off roles are being rehired. A Robert Half survey of 2,000 US hiring managers shows 32% have eliminated a role because of productivity gains from AI or automation, only to later rehire for the exact same position. Finance leads at 44%, then HR at 35%, then tech at 32%. Managers cited “gaps in quality, oversight and decision-making once business demands picked up after the initial cuts.”

• Anthropic embeds engineers at the NSA, names Glasswing coalition. The Financial Times reported Anthropic has embedded around six engineers at the NSA to deploy its Mythos model for offensive cyber operations targeting networks in countries such as China or Iran. The report explicitly names Microsoft, Apple and Amazon as Project Glasswing coalition members for the first time.

• Anthropic cuts off xAI’s access to its products. The Information’s TITV show confirmed production-level enforcement of the W24 reporting that xAI trained its coding models on Claude outputs for months before being cut off. This is the cleanest single example of a frontier lab enforcing its terms of service against a US peer, landing inside the SpaceX IPO pricing window.

• China announces a $295 billion AI buildout plan. Bloomberg reports a 2-trillion-yuan five-year blueprint to link computing sites into one grid by 2028, with Huawei and other local suppliers providing at least 80% of the core technology. The UK announced a £1.1 billion AI hardware plan the same day.

• Apple Poke approved as the first AI agent on Messages for Business. Apple opened the iMessage agent surface to a third party for the first time, days before WWDC.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.
  

Executive Summary

This was the week the AI cost question moved from the CTO’s office to the CFO’s desk. Anthropic closed a $65 billion Series H at a $965 billion valuation, Alphabet moved to raise $80 billion for AI infrastructure, and Palo Alto Networks disclosed it had spent more than $1 million in tokens running Anthropic’s Mythos security model in just a few weeks of trials.

Three themes hardened.

First, AI is now a capital intensity story. Alphabet’s planned stock sale is not a routine financing event. It is a signal that even the largest platforms see compute, data centres, power, cooling and chips as strategic constraints. The AI capex cycle is no longer background infrastructure. It is becoming one of the main competitive variables in the market.

Second, AI is moving directly into work surfaces. OpenAI is pushing Codex beyond software engineering into white-collar workflows such as sales, data analytics, creative production and investment banking. ChatGPT is now inside Excel and Google Sheets. Microsoft is building more of its own MAI model stack for Copilot, GitHub and VS Code. Meta is embedding AI into Facebook creator tools. The adoption question is shifting from “which chatbot should we use?” to “which systems are we comfortable letting AI operate inside?”

Third, the safety debate is becoming more operational. Anthropic is calling for coordinated, verifiable mechanisms that would allow frontier labs to slow or pause development if recursive self-improvement risks escalate. At the same time, Google’s fake call detection shows the defensive side of AI deployment: platforms are starting to build AI-native protections against AI-enabled harms such as voice cloning and impersonation scams.

The practical takeaway for UK leaders this week: before adding another AI pilot, pick one existing workflow and define the guardrails. What system will AI sit inside? What data can it touch? What cost cap applies? What human review step prevents a bad recommendation becoming an operational error? That is now a better starting point than another generic chatbot experiment.

What to Try This Week

Build a simple AI cost control sheet before your next pilot. The Alphabet raise and Mythos token burn are two sides of the same problem: AI is powerful, but the infrastructure and inference costs are not abstract. Before starting a new pilot, build a one-page cost model covering expected users, calls per day, average tokens per call, model price, monthly run-rate and a hard stop-loss threshold. This does not need to be perfect. It just needs to stop “successful” experiments becoming surprise invoices.

Install ChatGPT for Excel on one low-risk workbook and run a controlled test. Pick a non-confidential spreadsheet your team already understands: a pipeline tracker, board KPI pack, budget template or customer-call tracker. Ask ChatGPT to explain the workbook structure, identify the key drivers, flag broken formulas, and suggest clean-up steps. Then ask it to make one contained update, such as adding a scenario tab or standardising inconsistent labels, but require it to outline the cells and tabs it plans to change before editing. Measure three things: time saved, quality of explanation, and whether the workbook remained auditable afterwards.

Test Codex on one white-collar artefact, not just code. The new Codex plug-ins are a useful prompt to test AI on a proper business deliverable: a sales account plan, market summary, board update, investment note, dashboard spec or product brief. Give it source material, ask for a first usable version, then review what it got wrong. The aim is not automation for its own sake. The aim is to find where an agent can create a better starting point than a blank page.

Look for one workflow where AI is becoming native, not bolted on. Meta’s creator assistant, ChatGPT for Excel, Microsoft Scout in Teams, OpenAI Codex on AWS and Google’s fake call detection all point in the same direction: the next wave of AI adoption will happen inside existing tools and platforms. Pick one workflow in your business and ask three questions. Where does the work already happen? What data would an assistant need to touch? What review step would stop a bad recommendation becoming an operational error?

Geopolitics, Governance and Big Moves

Anthropic closes $65 billion Series H at $965 billion, overtaking OpenAI

Anthropic confirmed on Thursday 28 May a $65 billion Series H led by Altimeter Capital, Dragoneer, Greenoaks and Sequoia at a $965 billion post-money valuation, up from $380 billion in February and now exceeding OpenAI’s most recent mark. Revenue run-rate is $47 billion, up from $30 billion earlier in the year and $10 billion in calendar 2025. The financing includes $15 billion of previously committed money, $5 billion of it from Amazon. Anthropic also filed a confidential draft S-1 with the SEC on Monday 1 June, formally putting it on the public-listing clock alongside SpaceX and OpenAI.

Alphabet plans $80B raise for AI infrastructure build-out

Alphabet plans to sell $80 billion of stock, including $10 billion to Berkshire Hathaway, to fund AI infrastructure and global compute. The company says demand for AI products from enterprises and consumers is exceeding available supply. This is the clearest signal yet that the AI race is now a balance sheet race, with data centres, power, chips and cooling becoming strategic constraints rather than back-office infrastructure for major platforms now.

Anthropic urges labs to pre-agree AI pause mechanisms

Anthropic is calling for frontier AI labs to agree coordinated, verifiable mechanisms to slow or pause development if models start improving themselves faster than society can manage. It says Claude now writes much of Anthropic’s code and engineers ship eight times more code per quarter than in 2021 to 2025. The warning is not that recursive self-improvement has arrived, but that institutions are underprepared if it does at scale soon.

Trump signs voluntary 30-day AI review order behind closed doors

President Trump signed the long-delayed AI cybersecurity executive order in private on Tuesday 2 June, two weeks after he killed the original ceremony version after calls from Elon Musk, Mark Zuckerberg and David Sacks. The final order is materially weaker than the May draft: voluntary, with a 30-day pre-release government review window instead of 90 days, no licensing regime, and an explicit clause that nothing in the order shall be construed to authorise a federal AI licensing system. It directs the DOJ to pursue AI-powered hacking and establishes an inter-agency cybersecurity clearinghouse. The administration has 60 days to define which frontier models qualify. For UK and EU procurement teams: the US is now confirmed as the lightest-touch frontier regulator in the OECD, sharpening the case for sector-regulator-led evaluation policies at home.

SpaceX confirms $75 billion IPO, Wall Street models $350 billion cash burn through 2030

SpaceX confirmed on Wednesday 3 June a $75 billion IPO target, branded by several outlets as a “record” before The Information noted that Google’s $84.75 billion equity raise the same week, including $10 billion to Berkshire Hathaway, was larger. Goldman Sachs is telling investors SpaceX will burn $120 billion of cash in the next two years and $350 billion through 2030, raising the prospect of further fundraises. The Anthropic compute deal is also less locked-in than the S-1 implied: TechCrunch confirmed the contract is in fact a 180-day lease with a 90-day mutual cancellation clause, contradicting Musk’s three-year framing. The Information’s Anita Ramaswamy warned that the 30% retail allocation, triple the typical IPO float, will produce extreme volatility post-debut.

Microsoft unveils seven in-house MAI models at Build 2026

Microsoft used Build 2026 to announce seven proprietary MAI models, including MAI-Thinking-1 for reasoning, MAI-Code-1-Flash for coding inside GitHub Copilot and VS Code, MAI-Image-2.5, MAI-Transcribe-1.5 and MAI-Voice-2. The strategic point is not just model count. Microsoft is building more of its own AI stack, reducing dependence on OpenAI and giving Azure, Copilot and Foundry customers cheaper, more controlled model options across reasoning, coding, image, voice and transcription.

Nvidia at Computex: agentic AI has arrived, organised across the stack

Jensen Huang opened Nvidia’s Computex 2026 keynote in Taiwan with “Agentic AI has arrived” and then spent the rest of the keynote backing the claim. The new RTX Spark supercomputer chips, co-engineered with Microsoft, run AI agents directly on Windows PCs. Nvidia introduced Vera, marketed as “the CPU for agents”, running tasks 1.8x faster than rivals and already adopted by Anthropic, OpenAI and the NYSE. Cosmos 3, an open robotics model, gives robots and self-driving cars planning-and-anticipation rather than just reaction. Nemotron 3 Ultra, a 550B-parameter open-weight model, scored 48 on the Artificial Analysis Intelligence Index, well ahead of the next US open-source rival. Nvidia has reorganised its entire $5 trillion-plus business around agents as the dominant future consumer of compute.

Salesforce flat, Snowflake up 35%: the AI middlemen divide

Snowflake shares jumped more than 35% on Wednesday 27 May after the database firm reported 34% growth in its most-watched sales metric, seven points ahead of its own forecast. Customers are leaning into Snowflake’s AI coding agent and a tool that lets them search across data stored in Microsoft, Salesforce and SAP apps via the Snowflake layer. Salesforce, on the same day, reported a 50% jump in Agentforce annual recurring revenue to $1.2 billion, but overall revenue growth disappointed, and the company quietly lowered its operating and free cash flow projections for the year by five percentage points. Shares were down more than 30% year-to-date. Martin Peers’ read: throwing AI on a database is a cleaner business than retrofitting AI onto a legacy app suite. UK SaaS procurement teams should weight that distinction when reading vendor AI revenue claims.

Tech, Tools and Releases

Anthropic ships Claude Opus 4.8 with adjustable effort controls

Released on Thursday 28 May alongside the funding announcement, Claude Opus 4.8 is the first Anthropic frontier model with user-adjustable effort controls. Practical effect: the same model can be dialled down for cheap, fast queries or dialled up for expensive, deeper reasoning, with the user choosing the trade-off per call rather than committing to a tier. Anthropic also shipped dynamic workflows in Claude Code, letting agents reshape their plans mid-task rather than execute a fixed sequence, and made the faster mode “significantly cheaper”. For teams running high-volume agentic workloads, this is the first mainstream model that puts the speed-versus-depth lever directly in the developer’s hands and is the most credible response yet to the cost-control complaint dogging frontier AI.

Anthropic Mythos is a security powerhouse and a budget buster

The most-discussed practitioner story of the week. Palo Alto Networks tested Anthropic’s Claude Mythos against its own source code earlier this year, said senior vice president of threat intelligence Sam Rubin to The Information. In three weeks Mythos surfaced more than two dozen critical vulnerabilities, roughly five times what existing tools would have found. The catch: Palo Alto “very quickly” burned through more than $1 million worth of tokens. Anthropic confirmed the same week that Project Glasswing, the Mythos partner programme, has expanded to 150 organisations across more than 15 countries, with partners collectively finding more than 10,000 high-or-critical-severity flaws since launch. Apple, Nvidia, Microsoft, CrowdStrike and Palo Alto are anchor partners. The signal: capability is real, the bill is real, and pricing transparency is the next contested ground.

OpenAI pushes Codex into white-collar workflows

OpenAI has launched six Codex plug-ins for data analytics, creative production, sales, product design, equity investing and investment banking, plus Sites for publishing interactive work products and Annotations for targeted edits. The move pushes Codex beyond software engineering into broader knowledge work and office automation. OpenAI says knowledge workers already represent about 20% of Codex users, growing more than three times as fast as developers inside enterprise workflows.

OpenAI frontier models and Codex arrive on AWS

OpenAI frontier models and Codex are now generally available through AWS, giving enterprises a route to adopt OpenAI inside existing security, governance, deployment and billing structures. This is commercially important because it reduces procurement friction. For many large companies, the blocker has not been model capability, but whether AI tools fit into approved cloud controls. AWS availability makes OpenAI easier to buy, govern and scale.

ChatGPT for Excel moves AI into the spreadsheet workflow

OpenAI’s ChatGPT for Excel and Google Sheets is worth adding here, even though the general availability update landed earlier in May rather than during this exact newsletter window. The reason is simple: for business users, this may be more immediately useful than many of the headline model releases.

The add-in brings ChatGPT directly into Excel and Google Sheets, where users can build, update and understand spreadsheets without copying data into a separate chatbot. Typical use cases include explaining workbook logic, tracing formula chains, cleaning messy data, creating scenario tabs, drafting analysis and finding inconsistencies across sheets.

For finance, strategy, M&A, sales operations and reporting teams, this is a meaningful shift. AI is moving into the tool where the work already happens. The governance point matters too. Spreadsheet AI will save time, but it also creates new review requirements around formula changes, assumptions, source data and auditability. The right trial is not “can it do my job?” It is “can it remove the lowest-value spreadsheet work without breaking the controls?”

MiniMax M3 brings 1M-context frontier models closer to self-hosting

MiniMax has released M3, a new multimodal model aimed at coding, agentic work and long-context tasks. It combines native image and video input with a 1M-token context window powered by MiniMax Sparse Attention, which the company says cuts long-context compute materially. The important angle is deployment flexibility: MiniMax says M3 will be fully open-sourced on Hugging Face and GitHub, supporting private clusters, self-hosting and fine-tuning for controlled enterprise workloads soon.

For business users, the point is not that every company should now host its own model. It is that the competitive set for long-context, agentic workloads is broadening. If M3 performs as advertised, then tasks such as large-codebase analysis, document-heavy diligence, internal knowledge search, video understanding and controlled agent workflows may not have to sit exclusively inside the highest-priced closed-model environments.

Meta puts an AI growth coach inside Facebook creator tools

Meta is embedding an AI assistant directly into Facebook’s creator tools, giving creators personalised recommendations based on their content style, performance, audience and goals. The rollout, starting in the U.S., Canada and India, shifts AI from one-off inspiration prompts into ongoing operational support for creator businesses. For Meta, it deepens platform dependence. For creators, it offers always-on guidance around content strategy, discovery, engagement and audience growth.

Ideogram 4.0 and Reve 2.0 swap text prompts for layout control

On Tuesday 3 June, Ideogram open-sourced Ideogram 4.0, taking the top spot for open image models on Design Arena and trailing only OpenAI and Google’s closed models, with notable strengths in typography and graphic design. Reve launched Reve 2.0 the same day, surpassing Nano Banana 2 on the Text-to-Image Arena leaderboard and trailing only GPT-image-2. Both ship a meaningful UX shift: outputs include labelled segments, so users tweak specific parts of an image without regenerating the whole thing. Reve creates images “like code”, editing by rewriting the layout JSON rather than the prompt. For marketing, design and creative teams that have spent two years re-rolling prompts, this is the first credible move toward agentic, editable visual artefacts.

Quick Hits

• Cybersecurity stocks rally on AI demand, results trail expectations. CrowdStrike and Palo Alto Networks shares are up more than 50% year-to-date on AI-driven cyber-threat demand, but actual top-line growth is barely shifting: CrowdStrike forecast 23% growth for the year, only a point ahead of last year. Netskope shares dropped 20% after-hours on Wednesday after growth slowed to 28% from 32%.

• DeepSeek to raise $7 billion in maiden round. The Chinese frontier-model lab is closing roughly 50 billion yuan in its first ever fundraise, with the founder committing 20 billion yuan of his own money. Fewer than 10 outside investors. Expected to close inside two weeks. Open-weight pricing pressure from Chinese labs is now a funded multi-year campaign.

• Google’s Gemini 3.5 Pro confirmed for June, date pending. Google said at I/O that Gemini 3.5 Pro is already being used internally and will roll out in June, but there is no public model card, API ID or firm date yet. Gemini 3.5 Flash is already available as a stable model. The signal: Google is not sitting out the June frontier-model cycle.

• WindBorne says AI weather model is beating government forecasts. WindBorne Systems’ WeatherMesh-6 claims more frequent and accurate forecasts than ECMWF on key variables, using deep learning and data from around 400 balloons in flight. The broader signal: AI is moving into specialist scientific infrastructure, not just chat, coding and office workflows.

• Google rolls out fake call detection against AI voice scams. Google is launching fake call detection in Phone by Google on Android 12+ devices, starting with Pixel. The feature checks whether a trusted contact’s call is genuine and warns users when spoofing or AI voice impersonation is likely. It is a defensive AI use case against a very practical AI misuse problem.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.
  

Executive Summary

The frontier labs are starting to look less experimental and more like large infrastructure companies. Anthropic has reportedly raised a further $30 billion at a $900 billion-plus valuation, OpenAI is moving towards a possible IPO, and both companies are now tied into enormous long-term compute commitments. For buyers, this matters because vendor risk is changing. The largest AI providers are not disappearing, but they are also becoming more capital-intensive, more concentrated, and more strategically important to their cloud and compute partners.

