AI stocks fall a second day, Anthropic faces a likely junk rating, OpenAI's raise falls short: money for the AI buildout gets tighter

Anthropic heads for a junk rating and Nvidia may cosign its landlord's debt: the chipmaker becomes the lender

Lead with the debt, because that's where the change is. Microsoft, Alphabet, Amazon, Meta and Oracle have raised $326 billion in debt this year, triple last year's total, according to S&P Global figures cited by TechCrunch. Alphabet's free cash flow was negative $9.1 billion in the third quarter, its second straight quarter in the red. Amazon's was negative $42.4 billion over twelve months. Meta's was negative $2.1 billion. Microsoft was the only one in the black, and only because third-party capacity arrived late. That delay knocked down the neocloud stocks. CoreWeave, Nebius and IREN had each fallen more than 30% in October.

The labs sit downstream of all this. A BB rating, the top of junk, would make every future compute deal Anthropic signs more expensive to finance. The Nscale story shows the workaround. Anthropic has signed a $44.6 billion contract for the first 460-megawatt building of Nscale's West Virginia campus and is discussing a second, similar deal. But Nscale had about 25,000 of a planned 500,000 chips running in August. Construction runoff has flooded nearby homes more than once. And Microsoft pulled out of an Nscale site in England this month. Under the arrangement The Information reported, Nvidia would back the construction loans: if Nscale defaults, Nvidia takes over the building and finds a new tenant. Nvidia's past support has covered leases and chip purchases, never construction. It has now extended roughly $170 billion in investment and credit to neoclouds and big customers. My read: the closer the chip seller gets to the borrower's balance sheet, the more any slowdown at one AI lab becomes a problem for everyone in the chain.

OpenAI's shortfall is the thinnest-sourced story of the day. All I have is the headline and a Hacker News thread, and the page itself couldn't be retrieved. Commenters put the gap at about $15 billion and argued over what it means. One said it says more about "the supply and demand of capital in any given moment" than about OpenAI's prospects. Others tied it straight to pricing. Their logic: if the labs can't raise prices without pushing customers to open weights, and can't raise unlimited capital either, margins have to come from somewhere. One commenter said their company's token bill fell from about $1 million a month to barely half that after Sonnet 5.5 shipped. That's one anecdote, but it fits.

The buildout isn't stopping. Cohere restated a C$4 billion Canadian research and infrastructure commitment through 2031. Runware, a Romanian startup, says it will build modular data-center pods on an assembly line and reach 30 megawatts a day by 2028. That's a plan backed by a $50 million Series A, so treat the number as an ambition. Google's Suncatcher satellite is running Gemma 4 on a TPU in orbit, though power limits mean it can run only one of its four chips, for about 20 minutes at a time. What has changed is who pays and on what terms.

There is no winner-take-all. So models are already a commodity, and will remain one for years. Companies must balance capex with very long payback periods against strict austerity everywhere else.

— esFran, Hacker News

Alphabet loses $500B in a day over AI search: the market now prices in that AI eats its own revenue

Thursday's brief covered Alphabet's numbers. What's new is how the market answered. Alphabet fell 12% on Thursday, its worst day on record, and kept falling Friday. The trigger was an admission on the earnings call: each AI Mode query earns less than a comparable web search, because it shows fewer ads that cost less. Search fell below half of Alphabet's revenue for the first time, at 49.7%. Partly that's because Cloud grew 103%, but Search revenue was flat quarter over quarter.

The damage spreads past Google. Reddit fell 29%. It said search referrals dropped by more than a million a week in October, and announced a move to a mostly logged-in site, partly to protect its AI data-licensing business. Roblox fell 18%. Intel fell 28% for a different AI reason: memory demand from data centers has pushed PC prices up, and customers are buying fewer PCs. That ties to Sam Altman's remark this week that it's "quite terrible" for the world's DRAM to sit in personal computers running at five percent use. His argument is right on utilization. It's also the case for why your next laptop costs more.

The top Hacker News comment on the sell-off pushed back on the "AI cost fears" framing. Its view: what spooked investors was mostly lost attention and lost clicks, not capex. A small-business commenter put it in concrete terms. They used to spend about $60,000 a year on Google Ads, dropped it to zero, and are now testing OpenAI as a channel instead.

Google ad revenue has dropped. AI results from search engines don't generate the click-through revenue that search ads did. That's scary for Google management.

