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