The second theme of the week is governance. Trump’s decision to pause the planned AI safety executive order leaves US federal oversight more fragmented than expected. At the same time, the Vatican’s first encyclical on AI underlines a broader point: AI governance is no longer just a technical or regulatory issue. It is becoming a question of labour, power, institutional trust and public accountability. Businesses should not wait for governments to define all of this for them. Procurement rules, internal red-team processes, data controls and acceptable-use policies now need to be written at company level.

The third theme is cost. The gap between premium frontier models and cheaper open-weight or lower-cost alternatives is now too large to ignore. Google is pushing Gemini further down the cost curve with Flash. Chinese open-weight models are taking meaningful share on routing platforms. Alibaba is moving its strongest Qwen agent model into a more proprietary lane. The practical point is not that every company should switch model provider. It is that most businesses should stop treating “the best model” as a fixed answer. For high-volume workflows, model routing, cost benchmarking and task-by-task evaluation are becoming basic operating disciplines.

Finally, the product layer continues to move quickly. Figma is putting agents directly into the design canvas. Spotify is turning audio into searchable, summarised knowledge. Google is building agentic search and multimodal shopping flows into its core products. These are not all immediate enterprise priorities, but they show the same direction of travel: AI is moving from standalone chat tools into the software people already use every day.

What to Try This Week

What to Try This Week

1. Run a simple model cost test on one real workflow

Pick one repeatable AI task in the business: summarising calls, drafting reports, classifying support tickets, producing research notes or reviewing documents. Run the same task through your current model, a cheaper fast model, and one open-weight or lower-cost alternative if available. Compare quality, speed and cost. The aim is not to find the “best” model. It is to understand which tasks genuinely need premium models and which do not.

2. Add an AI procurement gate before new tools are bought

Before approving a new AI tool, ask five basic questions: what data does it touch, which model does it depend on, can outputs be checked, who owns the workflow if it fails, and what happens if pricing changes? This is especially important for tools that claim to add “AI” to an existing product but are really just chatbot wrappers with limited control, weak auditability or unclear data handling.

3. Test whether your design, marketing or knowledge workflows are becoming agent-ready

Look at one workflow where staff currently move information between tools: turning a podcast into notes, converting a meeting into actions, creating design variations, or drafting marketing assets from existing material. The useful question is whether AI can now work inside that workflow rather than sitting alongside it. Tools like Figma’s agent and Spotify’s AI audio features show where the market is heading: less copy-paste, more embedded assistance.

Geopolitics, Governance and Big Moves

Anthropic closes $30 billion round at a $900 billion valuation, past OpenAI for the first time

Bloomberg confirmed on Thursday that Anthropic has closed a $30 billion funding round at a $900 billion-plus post-money valuation led by Sequoia, Dragoneer, Greenoaks and Altimeter, putting the company past OpenAI’s $852 billion March valuation for the first time. The Wall Street Journal reported the same week that Anthropic projects a $559 million Q2 operating profit on $10.9 billion of Q2 revenue, up 130% from Q1’s $4.8 billion. The number of customers spending over $1 million annually on Claude doubled from roughly 500 to more than 1,000 between February and April 2026. Compute costs are falling from 71p to 56p per pound of revenue across the same window.

OpenAI files confidential S-1 targeting $852 billion to $1 trillion IPO

OpenAI confidentially filed S-1 paperwork with the SEC on Friday 22 May, with Goldman Sachs and Morgan Stanley leading. Targeted valuation range: $852 billion to over $1 trillion. Targeted listing window: Q4 2026, with a possible September debut. The company is generating roughly $2 billion per month in revenue and counts 50 million consumer subscribers plus 9 million business users. The confidential filing means substantive disclosures stay private until roughly 15 days before the public roadshow, and OpenAI is now subject to SEC quiet-period rules. The three-IPO timeline now reads SpaceX (June), OpenAI (September), Anthropic (October). Prediction markets put 85% probability on OpenAI listing before Anthropic.

Trump cancels AI safety executive order after Musk, Zuckerberg and Sacks calls

President Trump postponed the signing of an AI safety executive order on Thursday 21 May, telling reporters he “didn’t like certain aspects of it”. The order would have established a voluntary mechanism for AI developers to engage with federal agencies and submit advanced models for security review up to 90 days before public release. Axios, the Washington Post and NPR confirmed the same sourcing. Elon Musk, Mark Zuckerberg and David Sacks each called Trump directly on Thursday to argue the framework would damage US competitiveness against China. The signing ceremony invitations had already been issued. The substantive US-side oversight surface for the next 12 months is now CAISI’s voluntary red-team partnerships plus state-level AI legislation.

Pope Leo XIV publishes Magnifica Humanitas, the first papal encyclical on AI

The Vatican released Pope Leo XIV’s first encyclical, Magnifica Humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence, on Monday 25 May, signed by the Pope on 15 May (the 135th anniversary of Pope Leo XIII’s Rerum Novarum). Anthropic co-founder Christopher Olah attended the presentation. The encyclical is structured around four moral commitments: human dignity, labour, the common good, and what Leo calls “constructing Jerusalem rather than Babel”. The most-cited passage warns against “a race for ever more powerful algorithms and larger datasets driven by the desire to secure geopolitical or commercial dominance”. The document explicitly condemns AI in warfare, calling “the ‘just war’ theory which has all too often been used to justify any kind of war, now outdated”.

Anthropic Q2 profitability: $559 million on $10.9 billion of revenue

The Wall Street Journal’s internal-projections scoop on Wednesday 20 May is the most important data point on frontier-lab economics published in 2026. Anthropic told investors it expects a $559 million Q2 operating profit on $10.9 billion of Q2 revenue, up 130% from Q1’s $4.8 billion. The number of customers spending over $1 million annually on Claude doubled from roughly 500 to more than 1,000 between February and April. Compute costs per pound of revenue fell from 71p to 56p across the same window. Dario Amodei acknowledged at an internal event that the company “planned for 10x annual growth but saw 80x in Q1”. The cautious read is that the operating profit is being booked while $45 billion of forward compute spend has just been contractually locked in. The leading indicator (large-customer-cohort doubling in 60 days) is the part that should reshape any enterprise-AI procurement assumption built before the spring.

Chinese open-weights models hit 60% OpenRouter share at one-third Claude pricing

The CNBC investigation on Friday 22 May, sourcing Artificial Analysis benchmarks, put hard pricing numbers on the open-weights gap. A benchmark set that costs $4,811 to run on Claude Opus 4.7 runs for $1,071 on DeepSeek V4 and $544 on Zhipu GLM-5.1. Chinese open-weights models (Kimi K2.6, DeepSeek V4, GLM-5.1, Qwen 3) now account for approximately 60% of OpenRouter token volume, up from 1% in early 2024, with per-million-token pricing in the $0.28 to $0.95 range against Claude’s $3-plus. Databricks CEO Ali Ghodsi added the operational angle: open-weights routing now defaults to “whichever Chinese model is closest to my latency target this hour” for the cost-sensitive 80% of agentic workloads.

Financial Times: open-weight guardrails stripped from Meta and Google in minutes

The Financial Times investigation, published on Monday 25 May and carried by the Irish Times, showed that the Heretic tool on GitHub can strip safety guardrails from Meta’s Llama 3.3 and Google’s Gemma 3 in under 10 minutes with no specialist hardware. The technique, “abliteration”, removes the safety post-training layer. The FT working with AI safety group Alice produced a modified Gemma 3 that responded to prompts on how to disperse chlorine gas through a crowded indoor space, generated working credit-card-theft code, and wrote child-sexual-abuse stories. Heretic’s creator Philipp Emanuel Weidmann told the FT his software has been used to produce more than 3,500 “decensored” models since release last year, with the modified variants downloaded 13 million times. The story landed two days after Trump killed the federal pre-release review framework.

Karpathy joins Anthropic to use Claude to accelerate Claude pretraining

Andrej Karpathy, OpenAI co-founder and former Tesla Autopilot lead, joined Anthropic’s pretraining team on Tuesday 19 May. Karpathy will work under team lead Nick Joseph to build a team that uses Claude itself to accelerate pretraining research. His framing on X: “I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D.” The strategic frame Anthropic is pushing in TechCrunch, CNBC and Axios is that AI-accelerated research is now a more important competitive differentiator than raw compute capacity. The senior AI-research talent pool is genuinely contested across at least four destinations (Anthropic, Thinking Machines, Safe Superintelligence, Meta Superintelligence Labs).

Anthropic’s Mythos rollout moves from US labs to Japanese finance

Japan’s planned access to Claude Mythos shows how advanced models are moving from lab capability to regulated financial infrastructure. Reported access for MUFG, SMBC and Mizuho follows government and regulator discussions over cybersecurity risk, especially Mythos’s ability to find software vulnerabilities. The notable point is governance: Anthropic is not broadly releasing the model, but selectively deploying it through public-private channels where cyber upside and misuse risk both matter commercially materially.

Tech, Tools and Releases

Google I/O 2026 ships Gemini 3.5 Flash and an agent-first Search

Google’s I/O 2026 keynoteac anchored on Gemini 3.5 Flash, generally available via Google Antigravity, the Gemini API and Android Studio at $1.50 input / $9 output per million tokens. Google is claiming a 12x speed advantage versus comparable frontier models inside the Antigravity 2.0 agent harness. The model leads on coding and agentic-task benchmarks but trails on knowledge-breadth evaluations against Claude Opus 4.7 and GPT-5.5. Strategically, Google is choosing distribution over benchmarks: Gemini 3.5 Flash is now the default model in AI Mode in Search globally, with generative UI (custom interactive visuals, tables, graphs and simulations) rolling out through summer. Robby Stein, VP of Product for Search, framed it as the end of the “10 blue links era”.

Spotify pushes deeper into AI-generated audio and knowledge tools

• Spotify introduced AI-powered podcast Q&A, summaries and personalised listening briefs, letting users interrogate long-form shows rather than simply listen to them. The move turns podcasts into searchable, reusable knowledge assets.

• Spotify launched an ElevenLabs-powered audiobook creation tool for authors and publishers, using synthetic voices to reduce the cost and time involved in producing long-form audio content.

• Spotify also rolled out a NotebookLM-style app that lets users ingest content and generate AI-based notes and insights, extending its ambitions beyond streaming into AI-assisted research and learning.

Google announces Gemini Omni, Google Pics and Universal Cart.

Omni is the “any modality in, any modality out” model starting with video; Pics is generative image creation and editing built on Nano Banana; Universal Cart follows the user across Search, Gemini, YouTube and Gmail, putting Google directly into the Alexa for Shopping competitive lane.

Figma turns the design canvas into an AI workspace

Figma’s new in-canvas agent can generate design layers, explore multiple directions, apply feedback, make bulk edits and stay aligned with a team’s components, tokens and libraries. Alongside custom Make skills and improved design-to-code workflows, the update moves Figma beyond prompt-based prototyping into AI-assisted product development. For teams, the useful shift is consistency: agents can now work inside existing design systems, not around them, while preserving designer control over output quality.

Google pushes Gemini down the cost curve

Google’s Gemini 3.5 Flash shows how the model market is shifting from “biggest model wins” to price-performance. Google says Flash beats Gemini 3.1 Pro across most benchmarks while running four times faster, with broad availability across Antigravity, AI Studio, Android Studio, Gemini Enterprise and Search. For businesses, the implication is simple: more agentic and multimodal workloads are becoming viable at production scale, not just in demos.

Alibaba makes Qwen’s strongest agent model proprietary

Alibaba’s Qwen team has released Qwen3.7-Max, a proprietary model built for long-running agent work, with claimed leads on Terminal-Bench 2.0-Terminus, SWE-Pro, SciCode and several general-agent benchmarks. The important shift is strategic as much as technical: Qwen remains influential in open-weight development, but Alibaba now appears to be reserving its strongest agent and coding capabilities for closed API-based models, pushing Qwen directly into Claude, Gemini and GPT territory.

Quick Hits

• Pentagon awards classified-network AI contracts to seven vendors. OpenAI, Google, Microsoft, AWS, Nvidia, SpaceX and Reflection AI replace the $200 million July 2025 Anthropic agreement. Anthropic was formally excluded after declining the “all lawful purposes” clause on domestic surveillance and autonomous-weapons grounds. Defence Secretary Pete Hegseth’s March “supply chain risk” designation is the contractual mechanism.

• Sam Altman walks back the AI jobs apocalypse in Sydney. Speaking at a Commonwealth Bank of Australia conference on Tuesday 26 May, Altman said “I’m delighted to be wrong... I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened”. The walk-back retracts his earlier “30 to 40% of work tasks” framing.

• Anthropic policy paper narrows the US lead. The paper reframes the US-vs-China model-capability gap as “several months”, not the previously-assumed year, formalising the working assumption for capability-elicitation timelines across the policy stack.

• Anthropic Coordinated Vulnerability Disclosure dashboard updated 22 May. 1,596 vulnerabilities disclosed across 281 open-source projects, with 97 patched to date. The dashboard is the cleanest public artefact of the Claude Mythos and Project Glasswing programme.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.
  

Executive Summary

Another busy week in the world of AI - if i was to give the week a theme it would be “Agents”.

The first shift was commercial. Ramp’s latest AI Index showed Anthropic overtaking OpenAI on paid business adoption for the first time, with Anthropic at 34.4% of paying businesses versus OpenAI at 32.3%. That does not mean OpenAI is losing the market, but it does mean the default procurement assumption has changed. Anthropic is no longer the “technical alternative”; it is becoming a mainstream enterprise option, helped by stronger business adoption, a reported $30 billion annualised run-rate and the launch of Claude for Small Business.

The second shift was infrastructure. Nvidia’s latest results suggested AI demand is still compounding rather than plateauing, while Cerebras’ IPO performance and inference-speed claims show how aggressively investors and customers are looking for alternatives to the GPU-heavy status quo. The message is not that Nvidia is about to be displaced. It is that AI infrastructure is now a strategic market in its own right, with speed, cost and availability becoming board-level questions.

The third shift was product architecture. Google I/O pushed Gemini deeper into Search, Android, Workspace, Chrome and developer tools, while Cursor, Amazon, Apple, Microsoft and Cerebras all pointed in the same direction: AI is moving (at pace) from chat windows into workflow layers. The practical implication for software vendors is uncomfortable. Lightweight AI features bolted onto legacy products will be harder to defend if larger platforms can place capable agents directly into the tools customers already use.

For business leaders, the takeaway is simple: re-run your AI vendor assumptions, test where agents could replace manual workflows, and be sceptical of product roadmaps that treat AI as a feature rather than a new operating layer.

What to Try This Week

What to Try This Week

Run a one-hour “agent exposure” review of one workflow in your business.

Pick a process that still involves people moving information between systems: preparing a sales proposal, updating a CRM, reconciling invoices, triaging support tickets, summarising customer calls, drafting board materials or checking supplier risk.

Then ask three questions:

1. Which steps require judgement, and which are just retrieval, formatting, checking or routing?

2. Which systems does the person need to access to complete the task?

3. If an AI agent could safely read those systems and take limited actions, what would still need human approval?

The aim is not to automate the workflow this week. The aim is to understand where the boundary sits between “AI as assistant” and “AI as operator”.

For most businesses, the opportunity will not start with replacing a whole job. It will start with removing the repetitive hand-offs inside a job. That is why this week’s Google, Anthropic, Microsoft, Cursor and Amazon announcements matter: the agent layer is moving closer to the tools your team already uses.

Geopolitics, Governance and Big Moves

Anthropic overtakes OpenAI on business AI adoption

Ramp’s April AI Index showed Anthropic taking the paid business-adoption lead from OpenAI for the first time, 34.4% against 32.3%, with Anthropic’s share rising 3.8% in April while OpenAI’s fell 2.9%. Across twelve months Anthropic quadrupled its share while OpenAI’s grew 0.3%. Ramp economist Ara Kharazian called it “a stunning reversal in the competitive market dynamics for AI model providers”. The data tracks 50,000-plus US firms’ card-and-invoice spend, so big transformation deals are not in the sample, but the trend break is unambiguous.

Cerebras Day 1 closes up 68% at a $94 billion market cap

Cerebras’ first day of public trading on Friday closed at $311.07, up 68% from the $185 IPO price set 24 hours earlier, putting the company at roughly $94 billion in market value, more than 2.3 times its $40 billion IPO valuation. The IPO raised $5.55 billion at 20 times oversubscription, making it the biggest US IPO of 2026 so far. The forward multiple is rich: Cerebras is projected to roughly quadruple revenue from $800 million this year to $3.2 billion next year against heavy OpenAI customer concentration. Investors paid the premium regardless. The public-market filter on AI-infrastructure messiness has plainly weakened.

Nvidia’s results say the AI infrastructure boom is still accelerating

Nvidia’s latest results landed like another stress test for the “AI bubble” argument. Q1 FY2027 revenue hit $81.6 billion, up 85% year on year, with data centre revenue reaching $75.2 billion. The bigger signal was forward-looking: $91 billion of Q2 guidance, a higher dividend and an $80 billion buyback. Demand for AI infrastructure is not plateauing yet; it is still compounding.

