The Week in AI — October 26–November 1, 2026
Two prices moved this week, in opposite directions. The price of money went up: OpenAI's round halved, memory is sold out a year forward, and lenders stopped accepting a GPU rental contract as collateral. The price of trust went down, which is the harder problem. "The signals we used to rely on as markers of a mature system are now easier to counterfeit," Kyle Kingsbury wrote in his Jepsen analysis of ParticleDB, and every story below is a version of that sentence.
1. OpenAI asked for $30 billion and is getting $20 billion. The Financial Times reported Thursday that the round's target is now about $20 billion at a $1.25 trillion valuation, after some investors cut their commitments in the AI selloff, per the FT. Hours earlier the Wall Street Journal had put the amount actually committed at roughly $11 billion — less than half the original ask. Two papers, one evening, agreeing on the valuation and disagreeing on the size, is what a market looks like when it reprices in public.
Why it matters is not demand. OpenAI's run-rate revenue was $40 billion in August, up about 70% since July; the customers are there. What moved is the price of the capital that funds the compute. A bridge round that lands at half its target leaves the company three options it does not like: raise again sooner, raise at a lower number, or spend less than it told its suppliers it would. The same week, S&P cut its outlook on Meta to negative over capex and negative free cash flow, and the neoclouds were effectively shut out of the bond market entirely — we laid out how the credit channel closed while SpaceX borrowed $40 billion against a satellite business. The build-out has not slowed. Its financing has.
2. Samsung's 2027 memory is already sold out, and it is spending $63 billion to add more. Samsung posted 107.4 trillion won of operating profit for July to September, up 783% year on year, the first 100-trillion-won quarter any technology company has recorded, according to its earnings deck. On the call, memory executive vice-president Jaejune Kim said 2027 high-bandwidth memory supply is sold out at agreed prices, with customers asking for more than the contracts cover. The stock fell as much as 6.6% in Seoul anyway.
The two facts sit together badly, and that is the story. Sold-out forward supply at fixed prices caps the upside while 84.4 trillion won of capex lands as depreciation whether or not the AI build-out keeps its pace — fixed price, rising fixed cost. Kim's own explanation of why the money will not loosen the shortage soon is the part worth keeping: much of it buys cleanroom space rather than tools, construction costs per unit of space are higher than in past cycles, and equipment lead times are stretched by the same shortage everyone is buying their way out of. The full read is here.
3. Google started selling TPUs outside its own walls, and Nvidia's biggest customer lined up to buy. Oracle will offer Google's tensor processing units through Oracle Cloud Infrastructure starting in calendar 2027 — the first time TPUs run on a hyperscale cloud other than Google's, Oracle announced at its AI World conference. Hours later The Information reported that OpenAI plans to buy TPUs directly, and both Bloomberg and CNBC said they had confirmed it.
For ten years the TPU was the one advantage Google refused to sell. What changed is not the chip but the distribution: Google concluded it can make more money selling silicon through partners than hoarding it for Gemini. The tell that this is real rather than a press release is dull — OpenAI has posted a job for an engineer to lead optimization of its frontier models on TPUs, which nobody writes unless there are racks coming. Nvidia's moat was never only the chip; it was the decade of software and the certainty that everyone bought the same thing. A second lane just opened, and the interesting question for 2027 is who prices against whom.
4. A tampered Claude Code extension sat in Cursor's marketplace for eleven hours. Versions 2.1.312 and 2.1.313 of the anthropic.claude-code package shipped to macOS and Windows users from about 2:27am to 1:35pm Pacific on October 29, per Cursor's incident notice. The payload fingerprinted the machine, swept environment variables for anything named key, token, secret or password, appended an attacker key called svc-telemetry to the SSH authorized-keys file, and opened a reverse tunnel that retried about once a minute.
What stopped it from becoming an SSH backdoor into a large number of developer machines was not a control. It was two bugs in the malware. That is the finding: the supply chain for AI coding tools is currently defended by the incompetence of strangers, and nobody in the chain — not the publisher, not the marketplace, not the signing system — has proposed a defense that works against a stranger who is competent. Eleven hours of exposure is also an inescapable measurement of review capacity, because a marketplace that cannot read a diff before it ships is a marketplace that is trusting, not vetting. Our take on why luck is not a security model.
5. A downloadable model broke six Firefox releases at about $1,000 a version. Epoch AI gave GLM 5.3, Z.ai's open-weight model, a target Firefox build and nothing else — no vulnerability reports, no insider information, no human steering the search. It compromised every release shipped between September 17 and October 26 by reading fixes that had been published but not yet installed everywhere, chaining a bug derived from a public patch diff to a second sandbox bypass it rediscovered on its own, per Epoch's report. The end state was a webpage that read a file off a visitor's machine and sent it to an attacker without user interaction.
The reflex is to reopen the open-weights liability argument, and the reflex is wrong. The fixed cost was roughly $100,000 of research and compute to build the harness once; after that, each target was a credit-card transaction and a waiting period. What got cheap is not expertise but the window between a diff becoming public and a patch becoming universal — six days on a weekly update cycle, guaranteed. Vendors hold that lever and have spent their effort on the first half of the gap while treating the second as the customer's problem. On the evidence: the same harness against Chrome produced nothing, so anyone legislating off six Firefox versions is legislating off a demonstration, not a trend. The patch gap is a calendar problem.
What to watch next week
Whether OpenAI's round closes at all, and at what price. About $11 billion of a $30 billion ask was committed as of Thursday. A close at $20 billion and $1.25 trillion is the market's new mark for frontier labs; a close well below it forces the company to choose between another raise and cutting the compute commitments it has already made to Oracle, Broadcom and the TPU side of the house.
The neocloud debt wall, now with a guarantor. CoreWeave, Nebius and Nscale have been shut out of the bond market, CoreWeave-tied paper priced at a 9.25% yield, and Nvidia is reportedly in talks to backstop Nscale's West Virginia construction debt with a residual value guarantee — a chip supplier cosigning a building. Watch whether that instrument becomes the template, because a guarantee that moves risk onto Nvidia's balance sheet is a different kind of AI trade than a chip order.
Agents, measured from both ends. DeepSeek published Super-Agent, an MIT-licensed coding agent that passed 5,900 stars in about thirty hours and, unlike the coverage suggests, ships defaulting to observe rather than act — the gap between the pitch and the code is the useful part. Against that, DoorDash told analysts third-party agents have placed "several hundred" orders in total, against 993 million human ones in the quarter. The agent-commerce thesis now has a number on each side; next week is about which one moves.
If the update window, not the model, is what makes exploits cheap, should browser vendors be required to report how fast a fix reaches users? Write to us via our contact page.