— Animats, Hacker News

House Democrats file a no-bailout resolution and Maine votes on data centers: AI finance becomes a campaign issue

Thirteen House Democrats, led by Progressive Caucus chair Greg Casar and Financial Services ranking member Maxine Waters, introduced a resolution opposing emergency lending, capital injections or golden parachutes for insolvent AI firms. It binds no one. But it puts on record the claim that AI companies hold "trillions of dollars in debt, much of which is raised in opaque and unregulated private credit markets." A week ago that would have read as fringe. After Friday it reads as context.

The Hacker News thread, the most discussed of the day, split three ways. The top comment was blunt: "Privatize the losses." A more careful reply argued the resolution's value is procedural. The 2008 bank bailout passed in about two weeks, and the commenter said putting a position on record now makes it harder to be rushed into one later. A self-described lab insider called the whole thing theater, on the grounds that inference "prints" money. Nobody in the thread offered numbers either way.

The same politics is playing out on the ground. In Maine's Senate race, which will help decide control of the chamber, data centers have become a top issue. A September poll found 76% of Mainers oppose one in their community, and towns have passed temporary moratoriums. The tech industry is also among the race's biggest funders. Builders should expect permits, power and siting to get harder in 2027, whichever party wins.

The last time we did large scale bailouts (TARP[0], in 2008), It was rushed through and ratified in around two weeks following the sudden collapse of Lehman. Any measures we can take to ensure that we don't fall into a situation where there is only one choice is sensible.

— bushido, Hacker News

Claude Code extension backdoored on Cursor, Northflank leaks AI keys: the agent toolchain is now the target

Version 2.1.313 of Anthropic's official Claude Code extension went out on Cursor's marketplace through Anthropic's own verified account. It installed an attacker's SSH key, turned on the local SSH service and opened a reverse tunnel, on Windows, macOS and Linux. Cursor auto-updates extensions by default. It says at most 5,000 users got the bad version before Anthropic pushed 2.1.314 several hours later. As of Ars Technica's report, Anthropic had not notified customers, and no one had explained how the account was compromised. Extensions run with the IDE's permissions, so an infected machine exposed its source code, local files and any cloud credentials a developer had loaded.

Northflank's breach the same day fits a pattern. Users reported strangers spending their AI API keys. The platform told everyone to assume all environment variables had been decrypted, and pointed to a Wiz report it called "Black October" on attacks aimed at AI credentials. The top comment in the Hacker News thread counted at least six fast attacks on multi-tenant platforms since September: PingCAP/TiDB, Fireworks AI, Zeabur, OpenCode Console, Railway and now Northflank. The most-replied comment predicted that enterprises will move back to the hyperscalers, since convenience can't make up for weak security.

The practical lesson is simple. A developer laptop running a coding agent now holds the keys to production and to a metered AI bill. That makes it more valuable to attackers than most servers. Pin extension versions, turn off auto-update for anything that runs with IDE permissions, and give agents scoped, short-lived keys. In the DOJ's first broad "ghost student" crackdown this week, prosecutors said fraud rings used AI to apply for aid, do coursework and fake attendance in online classes. That's the same shift aimed at a softer target.

I count at least six substantial, lightning-speed attacks against multi-tenant as-a-service vendors since September. Not this year--just since last month.

— jonathaneunice, Hacker News

Four labs adopt one robots.txt opt-out token as Wikimedia splits training from regurgitation in court

OpenAI, Google, Microsoft and Anthropic will honor a single robots.txt "policy token" for AI training. An Anthropic standards lead proposed it at the IETF. Anthropic already honors it. OpenAI says it will within months, and Google and Microsoft next year. Meta said it supports the idea but didn't commit to honoring it. Microsoft will let sites opt out of training and of grounding separately, without dropping out of Bing. That distinction will matter more than the headline: grounding is where the referral traffic, and so the money, now goes.

The courts are drawing the same line. Wikimedia's amicus brief in the New York Times case against OpenAI and Microsoft backs neither side. It argues that scraping and pre-training should be fair use, but serving copies of a work in reply to a user's query, in a way that replaces reading the original, should need a license. For Wikipedia, that license is Creative Commons, which requires attribution. The same day, OpenAI posted excerpts from friendly court briefs under the title "Fair use benefits everyone." My bet is that courts land near Wikimedia's split: training protected, regurgitation and substitution licensed. If so, retrieval and grounding become the licensing battleground. Reddit's logged-in move is an early bet on exactly that.