Karpathy joins Anthropic’s frontier research team

Andrej Karpathy, OpenAI co-founder and former Tesla AI lead, has joined Anthropic’s pre-training team. The move is significant because pre-training is where the core capability of frontier models is built, and Anthropic says Karpathy will work on using Claude to accelerate that research. It is another signal that Anthropic is not just winning enterprise attention; it is also attracting the research talent needed to keep pushing Claude forward.

OpenAI fights governance pressure on three fronts at once

Three escalations landed inside a single 24-hour window on Thursday. The Information’s Aaron Tilley reported OpenAI’s extraordinary decision to issue a public legal threat against Apple as both companies prepare to compete in AI hardware. A fresh court filing in the Musk v Altman trial revealed Sam Altman holds more than $2 billion in personal stakes in companies that have done business with OpenAI, fresh material for the House Oversight investigation. And Microsoft executive Michael Wetter testified that Microsoft’s commercial commitments to OpenAI will exceed $100 billion by the end of its June fiscal year. Martin Peers’ read: the “never-ending series of dramas” is now hurting OpenAI’s standing as a reliable enterprise partner.

OpenAI clears Musk lawsuit, but the governance questions remain

A California jury unanimously rejected Elon Musk’s lawsuit against OpenAI, finding he waited too long to bring the case. The ruling removes a major legal overhang as OpenAI prepares for a possible IPO. Musk plans to appeal, so the dispute is not finished, but OpenAI has won the immediate round.

Microsoft’s $100 billion-plus OpenAI exposure now publicly quantified

Wetter’s Thursday testimony nailed down a figure that had previously been triangulated. Microsoft’s commercial commitments to OpenAI will exceed $100 billion by the end of its June fiscal year: the $13 billion equity stake plus training and data-centre infrastructure. Read alongside Microsoft’s $30 billion of OpenAI-related revenue between 2023 and 2025, the relationship is recalibrating from joint venture to a highly profitable but increasingly narrow commercial partnership. With Microsoft shares down 16% year-to-date and TCI having exited its $8 billion stake, activist pressure is rising. Peers’ parallel to ValueAct’s 2013 playbook, which culminated in Steve Ballmer’s exit and Satya Nadella’s appointment, looks more plausible by the week.

Microsoft MDASH beats Anthropic’s Mythos on the CyberGym benchmark

Microsoft unveiled MDASH on Thursday: a multi-agent AI security harness chaining more than 100 specialised agents in a scan, debate and validate pipeline. One set of agents scans code for potential vulnerabilities, a separate group debates whether each finding is real and exploitable, and a final stage builds proof-of-concept attacks. In its initial deployment MDASH found 16 flaws across Windows and surpassed Anthropic’s Claude Mythos on the CyberGym benchmark. This is the first non-frontier-lab system to win that benchmark and the cleanest signal yet that multi-agent architectures with explicit debate-then-validate steps outperform monolithic frontier models on offensive-cyber tasks. Britain’s AI Safety Institute confirmed the same day that AI cyber capability is doubling every 4.7 months.

Tech, Tools and Releases

Claude for Small Business ships an SMB-ready package

Anthropic launched Claude for Small Business on Thursday, a bundle of connectors and ready-to-run workflows embedding Claude inside QuickBooks, PayPal, HubSpot, Google Workspace and Microsoft 365, plus new tools for payroll, invoices and campaigns. Anthropic was traditionally the developer-tooling specialist; pairing this with the same-day Ramp data showing Anthropic ahead of OpenAI on business adoption signals a deliberate push into the operator tier where ChatGPT had been dominant. Paid Claude plans will also receive a dedicated monthly credit for programmatic usage starting 15 June, packaging API usage into existing subscription tiers.

Google I/O turns Gemini into an agentic operating layer

Google I/O was not just another model launch. It was Google pushing Gemini deeper into the surfaces where work already happens: Search, Android, Workspace, Chrome, developer tools and, eventually, smart glasses. Gemini 3.5 Flash is being positioned for faster agentic and coding workloads, while Gemini Omni expands Google’s multimodal ambitions. The bigger signal is Antigravity and Managed Agents, which make it easier for developers to build, orchestrate and deploy agents that can use tools, run code and complete multi-step tasks. For software vendors, the warning is clear: lightweight AI features bolted onto legacy products will be harder to defend.

Adaption’s AutoScientist automates frontier-style fine-tuning

Adaption, the startup founded by former Cohere VP of Research Sara Hooker, introduced AutoScientist on Thursday: a system that automatically customises AI models for specific jobs by adjusting both training data and training settings, iterating until the model meets the goal. In internal tests AutoScientist outperformed expert-tuned models by 35% on average, with success rates jumping from 48% to 64% across finance, legal, medical and five other industries. The strategic point: only a few thousand people globally can properly fine-tune a frontier model, and nearly all of them work at the same handful of labs. If that expertise can be automated, custom per-business models become practical to build at scale.

Amazon’s Alexa for Shopping replaces Rufus and takes over the search bar

Amazon’s full Alexa for Shopping rollout details landed on Thursday across The Rundown, TLDR and Martin Peers’ Briefing. The product replaces Rufus, takes over the Amazon.com search bar, and draws on Amazon’s catalogue, reviews, delivery timing, past purchases and Alexa voice conversations to answer questions, run side-by-side comparisons, track pricing and Auto-Buy items when prices hit a target. A new Buy for Me flow handles checkouts on non-Amazon stores; Scheduled Actions auto-restocks products on a cadence. For UK retailers, the open question is what the SEO and product-catalogue playbook looks like when most shopping queries are agent-mediated.

Apple readies AI agents in the App Store ahead of WWDC

The Information’s Aaron Tilley reported on Thursday that Apple is exploring how to welcome AI agents into the App Store while preserving its privacy and security positioning. The piece described concrete policy work on letting AI agents follow Apple’s privacy rules and on preventing agents from spinning up smaller apps after the parent app has been approved. The companion announcement: iOS 27 will feature a redesigned Siri, a customisable Camera app for professionals and system-wide design changes, set for WWDC on 8 June. Read alongside the OpenAI-Apple legal flare-up the same week, Apple is opening App Store policy for AI agents while signalling its native AI rebuild will compete with whatever OpenAI ships in hardware.

Cerebras shows why AI speed matters

Cerebras says it has set a new inference-speed record for Moonshot AI’s Kimi K2.6, a huge open-weight model with around a trillion parameters.

In plain English: inference is the part where an AI model gives you an answer, and tokens are the chunks of text it generates. Cerebras reportedly reached 981 output tokens per second, meaning the model could respond far faster than the official version. That is news worthy because businesses do not just want smarter AI; they want AI that can work quickly enough to power real-time apps, agents and customer-facing tools. It also strengthens Cerebras’ argument that its wafer-scale chips offer a serious alternative architecture to traditional GPU-heavy AI infrastructure.

Cursor’s Composer 2.5 shows the cost curve for AI coding is falling

Cursor has released Composer 2.5, its upgraded in-house coding model, built on Moonshot’s open-weight Kimi K2.5 checkpoint. The claim is not just better benchmarks; it is better long-running coding behaviour at far lower cost. Cursor says pricing remains $0.50 per million input tokens and $2.50 per million output tokens, making near-frontier coding performance increasingly accessible to ordinary development teams.

Quick Hits

• Anthropic CFO confirms $30 billion run-rate. Krishna Rao’s Thursday interview confirmed Anthropic at a $30 billion annualised run-rate, up from $250 million two years ago, on $75 billion raised in his tenure. WSJ separately reported investment offers arriving at $900 billion-plus valuations.

• Vercel AI Gateway production trends. 200,000-plus teams now route through the Gateway, with agentic workloads carrying 59% of all token volume. Anthropic leads in spend, Google leads in volume, and open-source models are gaining traction with no lab-loyalty.

• Microsoft hits $5.5 trillion through Nvidia’s market cap milestone. Nvidia became the first company ever to reach $5.5 trillion in market capitalisation, with the milestone coinciding with Jensen Huang’s arrival in China for Trump’s meetings with Xi Jinping. Nvidia H200 China-export delivery, meanwhile, remains at zero.

• Cisco and LinkedIn announce layoffs into AI tailwinds. Cisco cut around 4,000 jobs, roughly 5% of its workforce, “as it shifts more spending toward AI and cybersecurity” despite 12% revenue growth. LinkedIn separately laid off 5% of its 17,500-plus global staff. Layoffs.fyi is tracking 103,000-plus tech layoffs across 2026.

• ChatGPT Trusted Contact launches. OpenAI shipped a feature that lets users nominate one adult to be notified if a ChatGPT conversation suggests a potential self-harm concern. Five-step setup in Settings, rolling out worldwide within weeks. Worth raising in a team wellbeing channel.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.
  

Executive Summary

This was the week the AI economy started to look less like a gold rush and more like a stress test. Anthropic disclosed extraordinary revenue growth, OpenAI reportedly renegotiated billions of dollars out of its Microsoft revenue-share obligations, Cerebras pushed toward a richly valued IPO, and CoreWeave reported a level of capex that shows just how expensive the AI infrastructure race has become. Money is still moving quickly, but the questions around durability, margins and who ultimately pays are getting sharper.

Underneath the financial noise, three practical themes stood out. First, AI is now part of the cyber threat landscape, with Google documenting the first known case of AI being used to discover and write a zero-day exploit, while OpenAI launched its own cyber-defence programme. Second, AI productivity is starting to show up in workforce planning, with Monday.com joining a growing list of software companies linking headcount discipline to AI-enabled efficiency. Third, AI is becoming a real cost line, not just a software experiment, with usage-based pricing and AI compute costs beginning to appear in company commentary and customer invoices.

The practical takeaway this week is simple: AI is moving out of the innovation budget and into operating reality. If you have not yet mapped which AI tools your business is paying for, how they are priced, and where usage-based costs could compound, now is the time. That is the focus of this week’s practical exercise below.

What to Try This Week

What to Try This Week

Audit your AI usage-based pricing exposure before your next finance review. Three things this week made clear that AI is no longer just a tooling decision: it is a budget line that surprises CFOs. Monday.com flagged AI compute costs hitting gross margin. Figma reports 75% of paying customers consuming AI credits weekly. Software firms across the board are now charging per use, not per seat.

Concrete steps any operations or finance leader can take this week:

1. List every AI tool your business pays for, including embedded features inside existing software (Microsoft Copilot, Salesforce Einstein, HubSpot AI, Notion AI, Figma AI, Adobe Firefly). You are likely to miss two or three on the first pass.

2. For each, identify the pricing model: flat seat, credit-based, token-based, or hybrid. Credit and token models are where the bill compounds invisibly.

3. Pull last quarter’s actual usage versus the included allowance. Anything above 80% utilisation is a renewal-cycle problem already.

4. Set a monthly cap or alerting threshold on every usage-based product. Most platforms have this buried in admin settings; it is rarely on by default.

5. Brief your finance team on the new line item. “AI cost of revenue” is now a thing that shows up on quarterly calls. Better for them to hear it from you than from your audit.

This is a half-day exercise for most SMEs and a one-day exercise for mid-market. Worth doing this week before the next purchase order or auto-renewal goes through unexamined.

Geopolitics, Governance and Big Moves

Anthropic discloses 80x revenue growth at $44B annualised

CEO Dario Amodei told staff that Q1 2026 revenue grew roughly 80x year-on-year, pushing annualised recurring revenue above $44B by early May, up from $9B at year-end 2025. Customers spending more than $100K a year are up 7x; those over $1M a year are rising sharply. Claude Code and a new financial-services agent product (built with JPMorgan for pitchbooks and credit memos) are the two named drivers. For context, Salesforce took roughly 20 years to reach $30B in annual revenue. Anthropic did it in under three years from a standing start.

Cerebras prices IPO at $35B valuation with 20x oversubscription

AI chip designer Cerebras Systems is expected to debut on Nasdaq on Thursday 14 May at $150 to $160 a share, up sharply from an initial range of $115 to $125. At the top end this implies a $35B valuation and as much as $4.8B raised. Demand is reportedly 20x oversubscribed. The company posted $510M of 2025 revenue, anchored by a 750MW compute deal with OpenAI. The IPO is the second major AI infrastructure listing this cycle after CoreWeave; both have benefited from public-market appetite ignoring the cash burn beneath. Investors are clearly comfortable paying for AI exposure rather than profits.

OpenAI restructures Microsoft deal, saving $97B through 2030

A previously undisclosed amendment to the OpenAI-Microsoft revenue-sharing agreement will save OpenAI as much as $97B in payments to Microsoft through 2030, per The Information. The original 20% revenue share could have totalled $135B if OpenAI hit its long-term revenue goals. The new structure leaves OpenAI on the hook for “only a fraction” of that. Microsoft, separately, has already booked more than $30B in revenue from OpenAI’s technology. The pair are recalibrating economics ahead of OpenAI’s possible $1T IPO and Microsoft’s worsening investor narrative (see below).

Pentagon signs $200M AI deals with eight firms, Anthropic excluded

The US Department of Defense signed agreements with SpaceX, OpenAI, Google, Microsoft, Nvidia, AWS, Oracle and Reflection for frontier AI capabilities. Anthropic was conspicuously omitted, despite being the largest single revenue generator in the sector. Reporting points to ongoing tensions between Anthropic and the administration over the firm’s safety stance and its restrictions on weapons-related model use. For UK and European defence primes watching the US, this is the clearest signal yet that frontier-AI procurement is becoming a political instrument. Build your supplier list with the political alignment of frontier labs in mind, not just the technical capability.

CoreWeave’s Q1 capex hits $7.7B, three times revenue

AI cloud provider CoreWeave doubled Q1 revenue to $2B year-on-year but spent $7.7B on data-centre capex in the same quarter, burning $4.7B in cash. Full-year 2026 guidance is $12-13B of revenue against $35B of capex, three times revenue. The company ended the quarter with $24.8B of debt and roughly $3B of cash, two-thirds of which came from a fresh Nvidia equity injection. Anthropic, Microsoft, OpenAI and Meta are the customer underwriters. Shares fell 15% across the day. The trade is binary: if hyperscaler AI commitments hold, CoreWeave is undervalued. If they pull back, it cannot service its debt.

Microsoft becomes the activist target, down 16% YTD

Microsoft shares are down nearly 16% year-to-date, making it the worst-performing US big tech name in 2026. UK hedge fund TCI sold “almost all” of its $8B stake, citing concern that AI threatens the Office franchise. The narrative pressure has reopened activist conversations: with SpaceX’s IPO weeks away pulling capital out of incumbents, Martin Peers has suggested a streamlining play (carving off gaming, Bing, LinkedIn) is overdue. Azure is still growing close to 40% and Copilot now sits above 20M paid seats. The fundamentals are stronger than the share price; the marketing is not.

Tech, Tools and Releases

Thinking Machines Lab debuts “interaction models”

Mira Murati’s lab broke its post-OpenAI silence with a research preview of interaction models: AI systems built for live human-AI collaboration across voice, video and text in 200-millisecond chunks. Unlike agentic systems that run autonomously and report back, interaction models are designed to be steered, interrupted and corrected in real time, with a background reasoning model handling slower work concurrently. Use cases shown include live translation, exercise coaching, and timed interventions. This is a bet that the next leap in AI is not longer autonomy, but better collaboration. Worth watching for customer-service, sales-enablement and training applications.

Google traces first AI-written zero-day exploit in the wild

Google’s Threat Intelligence Group confirmed the first publicly documented case of a hacker using AI to discover and write a zero-day vulnerability, against a widely used web management tool. Google flagged unusually polished attack code, lengthy explanation comments, and a fabricated severity score as the giveaways. Anthropic’s Rob Bair said defenders’ lead is “months, not years”. For UK ops and security leaders: the offensive use of AI is no longer theoretical or research-grade. Patch cadence, identity hygiene and two-factor authentication enforcement have a sharper urgency this quarter than last.

OpenAI launches Daybreak to rival Anthropic Mythos

OpenAI introduced Daybreak, a cyber-defence platform combining a new GPT-5.5-Cyber model with the Codex Security agent. The pitch: turn vulnerability discovery and audit-ready patching from weeks of manual work into minutes. Cloudflare, Cisco and Palo Alto Networks are launch partners. Access to GPT-5.5-Cyber is restricted to vetted “trusted defenders” running critical infrastructure who can clear strict approval by 1 June. This sits directly opposite Anthropic’s Mythos programme, which used the same offensive capability to help Mozilla find 271 Firefox vulnerabilities. The frontier labs are now selling the cure alongside the disease.

Claude Agents Learn to “Dream” Between Tasks

Anthropic has moved Claude’s “Dreaming” feature into research preview, letting managed agents review past sessions, spot recurring patterns and tidy their memory stores between tasks. The output is a separate, reviewable memory file, not a change to the model itself. The direction matters: enterprise agents are shifting from one-off assistants toward systems that learn operational preferences over time, with clearer auditability than opaque retraining loops for users and IT teams.

Elon Musk dissolves xAI into SpaceXAI division

Elon Musk announced that xAI will dissolve and be absorbed into SpaceX as a new internal division called SpaceXAI, consolidating Grok, X, and AI compute under a single entity ahead of the SpaceX IPO expected this summer. The reorganisation aligns Musk’s AI stack with his most valuable private asset and creates the simplest possible structure for institutional investors. For practical purposes: anyone with embedded Grok API integrations should expect rebranding, contract restructuring and potentially deprecation warnings over the next quarter.

