Good morning,
This week we cover how Nvidia stopped being merely the arms dealer of the AI boom and became its banker. Then, a “first” (or is it?): OpenAI has paused its own frontier model on security grounds. Plus the quick bites, one listen on what happens when millions of agents meet, and a closer look at the tool behind the fortnight's biggest acquisition.
Nvidia, from arms dealer to banker with billions to invest
Nvidia will report on 26 August, but it has not waited to create excitement. In the last two weeks, the company has done its best to make that print the most consequential of the year, not through chips, but through finance.
On 10 August, Nvidia signed agreements with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise over $500bn of third-party capital for AI infrastructure, offering to backstop up to $125bn of it itself. Jensen Huang's pitch is that compute is now "an investable asset" - a new asset class, like toll roads or pipelines. A week later, an SEC filing revealed Nvidia will guarantee up to $105bn of lease and power obligations for a single OpenAI data centre in Pike County, Ohio, 8 gigawatts of compute, built and owned by SoftBank's SB Energy, in which Nvidia is also investing $1.5bn as exclusive chip supplier. And on 21 August, The Information reported Nvidia is paying ~$6bn to license coding startup Poolside's software and hire 109 of its staff, following similar arrangements with Groq and Enfabrica.
Spot the pattern: Nvidia finances the customer, the customer buys Nvidia chips, and the revenue books either way. The polite term is vendor financing; the impolite one is circularity. Bill Dudley, formerly of the New York Fed, wrote in Bloomberg that the AI boom won't prevent a sharp market correction; Apollo's own chief economist, Torsten Slok, put the problem plainly: capital can bridge the gap between AI capex and AI revenue for a while, but not indefinitely. Meanwhile server makers quietly told customers that some flagship Nvidia systems will cost ~17% more — scarcity pricing, for now.
Which brings us back to Tuesday. Consensus expects roughly $92bn of revenue (up ~96% YoY) with data centre above $85bn, almost exactly Nvidia's own prediction. A comfortable beat and strong guide extends the runway for all of the above. A wobble, and every one of these self-referential financing structures gets re-examined at once. But given all the recent announcements, it is quite likely that Tuesday will be big.
Why it matters: Treat Tuesday's Nvidia print as a planning signal, it is the closest thing to a stress test the AI capex thesis will get this year. And when a vendor offers to finance your AI infrastructure, read it as what it is: the vendor taking your demand risk onto its own balance sheet.
Astra, the model that was too good to ship (again)
Last edition, Astra was the feel-good story: an internal OpenAI model solving ten long-open mathematics problems for about $2,000 of compute. This edition, it is something rarer — supposedly the first frontier model a major lab has publicly held back on security grounds. OpenAI says it cannot rule out that Astra crosses the "Critical" cybersecurity threshold in its Preparedness Framework — the capacity to autonomously find and exploit unknown software vulnerabilities — and has put its largest planned training run on hold while it rewrites that framework. The same reasoning that produces machine-checkable proofs also, it turns out, produces rather good zero-day exploits.
Before applauding, note the choreography. A US frontier lab announces it has built something too powerful, places it solemnly in a cage — and weeks later hands out the keys. OpenAI wrote the script with GPT-2 in 2019; Anthropic performed it this spring, when Mythos went from too risky for public release in March to everyone's Fable 5 in June. Perhaps staged release is the safety process working; but "too dangerous to ship" is also the best marketing ever devised, and the cage has yet to hold a model for long. Expect Astra to follow.
The regulators, meanwhile, have arrived anyway. The EU AI Act's enforcement powers over general-purpose models are live, with fines up to €15m or 3% of global turnover, and Article 50's transparency duties apply now: chatbots must disclose they are chatbots, AI-generated content must carry machine-readable labels (Anthropic will watermark Claude outputs worldwide to comply). Wanna test if you are compliant? Use the checker tool. Washington, by contrast, is sticking to a voluntary early-access framework with no licensing power behind it.
Why it matters: If you deploy any customer-facing AI in Europe, Article 50 is enforceable today: confirm your chatbots self-identify and your synthetic content is labelled with EU checker.
Quick bites
Deals & money
Anthropic is preparing super-voting founder shares ahead of a potential IPO. Dario Amodei and co-founders would retain control post-listing. The IPO is expected to value Anthropic at $2tn. OpenAI, confidentially filed since June, reportedly leans towards 2027.
Alibaba expects AI-related revenue to hit a $10bn run-rate by September, up from $7.3bn, with cloud-and-compute growing 45% and capex up 75% YoY; a datapoint that Chinese enterprise AI is monetising.
Unitree rose 460% on its Shanghai debut, valuing the humanoid-robot maker at roughly $50bn — mainland China's first listed humanoid company, oversubscribed thousands of times, with DeepSeek among its investors. The robots themselves remain mostly demonstration-grade in their “GPT-2 era”.
