AI-Proof - Weekly AI Pulse
A concise summary of the week’s most important AI developments
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.