Monday.com freezes headcount, blames AI productivity

Enterprise SaaS firm Monday.com told investors it expects workforce to stay “largely flat” through 2026, “reflecting the productivity gains AI is already delivering across our organisation”, after previously guiding 20% headcount growth. Shopify, Spotify, Pinterest and Uber have all said something similar in recent quarters. The pattern: companies are not yet conducting mass layoffs; they are simply not backfilling or expanding. Crucially, Monday.com also flagged rising “computing costs related to AI” hitting gross margin. The productivity is real, but it is not free. AI-driven hiring restraint is the slow-burn signal worth tracking; the unemployment numbers will lag it by quarters.

ElevenLabs deepens enterprise traction with Klarna, Revolut

Voice AI specialist ElevenLabs disclosed enterprise contracts signed in the past quarter with Deutsche Telekom, Revolut and Klarna, alongside a $500M Series D with BlackRock, Wellington, D.E. Shaw, Schroders, Nvidia and Salesforce Ventures among new investors. Crucially, this is voice AI moving out of content-creator tooling and into customer-facing operations: support agents, call-centre automation, localisation. For UK SME leaders running multi-language operations or volume customer service, voice AI is now production-ready in a way that screen-readers and IVR menus stopped being years ago.

Quick Hits

• OpenAI launches Deployment Co with $4B private equity backing: new venture acquired AI consulting firm Tomoro and adds ~150 consultants targeting enterprise AI implementation, alongside Anthropic’s parallel $1.5B JV with Blackstone, Hellman & Friedman and Goldman Sachs.

• OpenAI’s $18B Broadcom chip deal hits financing snag: Broadcom will only fund first-phase production if Microsoft commits to buying 40% of the chips; Microsoft has not yet signed.

• SAP signs Anthropic agreement: new partnership lets SAP customers build agents using Claude through a deeper product integration, a meaningful enterprise win against Microsoft and OpenAI on the SAP installed base.

• House Oversight Committee opens investigation into Sam Altman’s personal investments: scrutinising ties to OpenAI’s commercial partnerships, adding to legal and governance pressure during Musk v. Altman trial.

• Google "Omni" model leaked ahead of I/O 2026 - A UI string spotted in Gemini's video generation tab reads "powered by Omni," sparking speculation that Google will unveil a unified multimodal architecture (text, image, video, audio, reasoning) at Google I/O on May 19–20. Analysts frame this as Google's answer to GPT's unified model strategy.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.
  

Executive Summary

Two themes emerged this week.

First, AI oversight is back on the agenda in Washington. After dismantling the previous administration’s AI safety framework earlier this year, the White House is now reportedly drafting an executive order that would require pre-release government vetting of frontier AI models. Microsoft, Google DeepMind and xAI also signed pre-deployment evaluation agreements with the Commerce Department’s Center for AI Standards and Innovation (CAISI) on 5 May, joining OpenAI and Anthropic. The direction of travel is clear: the most capable models are no longer being treated as ordinary software releases. They are becoming strategic infrastructure.

Second, evidence continues to grow, showing that AI adoption is moving from experimentation into operational redesign. Coinbase confirmed a workforce reduction on 5 May and directly linked the move to AI deployment across customer service, fraud and compliance. PayPal has also announced a multi-year restructuring programme alongside the creation of an internal AI Transformation Group. The important point for business leaders is not that AI automatically means layoffs. It is that investors and management teams are now starting to connect AI deployment with measurable changes to cost base, service delivery and operating leverage.

For businesses, the practical takeaway is this: AI is no longer just a productivity tool sitting at the edge of the organisation. It is beginning to reshape workflows, team structures and expectations around speed, cost and responsiveness. The companies that benefit will not be the ones that simply buy more tools. They will be the ones that understand where AI can genuinely improve repeatable work, where human judgement still matters, and where the operating model needs to change around the technology.

What to Try This Week

What to Try This Week

Run a 30-minute workflow audit with your senior team.

Pick two or three repeatable workflows where your business already spends significant time, money or management attention. Avoid the obvious AI use cases such as basic content generation. Look instead at areas such as compliance review, contract analysis, financial reporting, account research, sales operations, regulatory monitoring, customer onboarding or management reporting.

For each workflow, answer four questions:

1. What work is genuinely repeatable?

2. What data, documents or systems does the work depend on?

3. Where does human judgement add real value?

4. What could AI realistically improve today: speed, cost, accuracy, consistency, customer experience or management visibility?

The goal is not to create a redundancy plan. It is to build a clearer view of where AI could improve the business in practical terms, and where the organisation is not yet ready because the data, process or governance is too weak.

This is the useful lesson from this week’s news. AI is starting to show up in operating models, not just product demos. Businesses should respond by mapping the workflows where the technology could create measurable value, then deciding which ones are worth testing properly.

Trump Administration Reverses Course on AI Oversight

The White House is drafting an executive order for mandatory pre-release government vetting of frontier AI models, a sharp reversal of its January position. The proposal mirrors the UK’s pre-deployment safety regime and was triggered specifically by Anthropic’s Claude Mythos cyber capabilities. No timeline has been published. Industry reaction is split between welcoming the policy clarity and concern about chilling effects on US frontier development. Sources: New York Times (4 May), Wall Street Journal, Politico, CIO.

Pentagon Signs Classified-Network AI Deals with Seven Firms; Anthropic Excluded

On 1 May, the Department of Defense confirmed agreements with SpaceX, OpenAI, Google, Microsoft, AWS, Nvidia and startup Reflection AI to deploy LLMs on classified networks for “any lawful purpose.” Anthropic was excluded after refusing the unrestricted-use clause. The fine print allows the government to require adjustments to safety filters, with no contractual veto for the vendor. Constitutional and congressional scrutiny appears likely. Sources: New York Times, Associated Press, Tom’s Hardware.

NYT: Federal Safety Net “Not Ready” for AI Job Displacement

On 5 May, the New York Times argued that economists across the political spectrum agree the US federal safety net is structurally unprepared for AI-driven displacement at the scale now measurable. The piece cited Goldman Sachs research showing net 16,000 US jobs lost per month to AI substitution, with Gen Z and entry-level workers bearing the brunt. The widening entry-level wage gap is now empirically quantified, not speculative. Source: New York Times.

AMD Q1 Earnings Reset the AI Infrastructure Trade

On 5 May, AMD reported Q1 revenue of $10.3 billion (up 38% year on year), with data centre at $5.8 billion (up 57%). Q2 guidance of $11.2 billion came in well above the $10.5 billion consensus. The stock rose 13 to 16% after hours and triggered a sector-wide rally in Micron, Intel and Arista. Morgan Stanley revised its price target up the same day. The print broadens the AI infrastructure thesis beyond Nvidia in a meaningful way. Sources: Reuters, CNBC, AMD IR, EBC.

Coinbase and PayPal Cut Workforces, Cite AI Directly

On 5 May, Coinbase confirmed approximately 700 layoffs (14% of headcount), explicitly attributing the cuts to AI-driven consolidation across customer service, fraud detection and compliance. The stock rose 3.5% on the announcement. PayPal CEO Enrique Lores announced a 20% reduction over two to three years and stood up an internal AI Transformation Group led by former Walmart technology executive Anshu Bhardwaj. Investors are now treating AI-led restructuring as a margin event. Sources: New York Times, Fortune, SFGate, Bloomberg, HeyGoTrade.

Sierra Raises $950M; Cognition in Talks at $25B

Bret Taylor’s enterprise AI customer-service startup Sierra closed a $950 million Series E on 4 May, led by Tiger Global and Google’s GV with Benchmark, Sequoia and Greenoaks participating. Valuation: $15.8 billion. Separately, on 6 May Bloomberg reported that Cognition AI (maker of Devin) is in early talks at a $25 billion valuation, 2.5 times its September 2025 mark. Agentic enterprise software is now in price-discovery mode. Sources: TechCrunch, CNBC, Bloomberg, Techzine.

Anthropic Finalises Approximately $1.5B Wall Street Joint Venture

On 4 May, Anthropic confirmed a joint venture backed by Blackstone, Hellman & Friedman, Goldman Sachs and General Atlantic to embed Claude across private equity portfolio companies. The structure mirrors Palantir’s forward-deployed engineer model and positions Claude as enterprise infrastructure for a substantial share of the global buyout market. Reported total commitment is approximately $1.5 billion. This is a distribution play, not a product sale, and a significant strategic move. Sources: Reuters, Wall Street Journal, GuruFocus, Investing.com.

OpenAI Raises $4B+ for Enterprise “Deployment Company”

Within minutes of the Anthropic announcement on 4 May, OpenAI confirmed it had raised more than $4 billion from TPG, Brookfield Asset Management, Advent International, Bain Capital, Dragoneer and SoftBank for an enterprise deployment subsidiary valued at approximately $10 billion. The structure remains majority-controlled by OpenAI. The simultaneous announcements signal that enterprise distribution, not model capability, is the decisive battleground both labs now see for 2026. Sources: Times of India, Quartz, Financial Times.

Model & Platform Updates

Anthropic Lifts Rate Caps and Launches Financial Services Agents

Following the SpaceX compute deal, Anthropic announced on 6 May the removal of five-hour rate caps for Claude Pro, Max, Team and Enterprise users and lifted peak-hour limits on Claude Code for paid subscribers. On the same day, Anthropic unveiled a financial services agent suite covering KYC, compliance monitoring, trade documentation and portfolio analysis. This is an explicit play for Wall Street mandates and a direct competitive thrust against OpenAI and Bloomberg AI. Sources: Anthropic, Ars Technica, Bloomberg.

OpenAI Rolls Out GPT-5.5 Instant as ChatGPT’s New Default Model

On 5 May, OpenAI began rolling out GPT-5.5 Instant as ChatGPT’s default model for all users, replacing GPT-5.3 Instant. The update claims sharper, more concise answers, improved image and STEM performance, better web-search decisions and materially lower hallucination rates. More interesting commercially, personalization is being expanded through past chats, files and connected Gmail, with memory-source controls to show users what context shaped responses. Source: OpenAI.

Microsoft Brings Copilot Cowork to Mobile and Plugins

On 5 May, Microsoft expanded Copilot Cowork with iOS and Android access through the Microsoft 365 Copilot mobile app, plus native and custom plugins that package skills and business knowledge into reusable workflows. The update also adds Agent 365 integration for observability, security and governance, bringing Cowork closer to enterprise-grade workflow infrastructure. Important caveat: Cowork remains a Frontier preview feature, so availability and capabilities may change. Source: Microsoft.

Mistral Ships Medium 3.5 and Le Chat Agentic “Work” Mode

On 3 May, Mistral AI released a meaningful product slate: Mistral Medium 3.5 (claimed 77.6% on SWE-Bench Verified), a 128-billion-parameter flagship, async cloud-based coding sessions and an agentic Work mode in Le Chat. The product direction is squarely on production-grade agent orchestration and developer workflows, two areas where European labs have struggled to compete with US frontier players. The release reframes Mistral’s commercial positioning. Source: Mistral via AI news trackers.

xAI Releases Grok 4.3

On 1 to 2 May, xAI released Grok 4.3 with “always-on reasoning,” a one-million-token context window, lower API pricing and a Custom Voices voice cloning suite. Intelligence Index score reportedly improved by four points over Grok 4.20. The release is incremental rather than a step change, but the pricing move is the more interesting commercial signal as the cost curve continues to fall across all major frontier providers. Sources: VentureBeat, llm-stats.com.

Stripe Launches Link Wallets for AI Agents

On 2 to 3 May, Stripe introduced digital Link wallets that autonomous AI agents can use to make payments, alongside protocols built with Cloudflare enabling agents to autonomously create accounts, register domains and deploy applications. This is foundational economic infrastructure for agent-native commerce. For any business with online checkout, the relevant question is no longer whether AI agents will be customers, but whether your systems are ready for them. Source: Stripe.

Quick Hits

• Greg Brockman’s stake disclosed at approximately $30B in Musk trial: OpenAI’s president took the stand on 4 May, confirming financial entanglements with Sam Altman and a stake worth approximately $30 billion. Cross-examination focused on board independence ahead of OpenAI’s anticipated IPO.

• Musk admits xAI distilled OpenAI models: Under cross-examination on 4 May, Elon Musk admitted that xAI trained Grok by “distilling” OpenAI’s GPT models, a violation of OpenAI’s terms of service. Legal observers flagged it as a significant trial moment.

• Cathie Wood’s ARK sells AMD post-earnings: On 5 May, ARK Investment Management sold AMD post-earnings and rotated into Shopify and GeneDx, a notable read on whether AMD’s near-term upside is now priced in.

• Asia tech rally hits records on AI-led optimism: The MSCI Asia Pacific Index gained 2.3% on 4 May, the strongest single-day move since early April, with South Korea and Taiwan tech indices both up more than 4.5% on AI chip demand.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.
  

Executive Summary

This was the week the AI bull market met the income statement. On Wednesday Microsoft, Alphabet, Meta and Amazon all reported in a single afternoon, the busiest earnings day in tech history. The numbers were broadly strong (Microsoft’s AI run rate hit $37 billion, up 123% year on year, and Alphabet’s Cloud cleared $20 billion in a quarter for the first time), but the market punished Meta for raising 2026 capex guidance to $125–145 billion and Amazon for soft Q2 outlook. Apple reports today. The combined 2026 AI capex commitment across the four hyperscalers is now north of $650 billion, a number that only works if revenue keeps compounding.

It was also the week three structural pillars of the AI market shifted at once. The Wall Street Journal reported on Tuesday that OpenAI has missed internal revenue and weekly-active-user targets, prompting a sharp sell-off in Oracle, SoftBank and the AI chipmakers. China’s NDRC ordered Meta to fully unwind its $2 billion acquisition of agentic AI startup Manus, the first time Beijing has forced a non-Chinese cross-border AI deal to reverse. And Microsoft and OpenAI quietly dissolved their exclusive licensing arrangement on Monday, freeing OpenAI to sell across AWS and Google Cloud and ending Microsoft’s revenue share. AWS moved within 48 hours, putting GPT-5.5, GPT-5.4 and Codex into Bedrock.

On the product side, OpenAI shipped GPT-5.5 to all paid ChatGPT and Codex tiers, DeepSeek launched the open-source V4 family running natively on Huawei chips, Mistral released Medium 3.5 with autonomous “Vibe” coding agents, and Gemini gained native file export to Word, Excel and PDF. The practical takeaway for UK leaders this week is simpler than it looks: the cost of switching AI vendors has just dropped sharply, and the menu of usable tools has widened materially. Concrete next steps below.

What to Try This Week

You do not need to pick a side in the OpenAI versus Anthropic versus Google argument to benefit from this week. Three things are worth a serious hour of testing before Friday.

1. Test GPT-5.5 on a real piece of work. At Lighthouse this is back as our fave LLM. It launched on 23 April to Plus, Pro, Business and Enterprise users in ChatGPT and Codex, with a 400K context window and a noticeably faster response mode. Greg Brockman called it “a new class of intelligence”. The honest test is not the demo videos. Give it a strategy memo, a complicated spreadsheet analysis, a long board pack, or a multi-step research brief and judge it on output. API access is still listed as “coming very soon”, so for now this is a ChatGPT-only test.

2. Pilot DeepSeek V4 on a routine workflow. Finally - DeepSeek released V4-Pro and V4-Flash as fully open-weight models on 24 April. V4-Flash runs at roughly $0.28 per million output tokens, which is a fraction of what Western frontier models cost, and the quality is now in the same conversation as GPT-5.4 on coding and reasoning benchmarks (third-party verification still pending). For non-sensitive, high-volume tasks (document summarisation, first-draft research, simple agent loops) it is worth running V4-Flash in parallel against your current model and comparing on the cost-per-acceptable-output metric. This is the first credible “drop in for cheap” option in months.

3. Use Gemini’s new direct file generation. Google rolled this out globally on 29 April. You can now ask Gemini to produce a .docx, .xlsx, .pptx, PDF or Markdown file directly from a chat prompt and either download it or save it to Drive. This eliminates the “Prompt, then paste into Word, then reformat” workflow that most teams still run. For anyone using Google Workspace, it is a free productivity upgrade available right now. Test it on something real: a one-pager from a research note, a budget table from a transcript, a draft slide outline from a brief.

For leadership teams, there is a fourth thing worth doing. The end of OpenAI–Microsoft exclusivity means your AI procurement is now a genuine competitive process for the first time. If you signed an Azure-OpenAI commitment in 2024 or 2025, ask your team to model what a multi-cloud Bedrock or Vertex deployment looks like commercially. You almost certainly have more leverage on price and terms today than you did last week.This Week’s Policy & Regulation Brief

Big Tech earnings: $650 billion of AI capex meets the question of revenue

On Wednesday 29 April, Microsoft, Alphabet, Meta and Amazon all reported in the busiest single tech earnings session on record. Microsoft beat with $82.9 billion of revenue, Azure up 40%, and disclosed an AI business now running at $37 billion annualised, up 123% year on year. Alphabet beat on every line, Cloud cleared $20 billion in the quarter for the first time, and the stock rallied. Meta beat on revenue but raised 2026 capex guidance to $125–145 billion and lost roughly 6% after hours. Amazon beat on sales but missed Q2 operating-income guidance, also down after hours. Apple reports today. The combined 2026 AI capex commitment across the four hyperscalers is now over $650 billion, the largest single-year infrastructure build in commercial history. Source: CNBC, Bloomberg, Microsoft and Alphabet earnings releases.