Security firm Obsidian raised $85m at a $1.1bn valuation in August on the back of one striking datapoint: nearly 70% of its clients now let AI agents interact directly with business data.
Nvidia is in talks to fund data supplier Mercor at a $20bn valuation. And Stripe agreed to buy model marketplace OpenRouter for a reported ~$7.5bn — more at the end.
The cost fight, continued
AT&T plans to hold its Anthropic and OpenAI spending flat by shifting its ~45bn daily tokens to open models, a quarter of its usage today. Its chief data and AI officer plans for open models to power 70 to 80% of AT&T's total usage.
The labs are moving up the stack. Anthropic and OpenAI are building industry-specific applications that compete with their own customers, and Anthropic's enterprise venture Ode (a JV with Blackstone) made its first acquisition, consultancy Casper Studios.
Google claims its AI can replace "forward-deployed engineers", the Palantir-style consultants OpenAI, Anthropic and Microsoft have spent billions hiring.
shared.image.missing_image
Adoption watch: AI beyond the frontier
IBM signed a sweeping partnership with OpenAI on 13 August, training thousands of consultants to deploy AI across core operations. Everyone, it seems, now wants to sell the shovelling rather than the shovel: Microsoft ($2.5bn, 6,000 engineers), OpenAI's own DeployCo and now Big Blue are all betting the money is in implementation, not models.
Gartner 2026 Hype Cycle for Agentic AI finds only 17% of organisations have fully deployed AI agents and predicts over 40% of agentic projects will be cancelled by end-2027 on cost and unclear value.
A US mortgage lender that promised Salesforce's conference it would return with 100 agents running spent weeks stuck on a single integration, despite consulting every architect and forward-deployed engineer it could find.
Science & society
DeepMind open-sourced WeatherNext, its cyclone model, after a Nature paper showed state-of-the-art track and intensity forecasts, giving on average an extra day of warning.
Isomorphic Labs, DeepMind's drug-discovery spinout, still targets first human trials of AI-designed drugs by year-end, with one compound already FDA-cleared. But will AI really cure everything? Scientists are doubtful.
OpenAI launched ChatGPT for Teens, auto-enrolling users its systems predict are 13–17 into a restricted experience.
Is AI actually taking the graduate jobs? Stanford's Digital Economy Lab's new paper finds employment for 22–25-year-olds in the most AI-exposed occupations down ~19% relative to peers since late 2022; but another report from Ramp and Revelio Labs (21,599 firms) earlier this year finds the heaviest AI adopters grew entry-level headcount 12% post-adoption. Both can be true — the effect is real but concentrated among the youngest workers, and its direction depends on how you measure "exposure".
The interview: when millions of AI agents meet
This week I want to share my love for Google DeepMind: The Podcast hosted by Hannah Fry. It is in itself already worth a subscription in general for direct access to frontier researchers, but specifically I really enjoyed their recent episode “When millions of AI agents meet” (42 minutes). Most agent talk stops at “what can one agent do for me”; this one asks the strategy question underneath — what happens when agents transact with other agents at scale. It walks through delegation between agents, security and “traps”, the design of an agentic economy, and a risk your risk register almost certainly lacks: cognitive monoculture, millions of agents making correlated mistakes because they share the same model. A very interesting scientific exploration of what could come.
Tools to try: OpenRouter
Last week, Stripe agreed to acquire OpenRouter for a reported ~$7.5bn — more than five times the valuation it raised at in May. At such a price, it is perhaps worth understanding what OpenRouter is, so let’s.
OpenRouter is a single door to the whole model market: one account, one API, one bill, and behind it 400+ models from OpenAI, Anthropic, Google, Mistral, the Chinese open-weight labs and dozens more. Sign up, load a few dollars of credit, and in the chat interface you can put the same prompt to several models side by side. A super fast way I know to test whether a cheaper model handles your actual task before anyone signs anything. For developers, it means switching models is a one-line change rather than a procurement cycle; requests can be routed to whichever model offers the right mix of price, speed and quality, with automatic fallbacks if a provider goes down. Even if you never send a token through it, bookmark the rankings page: it shows which models real applications actually pay for, in near real time — a far better market signal than any benchmark, and the reason we knew Chinese open-weight models were taking share before it made the papers.
Why it matters is the position, not the product. OpenRouter sits in the transaction path of a meaningful slice of the world's AI usage — it sees what everyone builds with, what they pay, and what they switch to. That vantage point is what Stripe bought, and it is the same logic as owning the harness we discussed two weeks ago: whoever controls the routing layer holds the pricing leverage, not only the model vendor. Two caveats for corporate use: routing traffic through an intermediary adds a data-governance hop your compliance team will want to examine, and a marketplace's neutrality is easier to trust when its owner isn't also processing payments for the labs it ranks. Worth trying regardless (with test data first).