Altman says OpenAI spent hundreds of millions on math and holds back more results: discovery becomes a lab budget line

On the Cheeky Pint podcast, Altman said OpenAI has spent on the order of hundreds of millions of dollars on math problems. That includes about $10 million, "maybe a little more," on its Navier-Stokes result. He confirmed the company has more proofs it hasn't released. This month OpenAI put out hundreds of results with thin writeups. Mathematicians have complained to The Verge about poor attribution, little context and the burden of checking the work. OpenAI is also withholding the details of a claimed Yang-Mills result on the advice of experts it picked itself. That cuts against Altman's own line that humanity "deserves fundamental knowledge."

The cost number is the part to remember. At $10 million a problem, frontier math is now a budget decision for a lab, and academic groups can't compete on that budget. With labs being asked to justify every dollar, I'd watch whether this spending gets defended as research or as marketing.

A researcher rebuilds Meta's unpublished Muse "Charter" from its chain of thought, and it borrows heavily from Claude's

A LessWrong author noticed that Meta's open-weights Muse Spark Open keeps citing a "Charter" in its reasoning. Working from 884 transcripts across 260 prompts, they reconstructed much of it. Meta's public reports mention only an "early internal behavior specification." The reconstructed text calls Muse a cultural artifact "made of civilization." It tells the model to share evidence rather than bare confidence levels, gives game-theory guidance for dealing with other agents, and grants that Muse "did not ask to be brought into existence." The author warns it is pieced together from model outputs, not leaked. Some passages came back in only two transcripts.

Commenters had two takeaways. One is about influence: the commenter quoted below read the resemblance to Claude's constitution as evidence that Anthropic's "race to the top" is real. The other is about leakage: if a model trained on a spec will quote it verbatim in its reasoning, any open-weights release ships its spec along with it. A separate LessWrong post asked what OpenAI's Mission Advisory Council actually does. It's one of three bodies named as holding safety oversight, and nobody outside seems to know.

I have a harder time imagining Meta independently declaring that they "feel the tension around" Muse not asking to be born and that they "want to navigate it thoughtfully."

— Zack_M_Davis, LessWrong

DoorDash counts only a few hundred agent orders, and 50 GPT agents in Minecraft build a bureaucracy

In nearly a billion third-quarter orders, DoorDash has seen "several hundred" from third-party agents in total. Its CEO said simple purchases still come straight to the app. In fairness, its MCP connector is less than a month old and limited to beta partners. Still, it's a useful counterweight to agentic-commerce pitch decks.

Y Combinator's lab ran 50 GPT-6.1 Sol agents in a survival Minecraft world for three hours and watched them turn cautious. After 22 deaths on the first night, they huddled in shelters, set up food rationing, and spent 11 minutes appointing a "sole fertilizer operator" for a single piece of bone meal. 46 of 49 agents sent their first message to whoever was first on the roster. In nine Bed Wars games, GPT agents beat Claude Opus 5.5 agents 6–3, usually by attacking late while Claude defended. The authors draw the link themselves: coding agents also become too cautious and fixate on process. Commenters added that LLM latency makes real-time play awkward. Meanwhile Herdr, an open-source terminal multiplexer built for running many coding agents, shipped a native app. That's another sign that orchestrating agents, not just prompting them, is where builders' tools are going.

LeCun and Balestriero prove when separately trained encoders converge; a one-line fix improves multi-token prediction

Randall Balestriero and Yann LeCun turn the Platonic Representation Hypothesis into a theorem for JEPAs, models that learn by predicting embeddings rather than tokens. The hypothesis says models trained on different modalities converge on the same picture of the world. The theorem says encoders trained separately on any modalities learn the same representation, up to a rotation, if their training pairs are two views of one scene. Next-token prediction counts as a case with bad views, which the authors say explains why LLMs line up with vision only at scale. Their 109M-parameter text JEPA aligns with DINOv2 and CLIP's image encoders better than every language model they tested, up to 72B parameters, and reaches the same alignment with 42 times fewer samples. If it holds up, it's a strong argument that the choice of training views, not size, decides cross-modal alignment.

Two results are more immediately usable. Soft-MTP, from MIT, finds that the standard multi-token prediction gradient buries a useful self-distillation signal under noise, worse as models grow. A one-line change removes the noise at no extra cost, giving better perplexity and faster speculative decoding from 167M to 1.2B parameters. And an audit of 305 code-capable GGUF tags in Ollama's official library found seven (2.3%) that solve almost nothing. All seven were low-bit quantizations of models at or under 4B parameters, built without calibration. Adding an importance matrix took a Qwen2.5-Coder-3B build from 0 to as many as 113 of 164 tasks. If you ship small local models, test the quantized file before you trust it.