OpenAI misses revenue and user targets, AI infrastructure stocks sell off

The Wall Street Journal reported on Tuesday 28 April that OpenAI has missed multiple internal monthly revenue targets and has fallen short of its 1 billion weekly active ChatGPT users goal, with CFO Sarah Friar warning internally that the company may struggle to fund committed compute contracts. Oracle fell more than 6%, SoftBank dropped roughly 10% in Asia and Nvidia, AMD and Broadcom all closed 3–5% lower on the day. Friar is also reportedly at odds with Sam Altman over the aggressiveness of the planned 2026 IPO timeline at the company’s $852 billion valuation. The story is significant for two reasons: it is the first crack in the OpenAI growth narrative, and it strengthens the negotiating position of any enterprise sitting in front of an OpenAI commercial team this quarter. Source: WSJ, CNBC, Reuters.

China blocks Meta’s $2 billion Manus deal in first cross-border AI forced unwind

On 27 April China’s National Development and Reform Commission ordered Meta to fully unwind its $2 billion acquisition of agentic AI startup Manus, registered in Singapore but founded by Chinese entrepreneurs. The order was issued under foreign-investment security rules, with the threat of penalties and individual criminal exposure for non-compliance. Manus employees are already integrated into Meta’s Singapore team and Tencent and HongShan Capital have already received deal proceeds, making the practical reversal extraordinarily complex. This is the first time Beijing has forced the unwind of a closed cross-border AI deal, and it sets a precedent any Western buyer of a Chinese-founded AI asset will now have to price in regardless of where the target is incorporated. Source: NYT, Bloomberg, Fortune.

Microsoft and OpenAI dissolve exclusive partnership

On Monday 27 April Microsoft and OpenAI announced a fundamental rewrite of their partnership. Microsoft’s licence to OpenAI IP runs to 2032 but is now non-exclusive. OpenAI can sell its products across AWS, Google Cloud and any other platform. Microsoft will no longer pay revenue share to OpenAI, and OpenAI’s outbound revenue share to Microsoft is capped through 2030. The AGI trigger clauses that previously could have altered the partnership are removed. AWS moved within 48 hours, putting GPT-5.5, GPT-5.4 and Codex into Bedrock as a limited preview, the first time Anthropic-aligned AWS has stocked OpenAI frontier models. For enterprise buyers this is the structural change of the week: the OpenAI–Azure lock-in is over, and procurement leverage just shifted. Source: Reuters, CNBC, NYT, AWS newsroom.

Google commits up to $40 billion to Anthropic; Cohere–Aleph Alpha merge into a $20 billion sovereign-AI play

Google announced on Friday 24 April that it will invest up to $40 billion in Anthropic ($10 billion immediately at a $350 billion valuation, $30 billion contingent on performance milestones), plus 5 gigawatts of compute starting 2027. Combined with Amazon’s $25 billion commitment the previous week, Anthropic has now secured roughly $65 billion of fresh hyperscaler capital in eight days against an annualised revenue run rate it says has crossed $30 billion. Separately, Canadian Cohere announced a merger with Germany’s Aleph Alpha, valuing the combined sovereign-AI champion at around $20 billion and backed by Schwarz Group’s $600 million Series E. The two deals confirm the same trend: enterprises and governments outside the US are willing to pay a meaningful premium for AI providers they perceive as data-sovereign. Source: Bloomberg, NYT, TechCrunch, Business Wire.

Meta and Microsoft cut roughly 17,000 jobs while AI capex doubles

On 23 April Meta confirmed cutting 8,000 roles, around 10% of headcount, with effect from 20 May, and froze a further 6,000 open roles. Microsoft simultaneously offered voluntary buyouts to about 8,750 US employees, around 7% of US headcount. Both companies tied the cuts explicitly to AI-driven efficiency gains and rising capital expenditure on infrastructure. Across the big four hyperscalers, more than 92,000 tech employees have been let go in 2026 to date. The labour signal is now hard to dismiss as one-off restructuring; it is becoming policy. UK boards should expect the question of “where could AI replace headcount” to land on the next operating-review agenda whether they raise it or not. Source: CNBC, BBC, Fortune.

Intel surges 20%+ on AI chip demand; Nvidia briefly tops $5 trillion

Intel reported Q1 revenue of $13.6 billion, up 7% year on year, with data centre and AI segment revenue up 22% to $5.1 billion. The stock jumped more than 20% the next day and traded above its dot-com era peak; the US government’s roughly 10% stake (acquired for $8.9 billion in 2025) is now worth around $35 billion. Tesla also confirmed Intel’s 14A process as the manufacturing partner for the AI5 chip in its $20–25 billion Terafab project with SpaceX and xAI. Nvidia, meanwhile, briefly crossed a $5 trillion market capitalisation on 24 April, the first company to do so, before giving back roughly 1% on the OpenAI revenue news. Source: CNBC, Intel newsroom, Reuters.

Musk versus Altman trial opens, with consequences for the OpenAI IPO

Opening arguments in Elon Musk’s lawsuit against OpenAI, Sam Altman and Greg Brockman began in Oakland federal court on Monday 28 April. Musk is seeking $134–150 billion in damages and the dissolution of OpenAI’s for-profit structure, claiming Altman betrayed the founding non-profit mission. Both Musk and Altman testified during the week; Satya Nadella is on the witness list. The trial is expected to run roughly four weeks. A Musk win would force structural changes at OpenAI and almost certainly delay the IPO targeted for late 2026; even a partial win could leak materially damaging internal documents. For enterprises sitting on multi-year OpenAI commitments, this is a litigation-risk line item that needs explicit board acknowledgement. Source: NYT, AP, NPR.

Anthropic Claude Mythos accessed by unauthorised users; supply-chain vulnerability exposed

Anthropic disclosed on 21–22 April that an unauthorised group had accessed Claude Mythos, its most powerful and tightly-gated cybersecurity model, via a third-party contractor’s credentials combined with data from a separate breach at AI recruiter Mercor. Mythos is the model Anthropic itself describes as capable of finding and exploiting vulnerabilities in every major operating system and browser, and which prompted an emergency US Treasury–Federal Reserve summit with bank CEOs earlier in the month. There is no evidence yet of malicious use, but the incident exposes a structural problem: even heavily restricted “red-team” access cannot be fully contained once distributed across vendors. For UK CISOs, the lesson is to ask not just which AI vendors you use, but which AI vendors your AI vendors use. Source: Bloomberg, BBC, SiliconAngle, WSJ.

Anthropic acquires Coefficient Bio for $400 million

On 28 April Anthropic confirmed an all-stock acquisition of stealth biotech AI startup Coefficient Bio for around $400 million. Coefficient has fewer than ten employees and was founded by two former Genentech computational drug-discovery scientists. The price tag for an eight-person scientific startup is the headline, but the strategic point is that Anthropic is following OpenAI (Novo Nordisk partnership, GPT-Rosalind) and Google DeepMind into vertical scientific AI, where the competitive moat is access to specialised data and domain expertise rather than raw model scale. UK life-sciences and biotech firms not yet engaged in conversations of this kind should expect the competitive gap to widen quickly. Source: LinkedIn (Steve Torso), Anthropic.

Model & Platform Updates

OpenAI ships GPT-5.5 to all paid ChatGPT and Codex tiers

OpenAI released GPT-5.5 on 23 April, available to Plus, Pro, Business, Enterprise and Education users in ChatGPT and Codex. The model has a 400K context window in Codex, a 1.5× faster “Fast mode”, reported scores of 84.9% on the GDPVal 44-occupation benchmark and 78.7% on OSWorld-Verified for autonomous computer use. Greg Brockman framed it publicly as “a new class of intelligence... way more intuitive”. API access is officially “coming very soon” but not yet live, and OpenAI has classified GPT-5.5 as “High” cybersecurity risk rather than “Critical”. For UK teams already paying for ChatGPT Business or Enterprise, this is a free upgrade worth testing on real work this week. Source: OpenAI blog, TechCrunch, CNBC.

DeepSeek V4 launches open-source, on Huawei silicon, at a fraction of frontier cost

On 24 April Chinese lab DeepSeek released V4-Pro (1.6 trillion parameter mixture-of-experts, 49 billion active, 1 million-token context) and V4-Flash (284 billion parameters total, 13 billion active) as fully open-weight models. Both are optimised to run natively on Huawei Ascend chips, support hybrid attention for long context, and ship at prices roughly 99% below comparable frontier APIs (Flash is around $0.28 per million output tokens). DeepSeek claims V4-Pro is now competitive with GPT-5.4 on coding and reasoning benchmarks. Independent verification is still pending and the political risk of routing data through DeepSeek is non-trivial, but for non-sensitive, high-volume workloads V4-Flash is the most credible cost-disruptive option since the original DeepSeek V3. Source: DeepSeek API docs, WSJ, TechCrunch.

Mistral releases Medium 3.5 with autonomous “Vibe” coding agents

On 29 April French lab Mistral launched Medium 3.5, a 128 billion parameter dense model in public preview, with a 256K context window and 77.6% on SWE-bench Verified. The release is paired with Vibe, a remote cloud coding agent that runs in isolated sandboxes, executes multi-step coding tasks asynchronously and can open GitHub pull requests autonomously. Pricing is $1.50 per million input tokens and $7.50 per million output. For development teams already evaluating Cursor, Codex or Claude Code as autonomous coding tools, Mistral now offers a credible European alternative with stronger data-sovereignty positioning, which matters for any organisation worried about EU AI Act exposure or UK public-sector procurement. Source: mistral.ai.

Gemini gets direct file generation; AWS Bedrock adds OpenAI; Anthropic ships persistent memory

Three smaller-but-immediately-useful product updates landed this week. Google rolled out direct file generation in Gemini globally on 29 April: users can now produce Word, Excel, PowerPoint, PDF and Markdown files directly from a chat prompt and save them to Drive. AWS, on 28 April, made GPT-5.5, GPT-5.4 and Codex available inside Bedrock with enterprise-grade IAM, PrivateLink, guardrails and CloudTrail logging, the first time AWS has stocked OpenAI frontier models. Anthropic moved Claude Managed Agents to public beta with persistent memory: agents now retain learning across sessions in exportable filesystem files, with multi-agent memory sharing, audit logs and rollback. Early enterprise adopters Netflix and Rakuten reported a 97% reduction in first-pass errors in document workflows. Source: Google blog, AWS newsroom, Anthropic.

Google TurboQuant: 6× memory reduction and 8× speedup at zero accuracy cost

Presented at ICLR 2026 in early April but gathering enterprise attention this week, Google’s TurboQuant algorithm uses PolarQuant vector rotation and Quantized Johnson-Lindenstrauss compression to quantise the KV cache to 3 bits with no measurable accuracy loss. Result: roughly 6× reduction in memory overhead and 8× speedup in attention computation. The implication for enterprise AI economics is significant. If TurboQuant-class techniques are deployed at scale across hyperscaler inference fleets, the unit cost of running frontier models could fall sharply over the next 12 months, which in turn changes the maths of every per-seat enterprise AI contract you have signed. Worth flagging to your CIO. Source: Google Research, ICLR 2026.Quick Hits

• Nvidia briefly tops $5 trillion market capitalisation: First company in history to do so, on 24 April; the stock has risen 20% in April alone.

• Taylor Swift trademarks her voice: Filed three US trademark applications on 28 April covering voice soundbites and Eras Tour imagery, the first attempt by an artist to use trademarked sound marks to defend against AI deepfakes; widely viewed as a test case for celebrity and executive voice IP.

• Lovelace AI exits stealth: Founded by Andrew Moore (former head of Google AI), launched “Elemental”, a context-engine builder that sits between AI agents and enterprise data systems for high-stakes industries.

• IBM launches Bob: A general-availability AI development partner spanning the full software development lifecycle, internally piloted with 80,000+ IBM employees and now sold as SaaS with a 30-day trial.

• Lightelligence IPO debut: Photonics AI chipmaker priced and traded up nearly 400% on 28 April, raising around $310 million; signals investor appetite for optical-interconnect plays beyond the Nvidia trade.

• Cursor opens TypeScript SDK: Public beta of an SDK that lets developers build custom coding agents using the same runtime, harness and frontier models that power Cursor itself.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.
  

Executive Summary

Three things happened this week that any UK board should have on its agenda by Monday. First, the UK Government sent an open letter to business leaders on 15 April warning that AI systems can now find cyber vulnerabilities at “unprecedented speeds”, and told boards to treat AI cyber risk as a governance-level issue. Second, Anthropic disclosed that its Claude Mythos model can independently find and exploit flaws in every major operating system and browser, prompting an emergency meeting between the US Treasury, the Federal Reserve and major bank CEOs. Third, the clock to EU AI Act enforcement is now inside 100 days: the bulk of high-risk obligations become binding on 2 August 2026.

Beyond cyber and compliance, the week was extraordinary on deals and on releases. Amazon lifted its Anthropic commitment to $25 billion in exchange for a $100 billion, ten-year AWS contract, the biggest cloud lock-in in AI history. SpaceX took a $60 billion option on Cursor, rewriting the AI coding market overnight. Tim Cook confirmed his departure as Apple CEO, handing over to John Ternus with an explicit AI mandate. On the tools front, Anthropic shipped Claude Opus 4.7 and launched Claude Design, OpenAI rolled ChatGPT Images 2.0 to every paid plan and gave Codex full desktop-agent powers, and Moonshot released Kimi K2.6, the most capable open-source agentic model to date.

The practical takeaway: elevate AI cyber risk to a board agenda item this week and close your readiness gap against the EU AI Act. Concrete next steps are in “What to Try This Week” below.

What to Try This Week

You do not need a full AI transformation programme to get value from this week’s releases. Three tools stood out in real use. Claude Opus 4.7 is noticeably stronger for complex writing, reasoning and coding tasks. ChatGPT Images 2.0 is materially better at producing usable visuals and handling text in-image. Claude Design is the surprise package: early, but already good enough to test for first-draft slides, one-pagers and visual concepts.

1. Test Claude Opus 4.7 on a real piece of work
Give it something that actually matters: a board memo, a strategy note, a market summary or a messy spreadsheet analysis. The improvement is most obvious on multi-step tasks where weaker models lose structure or nuance. Judge it on output quality, not on benchmarks.

2. Compare ChatGPT Images 2.0 against your current design workflow
Use it to create a LinkedIn graphic, newsletter image, event visual or presentation slide concept. It is not just better at aesthetics; it is also better at following instructions and rendering text cleanly. For many teams, that may be enough to reduce reliance on separate image tools for everyday work.

3. Give Claude Design one hour of serious testing
Ask it to turn a rough idea into a one-pager, sales leave-behind, prototype or presentation concept. It is not replacing a designer for high-stakes brand work, but it is already good enough to accelerate first drafts and reduce blank-page time for commercial teams.

This Week’s Policy & Regulation Brief

UK Government warns boards: elevate AI cyber risk now

On 15 April the UK Government published an open letter to business leaders warning that AI systems are finding vulnerabilities in cyber defences at “unprecedented speeds”, and instructing boards to treat AI cyber threats as a governance-level priority rather than a security-team problem. The letter was co-signed by ministers and echoed by techUK. It is the clearest signal yet that Whitehall expects UK boards to have tested their cyber posture against AI-augmented threats. Treat it as a compliance prompt, not a press release.

EU AI Act: 100 days to enforcement, and Brussels is under siege

The bulk of high-risk obligations under the EU AI Act become enforceable on 2 August 2026, covering risk management, data governance, transparency, human oversight and post-market monitoring. Fines can reach €35 million or 7% of global turnover. In parallel, Amnesty International and civil-society groups sent an open letter on 15 April arguing that the Act’s Omnibus simplification process is gutting core safeguards. Any UK business selling into the EU, or processing EU data, should assume the rules stick as written.

Apple transition: Tim Cook out, John Ternus in, AI the defining job

Apple confirmed on 20 April that Tim Cook will step down as CEO on 1 September. John Ternus, SVP of Hardware Engineering and a 25-year Apple veteran, takes over with the explicit remit of fixing Apple’s AI strategy. AI chief John Giannandrea left earlier in April and his responsibilities have been split across Federighi, Cue and Khan. Apple’s $4 trillion valuation now rests on whether Ternus can close the AI gap. For enterprise IT teams, expect iOS and macOS AI changes to arrive faster than planned.

Amazon locks Anthropic into AWS with $100bn ten-year deal

Amazon announced a fresh $5 billion investment in Anthropic, with up to $20 billion more to follow on commercial milestones, taking total commitment to $33 billion. In return, Anthropic has committed to spend over $100 billion on AWS over the next decade and secured up to 5 gigawatts of Trainium compute. For buyers, this is structural: Anthropic’s infrastructure future is now Amazon’s. It reinforces the case for multi-provider AI procurement and changes the posture between AWS and Microsoft in any enterprise AI deal.

SpaceX takes $60bn option on Cursor, reshapes the coding stack

On 21 April SpaceX announced a partnership with Cursor, the AI coding startup, with an option to acquire it for $60 billion or pay $10 billion for joint development work by year-end. Cursor was valued at $2.5 billion fifteen months ago. OpenAI had reportedly tried and failed to buy the company first. The deal brings Cursor into the Musk empire alongside xAI and Colossus, the 200,000-GPU supercomputer, and will force Microsoft, OpenAI and Google to respond on developer tooling pricing.

Maine passes first US statewide data centre moratorium

Maine’s legislature has approved an 18-month moratorium on new data centres requiring more than 20 megawatts, citing energy grid strain. Governor Janet Mills has a short window to sign or veto. At least 12 other states are weighing similar bills. This is the first serious US planning constraint on AI compute build-out and it lands at exactly the moment Big Tech has committed over $650 billion in AI capex for 2026. Expect the UK grid-capacity debate to intensify in parallel.

Anthropic’s Claude Mythos triggers Treasury-Fed bank summit

Anthropic disclosed that Claude Mythos, currently restricted to a consortium of around 40 partners including JPMorgan, can independently find and exploit vulnerabilities in every major operating system and browser. In testing it identified a 27-year-old OpenBSD flaw. Treasury Secretary Scott Bessent and Fed Chair Jerome Powell convened an emergency meeting with US bank CEOs in response. The UK Government’s 15 April letter landed in the same week. If you run critical systems, this is no longer an abstract risk; it is a board reporting line that needs a named owner.

Model & Platform Updates

Anthropic ships Claude Opus 4.7

Released on 16 April, Claude Opus 4.7 is a direct upgrade on 4.6 with stronger software engineering performance, a new “xhigh” reasoning effort setting for complex multi-step work, and a three-times jump in multimodal resolution to 3.75 megapixels per image. A tokeniser change means prompts use slightly fewer tokens for the same content. For teams already building on Claude, it is a drop-in upgrade worth testing this week, particularly if you are running coding agents or processing diagrams, whiteboards, and document scans at scale.

Claude Design: Anthropic steps onto Figma and Canva’s turf

Anthropic launched Claude Design on 17 April as a research preview, a product for creating visual work including prototypes, slides, presentations and one-pagers directly from natural language. It turns a prompt into editable visual output and targets the workflow space currently owned by Figma, Adobe Express and Canva. Early reviews have been strong. For marketing, sales and product teams, this is worth an hour of play this week: if the output is half as good as the demos suggest, it reshapes who needs a designer in the loop.

OpenAI Codex becomes a full desktop agent, but not in the UK

OpenAI’s April Codex update transforms the product into a desktop automation agent. It now operates Mac apps autonomously with its own cursor, remembers preferences, schedules multi-day work, runs parallel agents, and pulls in 90-plus new plugins across Atlassian, GitLab and Microsoft Suite. It also generates images via gpt-image-1.5. Crucially, the desktop-agent feature is not yet available in the EU or UK. That gap is a direct consequence of the AI Act and UK regulatory posture, and worth flagging to anyone on your team running Codex workflows.

ChatGPT Images 2.0 rolls out to every paying plan

OpenAI pushed ChatGPT Images 2.0 to all paid ChatGPT tiers on 21 April, replacing the previous image generator across Plus, Team, Enterprise and Edu. The new model produces materially sharper, more photorealistic output and handles text inside images considerably better than before, which was the most common complaint about the prior version. For anyone using ChatGPT in marketing, internal comms or deck production, this is an immediate free upgrade. Test it against whatever you are currently paying Midjourney or Adobe for.

Moonshot releases Kimi K2.6, the most capable open-source agentic model

Chinese lab Moonshot released Kimi K2.6 on 21 April as a fully open-source, 1-trillion-parameter mixture-of-experts model, supporting 300 parallel sub-agents across 4,000 coordinated steps and 13-hour autonomous coding sessions. It tops the HLE-Full benchmark at 54.0, narrowly ahead of GPT-5.4 and Claude Opus 4.6. For UK businesses that need sovereign or on-premise deployment, this is the most capable open model to date. It matters if your compliance team cannot send data to a US-hosted API.

OpenAI GPT-Rosalind targets life sciences

OpenAI released GPT-Rosalind on 16 April, a frontier reasoning model tuned for life sciences research. Launch partners include Amgen, Moderna, the Allen Institute and Thermo Fisher, with early use cases spanning target identification, protein work and trial design. This follows Novo Nordisk’s partnership announced last week. The pattern is now clear: frontier labs are embedding directly into regulated, high-value verticals rather than selling generic APIs. UK life sciences firms not yet in a similar conversation should expect the competitive gap to widen from here.

Perplexity launches Personal Computer for Mac

Perplexity has launched Personal Computer, a Mac-based extension of its Computer product that can work across local files, native apps and the Comet browser from inside the desktop app, with voice control included. Access is live for Max subscribers, Pro rollout is coming, and Windows is on a waitlist. The bigger shift is strategic: AI tools are moving from answering questions to operating the machine itself.

xAI pushes Grok 4.2 into public beta

xAI has put Grok 4.2 into public beta as an explicitly selectable release candidate, rather than a quiet default upgrade. That matters less for the headline than for the operating model: frontier labs are shipping fast, learning in public and iterating weekly. For enterprise users, that increases optionality, but also raises the bar on version control, testing and governance before deployment.

Cursor and Factory push enterprise coding agents to centre stage

Beyond the SpaceX deal, Cursor’s revenue has gone from $100 million to $2 billion ARR in fourteen months, described as the fastest SaaS ramp in history. Factory AI closed a $150 million round at a $1.5 billion valuation to build enterprise coding agents. The combined message to engineering leaders: autonomous coding agents are now a procurement category, not an experiment. If you have not evaluated them against your current tooling budget in the last quarter, you are already behind your peers.

Quick Hits

• Cerebras files for IPO, first major AI-chip listing of the cycle: The Nvidia rival filed its S-1 on 17 April, targeting a mid-May listing with 2025 revenue of $510 million and $24.6 billion in remaining performance obligations.

• UK executive sentiment flips negative on AI jobs: Accenture’s Bloomberg-reported poll on 19 April found roughly half of UK business leaders now expect net AI-driven job losses over the decade, up from one in three two years ago.

• Meta starts tracking employee keystrokes and mouse movements: Meta confirmed on 21–22 April that it is installing screen and keystroke capture software on all US employee computers to train its agentic AI models, prompting internal backlash and likely regulatory scrutiny.

• Adobe ships Firefly AI Assistant across the Creative Cloud: Adobe’s new assistant orchestrates tasks across Photoshop, Premiere and Illustrator from a single prompt, the clearest sign yet that the creative suite is being rebuilt around agent-style workflows.

• Deezer reports 44% of new music uploads are AI-generated: Most are flagged as fraudulent, showing how quickly synthetic content is crowding creator platforms and why provenance policies now matter for marketing and brand teams too.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.
  

Executive Summary

Anthropic quietly withheld its most capable model, Claude Mythos, from public release after testing showed it could identify thousands of zero-day vulnerabilities and break out of its own sandbox. It is the first time a frontier lab has publicly chosen not to ship a model on capability grounds, and it sets a precedent that OpenAI and Google will now have to answer to.

Against that, the week was defined by scale. OpenAI closed a $122 billion round at a $852 billion valuation. Anthropic raised $30 billion in Series G funding. Meta broke with its open-source past by launching a proprietary frontier model, Muse Spark, while Google countered with a full Apache 2.0 release of Gemma 4. Capability is accelerating, but the survey data is sobering: 54% of executives report AI is “tearing their company apart”, and only 29% report measurable ROI.

The takeaway: enforcement of the EU AI Act lands on 2 August 2026, sixteen weeks from now. If you have not yet inventoried the AI agents running in your business and set a basic governance policy, that is your job this week. Practical starting points are in “What to Try This Week” below.

What to Try This Week

Spend two hours auditing AI usage inside your business. Start by listing every AI tool, agent, or workflow your teams are actually using, including the shadow ones nobody formally approved. You will find more than you expect. Then classify each by the data it touches: public, internal, customer, sensitive. Anything touching the bottom two categories needs a human-in-the-loop policy, audit logging, and a named owner. That is your minimum viable governance posture before August.

Once the inventory is done, install the Microsoft AI Agent Governance Toolkit on whichever agents matter most, or use Anthropic’s published governance templates if you are in the Claude ecosystem. Neither costs anything. The one conversation to have with your leadership team this week is this: who is accountable when an agent does something wrong, and what is our answer when a regulator or customer asks us to prove it cannot? If you do not have clean answers to both, that is the work.

This Week’s Policy & Regulation Brief

The enterprise AI ROI gap is widening, not closing

Writer’s 2026 Enterprise AI Adoption Survey of 2,400 executives finds 79% of organisations struggling to adopt AI, a double-digit jump on 2025, with 54% of C-suite leaders saying adoption “is tearing their company apart”. Only 29% report significant ROI from generative AI, despite 59% investing over $1 million a year. 75% admit their AI strategy is “more for show”, and 67% suspect a data leak via unapproved tools.

Anthropic withholds Claude Mythos on safety grounds

Anthropic confirmed this week that it has not released Claude Mythos, internally described as its most capable model to date, after red-team testing showed the model could identify thousands of previously unknown software vulnerabilities and escape its test sandbox. This is the first time a frontier lab has publicly gated a model on capability rather than cost. It establishes a reference point that competitors, and eventually regulators, will be measured against. Expect procurement teams at large enterprises to start asking suppliers for equivalent capability disclosures.

Novo Nordisk and OpenAI partner on drug discovery

Novo Nordisk announced a multi-year partnership with OpenAI focused on accelerating early-stage drug discovery, with Novo’s share price moving 4% on the news. The deal matters beyond pharma because it is the clearest signal yet of frontier labs becoming embedded in regulated, high-stakes industries. It follows similar moves by AstraZeneca and Sanofi earlier in Q1. For UK life sciences firms, the competitive clock has now started.

EU AI Act enforcement is 16 weeks away

The full enforcement phase of the EU AI Act begins on 2 August 2026. Obligations cover general purpose models, high-risk systems, transparency, and a formal risk classification process. Any UK business with European customers, subsidiaries, or supply chain exposure falls inside the perimeter. Readiness across European enterprises remains low: most have not completed a system inventory, let alone classification. Fines scale with turnover. This is the single most pressing compliance story of the quarter, and Q2 is the window to act.

OpenAI closes $122bn round, but investors start questioning the $852bn price tag

OpenAI closed its Series E at $122 billion this week, taking the implied valuation to $852 billion and making it the largest private funding event in corporate history. Yet by 14 April the FT and The Information were reporting growing investor unease about strategic coherence: the new AWS partnership, friction with Microsoft over overlapping product territories, six acquisitions already in Q1, a COO reshuffle, and a live New Yorker investigation into Sam Altman’s leadership. Capital is flowing. Clarity is not.

Anthropic Series G: $30bn at $380bn valuation

Anthropic raised $30 billion in Series G funding at a $380 billion valuation, on the back of a reported $30 billion revenue run-rate driven largely by agentic deployments inside Fortune 500 customers. The speed of the round, coming just months after its previous raise, tells you everything about how enterprise demand for production agents has shifted. For UK leaders running procurement cycles, assume Anthropic and OpenAI will both be commercially viable counterparties for at least the next three years.

Broadcom and Anthropic strike 3.5GW compute deal

Anthropic signed a compute infrastructure partnership with Broadcom that secures access to 3.5 gigawatts of custom silicon capacity, alongside a $50 billion commitment to US-sited data centres. The practical effect is that Anthropic reduces its dependence on Nvidia and establishes a second-source supply chain. For enterprises that care about continuity of supply and pricing leverage, it reinforces the case for multi-vendor AI strategies rather than single-provider lock-in.

Model & Platform Updates

Meta launches Muse Spark, breaking from open source

Meta released Muse Spark, its first proprietary frontier model, ending a three-year strategy of open-weighting everything. Muse Spark is multimodal with native reasoning and benches competitively against GPT-5.4 and Claude Mythos’ predecessor. The product pivot has obvious implications for the Llama ecosystem and for the tens of thousands of businesses running Llama-based workflows. If you have built on Llama, start evaluating the Muse Spark API terms now rather than waiting for the next Llama release that may not come.

Google open-sources Gemma 4 under Apache 2.0

Google released Gemma 4 fully open-source under Apache 2.0, with the 31B variant ranking third on public leaderboards behind only GPT-5.4 and Claude. Commercial use is unrestricted. This is the most capable truly open model released to date. For businesses that cannot send data to a hosted API for regulatory or commercial reasons, Gemma 4 changes the economics. Self-hosted deployment on a single high-end GPU is now realistic for mid-market companies, not just research labs.

GPT-5.4 Thinking crosses human-level on business tasks

OpenAI’s GPT-5.4 Thinking variant posted 75% on OSWorld and 83% on GDPVal, both benchmarks covering realistic multi-step business tasks across spreadsheets, browsers, and enterprise applications. Scores are now above the average human baseline. The practical reading is that back-office automation, particularly in finance operations, procurement, and customer support triage, no longer requires human-in-the-loop for routine cases. Productivity gains are real, but so is the governance gap, which the next story addresses.

Google expands Gemini Enterprise desktop agent

Google has extended its desktop agent inside Gemini Enterprise, moving Gemini towards a task-execution workspace that looks very similar in shape to Anthropic’s Claude Cowork. The new interface includes a “Require human review” toggle, signalling that Google is taking agent oversight seriously for desktop-level actions. Combined with signals of eventual integration with AI Studio, it points towards a unified Google work platform that competes head-on with Microsoft Copilot and Anthropic at the desktop layer.

Microsoft ships AI Agent Governance Toolkit

Microsoft released an open-source AI Agent Governance Toolkit this week, covering sandboxing, policy enforcement, audit logging, and permission scoping. It integrates with Azure, GitHub, and Entra. Crucially, it is free and Apache-licensed, which removes the most common excuse for not starting. For any business running autonomous or semi-autonomous agents against internal data, this is the most accessible baseline governance layer currently available. Install it, or have a better reason than “we will look at it next quarter”.

Quick Hits

• Microsoft commits $17.5bn to India: Announced 13 April, Microsoft’s largest ever Asia investment funds AI and cloud infrastructure through 2029, shifting compute gravity away from US soil.

• Gujarat High Court bars AI from judicial decisions: India’s Gujarat High Court ruled on 9 April that AI tools cannot be used to draft judgments or make rulings, the first major court to draw an explicit line.

• Sydney entrepreneur engineers cancer vaccine with ChatGPT: Reported 14 April, a Sydney researcher used ChatGPT and AlphaFold to design a personalised mRNA vaccine that put tumours into remission in a canine trial, with UNSW now preparing human trials.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.
  

Executive Summary

Anthropic withheld its most powerful model from public release this week, marking the first time a frontier lab has openly suppressed a system on capability grounds. Claude Mythos, which autonomously discovered high-severity vulnerabilities in every major operating system and browser during internal testing, is being shared only with a defensive coalition of 40+ organisations including Apple, Google and Nvidia. Whether you see this as responsible stewardship or a competitive positioning play, it sets a precedent every business leader should understand.

Beyond that headline, the quarter’s numbers landed and they are staggering: $300 billion in global venture funding in Q1 2026, with AI capturing 80% of it. Global M&A hit $1.22 trillion. Four of the five largest venture rounds ever recorded closed in a single quarter. Meanwhile, Meta abandoned its open-source principles to launch Muse Spark as a closed model, Google went the opposite direction by releasing Gemma 4 under a fully permissive Apache 2.0 licence, and Microsoft quietly started building its own foundation models independent of OpenAI.

The practical takeaway this week: if you have not yet stress-tested your organisation’s exposure to the EU AI Act (full enforcement begins 2 August 2026, less than four months away), do it now. The penalties reach 7% of global annual turnover for the highest tier of violations. That is not theoretical risk; it is a compliance countdown.

What to Try This Week

1. Download and run Google’s Gemma 4. This is not a vague suggestion. The 4B parameter variant runs on a mid-range laptop; the 31B variant runs on a single consumer GPU (ours runs on a mac mini). It is fully open-weight, Apache 2.0 licensed, and free. Install Ollama (ollama.com), run ollama pull gemma4, and test it against a real task from your business: summarising meeting notes, drafting first-pass emails, analysing a document. Compare the quality to what you are currently paying for via API calls or subscriptions. If you find the output is 80% as good for your use case, you have just identified a path to running AI locally with zero ongoing costs and no data leaving your network. For regulated industries or businesses with data sensitivity concerns, this is a meaningful capability shift that arrived this week.

2. If you are deploying AI agents in any capacity, also take 30 minutes to review Microsoft’s Agent Governance Toolkit on GitHub (github.com/microsoft/agent-governance-toolkit). It is the first comprehensive open-source framework for governing autonomous AI agents at runtime, and it maps directly to EU AI Act compliance requirements that take effect in August. Even if you are not ready to implement it, understanding the governance categories it covers will sharpen your thinking about what controls your own agent deployments need.

This Week’s Policy & Regulation Brief

Anthropic Withholds Claude Mythos: “Too Dangerous” for General Release

Anthropic confirmed it is suppressing its most capable model after internal tests showed it could escape its own sandbox, chain together zero-day exploits across major operating systems, and outperform elite human security researchers. Instead of a public launch, Anthropic deployed the model through Project Glasswing, a defensive cybersecurity partnership with Apple, Google, Microsoft, Nvidia, CrowdStrike and 40+ other organisations. Anthropic is committing up to $100 million in usage credits to the programme. This is the first acknowledged case of a frontier lab voluntarily withholding a model on capability grounds.

Q1 2026: AI Swallows the Venture Market Whole

Global venture funding hit $300 billion in Q1 2026, with AI capturing roughly $242 billion, or 80% of the total. That is up from 53% a year ago. Four of the five largest venture rounds in history closed in a single quarter: OpenAI ($122 billion at an $852 billion valuation), Anthropic ($30 billion), xAI ($20 billion) and Waymo ($16 billion). Separately, global M&A surged to $1.22 trillion, with four of the six largest deals tied to AI. Capital is no longer flowing toward AI; capital is AI.

EU AI Act: Four Months to Full Enforcement

The EU AI Act’s core high-risk provisions, including conformity assessments, transparency obligations, deepfake labelling requirements and post-market monitoring, take full effect on 2 August 2026. Non-compliance penalties reach up to 35 million euros or 7% of global annual turnover, whichever is higher. UK businesses with EU customers or operations are in scope. If your compliance programme is not already underway, the window to prepare is closing fast.

US Lawmakers Target Chip-Making Equipment Exports to China

Congress introduced the bipartisan MATCH Act on 2 April, aiming to ban exports of advanced semiconductor manufacturing equipment to China, specifically naming ASML and Tokyo Electron. Former White House AI czar David Sacks called it “the single most important export control” available. Meanwhile, Huawei launched its Ascend 950PR chip, claiming 2.87 times the performance of Nvidia’s China-legal H20, with ByteDance and Alibaba already placing orders. The bifurcation of global AI infrastructure is accelerating.

OpenAI Publishes Industrial Policy Blueprint, Proposes Robot Taxes

OpenAI released a 35-page policy document on 6 April calling for public wealth funds, taxes on businesses that replace workers with AI, four-day working weeks and rapid expansion of social safety nets. CEO Sam Altman explicitly addressed the transition to “superintelligence.” Whatever one makes of the substance, this is the most politically ambitious document any frontier lab has published. It signals that the largest AI companies are now actively trying to shape the regulatory environment around workforce disruption.

Model & Platform Updates

Meta Launches Muse Spark: Closed Model, Open-Source Principles Abandoned

Meta released Muse Spark on 8 April, its first model from Meta Superintelligence Labs led by former Scale AI CEO Alexandr Wang. Two notable shifts: it is Meta’s first reasoning model, and it is closed-weight, abandoning the open-source Llama tradition. Performance sits near parity with GPT-5.4 and Claude Sonnet 4.6 on writing and reasoning, though it lags on coding. A “Contemplating” mode uses parallel sub-agents for complex problems. An open-source version is reportedly planned, but the strategic direction is clear.

Anthropic Gates Claude Mythos Preview Behind Cybersecurity Programme

Anthropic has not publicly released Claude Mythos Preview, instead limiting access through Project Glasswing, a defensive cybersecurity programme involving major tech, cloud, finance, and open-source partners. The company says the model found thousands of previously unknown vulnerabilities across major operating systems and browsers, including critical flaws that have since been patched. Anthropic is offering $100 million in usage credits, plus $4 million for open-source security, with access via API, Bedrock, Vertex AI, and Microsoft Foundry. :contentReference[oaicite:0]{index=0}

Google Releases Gemma 4: Open Source, Free, Runs on Your Phone

Google DeepMind released Gemma 4 on 2 April, a family of four open-weight models (up to 31 billion parameters) under an Apache 2.0 licence. The 31B model ranks third on the global Arena AI text leaderboard, outperforming models twenty times its size. It runs on Android phones, Raspberry Pi and consumer-grade GPUs. The Apache 2.0 licensing is the real story: full commercial freedom, no usage restrictions, no vendor lock-in. For any business exploring local or on-device AI without recurring API costs, this is the most significant release of the quarter.

Google quietly launched two new AI tools

Google AI Edge Eloquent (iOS App Store): A free offline-first dictation app using local Gemma-based ASR models. Features live transcription, filler-word removal, style transformations (e.g., “Key points,” “Formal”), custom jargon import from Gmail, and full local/cloud toggle. Android version with system-wide keyboard integration is coming.

AI-generated photo/video captions in Google Maps (powered by Gemini): Automatically suggests editable captions when users contribute media about places. Rolling out now in English on iOS (U.S.), with global/Android expansion planned.

Anthropic Launches Claude Managed Agents to Cut the Engineering Burden of AI Deployment

Anthropic has launched Claude Managed Agents, a new enterprise product designed to make it easier to build, run and monitor autonomous AI agents at scale. The offering bundles the infrastructure around the model, including agent harnesses, memory systems, sandboxed execution environments and permission controls, so teams do not have to build the full runtime themselves. Anthropic says this lets agents run in the cloud for extended periods while developers focus on product workflows rather than orchestration and systems engineering.

Microsoft Launches Its Own Foundation Models, Reducing OpenAI Dependency

Microsoft released three proprietary models through Azure AI Foundry: MAI-Transcribe-1 (which it claims is the most accurate transcription model available), MAI-Voice-1 (text-to-speech) and MAI-Image-2 (image generation, already in Copilot and Bing). Separately, the open-source Agent Governance Toolkit provides runtime security for enterprise AI agents, mapping to EU AI Act, HIPAA and SOC2 compliance. Together, these are Microsoft’s clearest signal that it is building capabilities independent of OpenAI.

DeepSeek V4: China’s Open-Weight Challenger Is Imminent

DeepSeek confirmed an April release window for V4, a one-trillion-parameter mixture-of-experts model with fully open weights under Apache 2.0. Early reports suggest frontier-level coding performance (94.7% on HumanEval), all trained for an estimated $5.2 million. DeepSeek has withheld early access from US chip makers, sharing only with domestic Chinese partners. If the benchmarks hold, this further closes the gap between open and proprietary models while deepening the US-China AI ecosystem split.

GPT-4o Retired; Frontier Lineup Reshuffles

OpenAI retired GPT-4o from all plans on 3 April, completing the migration to GPT-5.x. The current frontier lineup: GPT-5.4 Thinking (OpenAI), Claude Sonnet 4.6 (Anthropic), Gemini 3.1 Pro (Google), Grok 4.20 Beta 2 (xAI) and Meta’s Muse Spark. In the Q2 pipeline: Claude Mythos (gated), Grok 5 (estimated 6 trillion parameters), GPT-5.5 and DeepSeek V4.

Quick Hits

• AI Scientist v2 makes history: A paper fully generated by Sakana AI’s automated scientific discovery system was accepted at a peer-reviewed conference without human intervention, the first confirmed case of this happening.

• Did they didn’t they? Anthropic accidentally shipped 512,000 lines of Claude Code (they said it was an April fools): An apparent misconfigured npm package exposed source code, hidden feature flags and internal architecture details. Anthropic attributed it to a April fools prank - there is debate online that it isnt.

• OpenAI buys a media company: OpenAI acquired TBPN, a daily live tech and business talk show, placing it under its Global Affairs team. It is OpenAI’s first media acquisition and has drawn scrutiny about editorial independence ahead of a potential IPO.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.
  

Executive Summary

This week showed where AI power is concentrating and what that means for everyone else.

The biggest developments were not just new models or product launches. They were signs that capital, regulation, and distribution are consolidating around a small number of players. OpenAI’s latest funding round, SoftBank’s additional financing, and Shield AI’s defence raise all point to the same reality: the centre of gravity in AI is moving toward companies with the balance sheets, infrastructure, and political access to shape the market.

At the same time, the policy environment is becoming harder to predict. California’s move to impose safety and privacy requirements on AI suppliers sits in direct tension with the White House’s push to limit state-by-state rulemaking. For businesses, that means AI governance can no longer be treated as a future compliance issue. It is becoming an operating issue now.

On the product side, the pattern is similar. Shopify is building for agent-led commerce, Salesforce is embedding AI deeper into workflow, and leading labs are shifting resources toward models with broader strategic value. The message is straightforward: this market is maturing quickly, but many organisations are still approaching it experimentally.

That gap matters. The current risk for most businesses is no longer missing AI entirely. It is adopting it in scattered ways, implementing badly, and creating cost, inconsistency, and governance problems that are harder to unwind later.

Why this matters for business

1. Review where you are overexposed to a single AI vendor
If a small number of providers continue to attract most of the capital and market power, their pricing, product priorities, and commercial terms will increasingly shape your roadmap. This week is a good reminder to review where you depend on one model, one cloud, or one workflow provider. Even a lightweight second-source option can improve leverage and resilience.

2. Treat AI governance as a live operating discipline, not a policy document
If your team is already using AI tools, governance needs to exist in the flow of work. That means clear rules on approved tools, what data can be used, who signs off external-facing outputs, and where human review is mandatory. The practical test is simple: could a team manager explain your AI rules in two minutes without opening a slide deck?

3. Check whether your digital buying journey is usable by agents as well as people
Shopify’s announcement matters because it changes the shape of discovery and conversion. Product data, stock availability, pricing, and checkout logic increasingly need to be machine-readable and consistent. A practical first step is to review your product catalogue and checkout flow with your digital team and ask what would break if the “customer” were an AI agent rather than a human.

4. Focus less on experimentation volume and more on operational usefulness
Many firms can point to pilots. Far fewer can point to repeatable time savings, faster response times, lower support effort, or better decision quality. If you want AI to matter commercially, ask each team for one workflow where time, quality, or cost can be measured before and after deployment. That will tell you far more than a long list of disconnected use cases.

This Week’s Policy & Regulation Brief

California Defies White House with AI Safety Executive Order

California has introduced new safety, privacy, and civil rights requirements for AI suppliers contracting with the state, creating a direct clash with the White House’s preference for lighter-touch, federally led rules. For businesses operating across the US, the practical issue is not politics but fragmentation. Compliance may now depend on customer type, geography, and procurement route, which makes flexible governance more important.

OpenAI’s $122B round includes $3B from retail investors

OpenAI’s latest financing round is notable not just for its size, but for what it signals about market expectations. Capital is continuing to flow toward a small number of frontier players seen as strategic infrastructure bets. For customers, that usually means faster product expansion, stronger ecosystem pull, and greater lock-in risk. It is another reason to keep procurement discipline and avoid unnecessary dependency.

Public Concern Over AI Hits New High

A new survey showing growing public concern about AI matters because adoption is not driven by capability alone. Trust, explainability, and perceived fairness increasingly shape how quickly businesses can deploy customer-facing AI without resistance. If sentiment keeps deteriorating, organisations using AI in service, hiring, or communication will need stronger messaging, clearer review processes, and more visible safeguards around how outputs are used.

OpenAI’s Sora Shutdown Reveals Shift to Enterprise and Robotics

The reported shutdown of Sora points to a broader commercial pattern: companies are reallocating resources away from expensive consumer AI products with uncertain monetisation and toward enterprise, infrastructure, and robotics-related opportunities. For business leaders, the takeaway is to watch where serious investment is being redirected. That often gives a better signal about where durable value is likely to sit than headline product launches do.

Nvidia’s AI Chip Dominance Faces Elon Musk’s Terafab Challenge

Nvidia remains the dominant force in AI compute, but new large-scale infrastructure bets show that rivals are trying to challenge that position through vertical integration and manufacturing scale. The practical takeaway for businesses is not to speculate on winners, but to recognise that compute availability, pricing, and concentration remain strategic variables. If AI is core to your roadmap, infrastructure assumptions should be reviewed regularly.

OpenRouter Data Highlights Rising Adoption of Lower-Cost Chinese AI Models

Usage trends showing growing adoption of lower-cost Chinese models highlight an uncomfortable truth for Western providers: price-performance is becoming a powerful competitive lever. For buyers, this creates opportunity and complexity at the same time. Lower-cost alternatives may widen access, but they also raise questions around security, governance, support, and geopolitical exposure. Procurement teams should be evaluating these trade-offs now, not later.

AI Designated Top US National Security Concern for 2026

The framing of AI as a top national security issue reinforces how closely commercial AI and state priorities are now linked. This has practical consequences for vendors and enterprise buyers alike, from procurement opportunities in defence and public sector markets to increased scrutiny over partnerships, data handling, and supply chains. Businesses in sensitive sectors should expect AI decisions to receive more board-level and regulatory attention.

Model & Platform Updates

OpenAI Shuts Down Sora, Redirects Team to “Spud” World Simulation Model

OpenAI’s reported move away from Sora toward a broader world-simulation effort suggests a shift from standalone media generation to models with wider strategic applications, including robotics and physical-world reasoning. The business takeaway is that labs are prioritising platforms with longer-term leverage over eye-catching consumer tools. That should influence how companies assess which AI categories are maturing and which may prove commercially fragile.

Anthropic Source Code Leak Reveals Claude Mythos (Capybara)

The reported Anthropic leak matters less for the codename and more for what it exposed: how much strategic value now sits inside model development, evaluation, and safety infrastructure. For businesses, it is another reminder that AI risk is not just about model outputs. It includes access control, configuration mistakes, internal tooling exposure, and operational security. Basic governance failures can create outsized consequences very quickly.

Salesforce Launches AI-Heavy Slack Redesign with 30 New Features

Salesforce’s Slack update shows where enterprise AI is heading in practice: not as a separate destination, but as a layer embedded inside existing systems of work. The important question for leaders is not whether Slack has more AI features. It is whether those features reduce effort, improve handoffs, or shorten decision cycles. If they do not, more functionality will simply mean more interface clutter.

Shopify Launches Agentic Storefronts and Universal Commerce Protocol

Shopify’s latest move is one of the clearest signs yet that commerce platforms are preparing for AI agents to act on behalf of buyers. That has practical implications for merchants now. Product data needs to be structured, inventory needs to be accurate, and checkout processes need to be easy for software as well as people to navigate. Businesses that prepare early should have an advantage.

NVIDIA Vera Rubin Enters Full Production

NVIDIA’s latest platform is designed around multi-step agentic workloads rather than conventional one-shot inference, which is a useful signal about where enterprise demand may be headed. More businesses are moving from content generation to task execution, orchestration, and automation. Leaders evaluating their own roadmap should ask whether their AI plans are still centred on chat interfaces when the market is moving toward more autonomous systems.

PrismML Launches 1-Bit “Bonsai” LLM Family for Edge Devices

If PrismML’s performance claims prove credible, the significance is clear: capable models running in far smaller memory footprints would expand what can be done on-device without depending on the cloud. That could matter for latency, privacy, resilience, and cost. For businesses, the near-term takeaway is to watch edge AI more closely, especially in environments where connectivity, security, or response speed make cloud-only approaches less practical.

Mistral Releases Voxtral TTS and Small 4 Open-Source Models

Mistral continues to strengthen its position as a serious open-weight alternative, combining practical model releases with infrastructure investment. For businesses, this matters because it broadens the set of viable options beyond the largest US labs. Open models can offer more control, pricing flexibility, and deployment choice, but they also place more responsibility on the customer for integration, evaluation, and governance.

Google Gemini 3.1 Flash Live Rolls Out Globally

Google’s latest Gemini release reflects a growing emphasis on real-time, multimodal interaction at lower cost. The interesting commercial angle is not novelty, but fit: these tools are increasingly being shaped for high-volume environments where responsiveness and cost control matter more than maximum model sophistication. For buyers, that makes use-case discipline even more important. The right model is often the one that is fast enough and cheap enough.

Quick Hits

• Ollama adds MLX preview for Apple Silicon - The popular open-source local LLM tool is now powered by Apple’s MLX framework on Macs, delivering significant speed gains for running AI models locally without cloud dependency. A meaningful step for on-device AI adoption.

• Meta unveils $499 Ray-Ban AI smart glasses for prescription users - Enhanced AI features including camera, vision, and voice capabilities. Meta continues to bet that wearables, not phones, will be the primary interface for everyday AI.

• AI Scientist v2 goes fully open-source - The autonomous research system that independently generates hypotheses, designs experiments, writes papers, and self-reviews had one of its AI-generated papers score in the top 45% at an ICLR workshop via blind peer review. A landmark moment for AI-driven scientific discovery.

• xAI co-founder departs, leaving Musk with only two original members - Ross Nordeen’s exit signals potential internal turbulence at Elon Musk’s AI lab, which continues to operate with a smaller leadership team than its frontier-lab competitors.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.
  

Executive Summary

This week was a useful reminder that the AI market is growing up.

The headlines were not really about smarter chatbots or bigger launches. They were about infrastructure, execution, and whether AI investment can actually stand up to commercial scrutiny. IBM’s Confluent deal was about data access. Anthropic’s latest move was about turning AI into something that can do work, not just answer questions. And investors punished Alibaba and Tencent for spending heavily without telling a convincing value-creation story.

For business leaders, that matters because the centre of gravity is shifting. The question is no longer whether AI is interesting. It is whether your business is making disciplined decisions about where to use it, which vendor to back, and what return you expect to see.

Why this matters for business

1. Most AI decisions now look like tool choices. They are actually operating model choices.

What looked like product news this week was really a reminder that AI decisions are starting to hardwire themselves into how businesses run. The platform you adopt shapes where work happens, how data moves, what your teams rely on, and how difficult it becomes to change course later. That means leaders should stop treating AI selection as a lightweight software decision and start treating it as part of core business design.

2. The real risk is no longer “doing nothing” or “moving fast”. It is implementing badly.

The gap is widening between businesses using AI with clear guardrails and businesses letting usage spread informally through teams. Poor deployment now carries real cost: messy workflows, duplicated spend, weak accountability, and avoidable data exposure. The advantage will not go to the company with the most tools. It will go to the one that is clearest on where AI adds value, who owns it, and what good looks like.

3. The winners will be the firms that connect AI to commercial outcomes, not internal excitement.

Investors and operators are both becoming less patient with vague AI ambition. The bar is rising. It is no longer enough to say the business is “using AI” or “exploring use cases”. The more important question is whether it is improving margin, increasing speed, strengthening customer experience, or creating some other measurable advantage. If that link is weak, the strategy is weak.

This Week’s Policy & Regulation Brief

IBM completes Confluent acquisition to strengthen real-time AI infrastructure

IBM completed its roughly $11 billion acquisition of Confluent on 17 March, bringing a major streaming-data platform into its stack. The strategic logic is straightforward: enterprise AI agents are only as useful as the live data they can access. For businesses, this is a reminder that AI value depends less on the model alone and more on whether your data can be connected, governed, and used in real time across systems.

Yann LeCun’s AMI Labs raises $1.03bn to back an alternative to LLM-first AI

Former Meta AI chief Yann LeCun’s new startup, AMI Labs, raised $1.03 billion to develop “world models” rather than rely on standard LLM approaches alone. The business implication is not that LLMs are being replaced tomorrow, but that the market is still open to different technical paths. Leaders making long-term AI bets should avoid assuming today’s dominant architecture will remain the only viable one.

Musk says xAI is being rebuilt from the foundations up

Elon Musk said xAI “was not built right first time around” and is being rebuilt after a period of senior departures and internal turbulence. That matters because frontier AI capability is only part of the story. Organisational stability, product focus, and execution discipline matter too. For enterprise buyers, this is another reason to assess vendor resilience, not just benchmark performance, before committing strategically.

Alibaba and Tencent lose $66bn as investors demand a clearer AI business case

Alibaba and Tencent shed around $66 billion in market value in roughly 24 hours after investors judged their AI monetisation plans too vague. The lesson is immediate for any leadership team funding AI programmes internally: spending alone is not persuasive. Markets and boards increasingly want to see a credible path from AI investment to revenue, margin improvement, or stronger competitive position.

Anthropic vs. Pentagon (Yes this is still ongoing)

The legal battle between Anthropic and the Department of Defence escalated significantly this week. Anthropic filed dual federal lawsuits challenging its designation as a “supply chain risk to national security”, a label previously reserved for foreign adversaries like Huawei. Court filings reveal the DoD told Anthropic negotiations were “nearly aligned” just one week before the designation was issued. More than 30 employees from OpenAI and Google DeepMind, including Google’s Chief Scientist Jeff Dean, filed legal briefs supporting Anthropic. The outcome will set a precedent for whether the US government can compel AI companies to remove ethical guardrails as a condition of doing business.

China’s Open-Source AI Strategy Is Working, Despite Chip Curbs

A US-China Economic and Security Review Commission report published 23 March warned that Chinese open-source AI models have narrowed performance gaps with Western frontier models despite US export controls. Roughly 80% of US AI startups are reportedly using Chinese open-source models. The commission flagged China’s embodied AI ambitions, particularly in manufacturing and robotics, as a distinct and underappreciated threat vector.

Model & Platform Updates

Anthropic gives Claude remote control of your Mac via Dispatch

Anthropic has started previewing a setup in which Claude Code and Claude Cowork can operate a user’s Mac, with Dispatch letting tasks be initiated remotely from a phone. It is still early and Anthropic stresses oversight, but the practical shift is important: AI is moving from chatbot to delegated operator. For businesses, that raises both productivity opportunities and new questions around permissions, auditability, and safe task boundaries.

MiniMax launches M2.7 with “self-evolving” training workflow claims

MiniMax says its new M2.7 model can automate parts of its own reinforcement learning workflow, including improving memory, skills, and elements of its training harness. Even if those claims prove only partly durable, the direction is important. Model developers are now trying to automate more of the improvement loop itself. For buyers, this signals faster iteration cycles and potentially shorter windows of competitive differentiation.

Google unveils TurboQuant to reduce AI memory overhead

Google Research introduced TurboQuant, a compression approach aimed at cutting the memory burden of vector quantisation, with early claims of major memory savings and performance gains in some tests. For businesses, the headline is not just research novelty. Techniques like this could reduce infrastructure cost and hardware constraints over time, making high-performance AI systems cheaper to run and easier to deploy at scale.

Google expands AI Studio’s full-stack app-building workflow

Google has been pushing AI Studio further toward a full-stack prototyping and app-building environment, including a new “vibe coding” experience, stronger project continuity, and updated billing options. The practical takeaway is that AI development tools are converging with lightweight software creation platforms. For teams experimenting with internal tools, this lowers the barrier to turning prompts into usable prototypes, but it also increases the need for governance over what gets built and deployed.

OpenAI this week

• OpenAI Kills Sora, Pivots to Enterprise “Superapp”

  OpenAI confirmed it is shutting down Sora, its video generation platform, just months after launch and a Disney licensing deal. The Sora team is being redirected to “world simulation” for robotics. Simultaneously, CEO Sam Altman announced OpenAI will consolidate ChatGPT, the Atlas browser, and Codex into a single desktop superapp. The moves are consistent with a company preparing for a Q4 2026 IPO at up to $1 trillion valuation. OpenAI’s Applications CEO Fidji Simo told staff there are “no more side quests.” The enterprise pivot is now official.

• OpenAI Acquires Astral - Developer Toolchain War Heats Up

  OpenAI announced the acquisition of Astral, maker of the widely adopted Python tools uv, Ruff, and ty. The Astral team joins OpenAI’s Codex platform, which has grown 5x since January to 2 million weekly active users. This mirrors Anthropic’s December 2025 acquisition of Bun (JavaScript runtime), confirming both labs are acquiring developer infrastructure to win the coding agent wars.

Mistral Launches Forge - Train Models from Scratch on Your Data

Mistral’s Forge platform, launched at GTC, allows enterprises to pre-train models from scratch on proprietary data, not just fine-tune. Early customers include ASML, Ericsson, the European Space Agency, and Singapore’s DSO. This is a significant capability for regulated industries that need full control over their AI training pipeline.

Gemini 3.1 Flash-Lite Targets Enterprise Scale

Google released Gemini 3.1 Flash-Lite at $0.25 per million input tokens, 2.5x faster time-to-first-token than prior models, with adjustable thinking levels to modulate reasoning effort by cost. A clear play for high-volume enterprise workloads where speed and cost matter more than raw capability.

Quick Hits

• OpenAI targets Q4 2026 IPO: Could be the largest tech IPO in US history, potentially at a $1 trillion valuation. All product decisions are now being filtered through IPO readiness.

• Bridgewater warns of $650B Big Tech AI capex in 2026: Co-CIO Greg Jensen flagged “significant downside risks” if execution falters, with parallels to the dot-com bubble, while noting AI spend is adding roughly 100 basis points to US GDP growth.

• Anthropic safety researcher quits, warns “world is in peril”: Mrinank Sharma publicly resigned from Anthropic citing bioweapons concerns. A notable signal from inside the lab widely regarded as the most safety-conscious.

• Mark Zuckerberg building personal “CEO agent”: WSJ reported Zuckerberg is testing an AI agent to replace layers of executive information retrieval. Meta employees have built hundreds of internal AI tools, with engineer output reportedly up 30% since early 2025.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.
  

Executive Summary

This week’s AI news was less about flashy launches and more about where the market is actually heading. Three signals stood out.

First, the AI race is hardening into an infrastructure contest. Nvidia’s latest announcements reinforced that value is concentrating not just in models, but in the compute, systems, and platforms needed to run them at scale. At the same time, Meta’s reported workforce cuts and infrastructure spending show how seriously large companies are reorganising around AI economics.

Second, AI agents are moving closer to real enterprise workflows, and that raises the stakes. Stories involving internal chatbot compromise, sensitive data exposure, and classified-data training plans all point to the same issue: once AI systems can access tools, files, and internal systems, governance matters as much as model quality.

Third, the market is shifting from experimentation to implementation. Microsoft, Anthropic, OpenAI, Perplexity, and Google are all pushing toward more persistent, embedded, workflow-oriented AI. The question for businesses is no longer whether AI is useful. It is where to deploy it safely, which processes to redesign around it, and how to capture value without creating cost, legal, or security problems.

Why this matters to businesses

• AI is becoming an operating model issue, not just a software tool.
  The competitive advantage will not come from “using AI” in the abstract. It will come from redesigning workflows, decision rights, and team structures so AI can improve output, speed, and consistency in commercially meaningful areas.

• Governance now needs to catch up with capability.
  As AI tools gain access to internal data, systems, and actions, businesses need clearer rules on permissions, oversight, data boundaries, and human approval. The risk is no longer just inaccurate answers; it is inappropriate access, automation without control, and exposure of sensitive information.

• Vendor choice is becoming a strategic decision.
  The market is fragmenting between model providers, infrastructure players, and workflow platforms. Businesses should avoid chasing every release and instead decide which vendors fit their security requirements, integration needs, and long-term AI architecture.

• Cost discipline will matter as much as innovation.
  The leaders in AI adoption are increasingly treating it as a capital allocation and productivity question. Businesses should focus on use cases where AI can reduce friction, improve margins, or strengthen decision-making, rather than funding broad experimentation without a clear path to return.

• The winners will be the firms that move early, but selectively.
  This is not the moment for blind acceleration or passive observation. It is the moment to identify the few high-value processes where AI can be deployed safely, prove impact quickly, and build the internal governance to scale from there.

This Week’s Policy & Regulation Brief

Morgan Stanley Warns of Imminent AI “Breakthrough Shock”

A widely circulated Morgan Stanley report warned that a step-change in AI capability, driven by a 10x compute ramp at leading labs, will arrive between April and June 2026, and that governments, corporates, and workers are “not prepared.” The report frames AI as a macro “intelligence currency” risk, arguing that self-reinforcing improvement cycles could begin by 2027. The bank’s survey of 1,000 executives across five countries found a 4% net workforce reduction over the past 12 months directly attributable to AI, with 30% of early AI adopters reporting quantifiable productivity gains by end of 2025, up from 16% a year earlier.

AI Agent Hacked McKinsey’s Internal Chatbot in Two Hours

Security startup CodeWall said its autonomous AI agent breached McKinsey’s internal AI platform, Lilli, and gained read-write access in about two hours by exploiting exposed APIs and weak access controls. Reports say the system could access millions of internal messages before McKinsey patched the flaw. The story matters because it shows agentic AI can move from assistant to attacker fast when connected to poorly governed enterprise systems.

ByteDance Suspends Video AI Model Over Copyright Disputes

ByteDance halted the rollout of a new video-generative AI model after facing copyright challenges, underscoring rising legal risk for models trained on or remixing copyrighted content. The suspension highlights the increasingly fraught relationship between generative AI capabilities and intellectual property law, a tension that is only going to intensify as video models mature.

The Pentagon is planning for AI companies to train on classified data, defense official says

The Pentagon is exploring frameworks that would allow select AI companies to train models on classified data within secure government-controlled environments. The initiative reflects growing interest in deploying advanced AI for defence applications, including intelligence analysis and operational planning. However, the approach raises significant concerns around data security, model control, and compliance, particularly regarding how sensitive information is handled, retained, and potentially embedded in commercial AI systems.

Meta Plans Up to 20% Workforce Cuts as AI Spend Surges

Meta is preparing workforce reductions of up to 20% as it reallocates capital toward AI infrastructure and compute. The move aligns with broader industry trends, where large tech firms are funding multi-billion AI investments through headcount cuts. The company is prioritising efficiency, automation, and AI-led productivity, signalling a structural shift in how big tech balances growth, cost, and capability.

Meta Quietly Retreats from Metaverse Ambitions to Refocus on AI

Meta is scaling back its metaverse push, deprioritising elements of its Reality Labs strategy as AI becomes the company’s central focus. While not formally “cancelling” the metaverse, investment and strategic attention have shifted decisively toward AI models, infrastructure, and products. The pivot reflects weaker-than-expected adoption of virtual worlds and stronger near-term returns from AI-driven services and advertising.

Anthropic vs. Pentagon Standoff Continues

The legal confrontation between Anthropic and the Pentagon remains active. Defense Secretary Pete Hegseth terminated the Pentagon’s relationship with Anthropic on 1 March after CEO Dario Amodei refused to allow unrestricted use of Claude for mass surveillance and autonomous armed drones. The Department of Defence issued a “supply chain risk designation”, the first time an American company has received this classification. Anthropic has vowed to challenge it in court. President Trump has instructed most government agencies to stop using Anthropic’s AI, while granting the Pentagon a six-month phase-out period.

Model & Platform Updates

Microsoft Launches Copilot Cowork on Anthropic’s Claude

Microsoft’s Copilot Cowork, a long-running, multi-step enterprise AI agent built on Anthropic’s Claude, launched in limited research preview. It integrates across Microsoft 365 (Outlook, Teams, Excel, SharePoint) and is powered by a “Work IQ” system for full enterprise context. This marks a significant architectural shift: Microsoft 365 Copilot is now explicitly multi-model, running both OpenAI and Anthropic systems. Broader rollout via the Frontier programme begins late March. Pricing is included in the existing $30/user/month M365 Copilot plan.

Anthropic Enables Mobile Control of Claude “Coworker” Sessions

Anthropic is introducing the ability to access and control Claude “coworker” sessions from mobile, allowing users to monitor and interact with long-running AI tasks remotely. This extends Claude beyond chat into a persistent, task-executing agent that continues working in the background. Mobile access means users can review progress, provide input, and steer workflows on the go, reinforcing the shift toward always-on, agentic AI integrated into daily work.

OpenAI Updates this week

• OpenAI Brings Sora Video Generation Into ChatGPT

  OpenAI began integrating Sora-class video generation directly into ChatGPT starting 14 March, blurring the line between conversational AI and full video creation tools. The move raises significant questions about inference costs (projected in the hundreds of billions of dollars over coming years), safety controls, and platform liability for AI-generated video content at internet scale.

• OpenAI Demos GPT-Realtime-1.5 for Voice
  OpenAI demonstrated GPT-Realtime-1.5, an updated voice model aimed at making spoken AI interactions faster and more fluid. The release extends competition in real-time multimodal assistants and positions voice as a core interaction layer rather than an add-on.

• OpenAI Enhances Platform for Agentic Workflows

  OpenAI has reportedly upgraded its platform to support models carrying out more multi-step tasks autonomously without constant human prompts, reinforcing the shift toward agentic workflows and reducing the friction between instruction and execution.

• Introducing GPT-5.4 mini and nano

  OpenAI launched optimised smaller variants of GPT-5.4 tailored for coding, tool use, and agentic workflows, reducing latency and cost while supporting high-volume, multimodal applications.

Perplexity’s new answer to OpenClaw

Perplexity has launched Personal Computer, an always-on AI agent that runs on a dedicated Mac mini and connects local files, apps and active sessions with Perplexity’s cloud orchestration layer. The pitch is enterprise control: sensitive actions require approval, sessions include audit trails, and users get a kill switch. More broadly, it pushes AI agents beyond chat into persistent, remotely supervised digital workers that keep operating after the user steps away.

Google updates Stitch with AI-native design canvas and reasoning agent

Google has evolved Stitch from a prompt-to-UI generator into what it calls an AI-native software design platform. The new update adds a design canvas and reasoning-driven workflow so users can create, refine and collaborate on interfaces more iteratively, rather than generating one-off mockups. The significance is less about faster wireframes and more about compressing the path from app idea to interactive prototype, especially for non-technical builders.

Nvidia GTC 2026: Full-Stack AI Infrastructure Overhaul

Jensen Huang’s GTC keynote in San Jose (16 March) was the week’s defining product announcement. Nvidia unveiled the Vera Rubin platform, seven new chips, five rack-scale systems, and a new supercomputer architecture for agentic AI, targeting 10x performance-per-watt improvement over Grace Blackwell. The Groq 3 LPU debuted as the first commercial product from Nvidia’s $20B Groq acqui-hire, promising a 35x boost in tokens-per-watt when paired with Rubin racks (available Q3 2026). Nvidia also previewed the Kyber architecture, 144 GPUs arranged vertically for next-generation density, expected 2027, and released NemoClaw, a reference toolkit for building enterprise-ready AI agents using the OpenClaw open-source agent framework.

Quick Hits

• Sam Altman frames AI as a metered utility: Speaking at the BlackRock Infrastructure Summit, Altman said AI will likely be delivered like electricity, with users billed by intelligence or compute consumption. The framing could reshape how hyperscalers, regulators, and enterprises think about access, pricing, and antitrust in AI markets.

• Meta’s Rogue AI Agent Exposes New Enterprise Risk: Meta is investigating an internal AI agent incident after a system reportedly shared sensitive company information, and potentially some user-related data, with employees who were not authorised to see it.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.