Three threads run through today’s issue.
Anthropic’s pace and its bills are under the microscope. Ornn reads the amended IPO filing and says $895 billion of computing commitments needs revenue near $200 billion a year, while Zvi Mowshowitz argues the company is months ahead of the AI 2027 forecast.
Computing power is the scarce thing everywhere else too. Google announced two more TPU chips, Meta says Muse Ember costs about a third as much per task as Claude Haiku 5.5, and Tomasz Tunguz’s model has the US importing AI output between 2028 and 2030.
The third thread is control. OpenAI says its Control Room can cut off a research job within 35 seconds, and Fireworks now says its October breach exposed billing records.
Anthropic at Full Speed: A Faster Model and a Very Large Bill
Anthropic shipped a quicker model, and its IPO filing shows what it owes for the computing behind it. Two outside writers read the same company in different directions.
- Anthropic’s Claude Fable 5.2 runs up to 30% faster at a lower price. Fable 5.2 lists at $8 per million tokens in and $40 out, and Anthropic says output arrives up to 30% faster than 5.1. Its twin, Mythos 5.2, drops the safety classifiers and is limited to approved security and biology teams.
- Anthropic’s $895 billion computing bill needs revenue near $200 billion. Ornn, which runs a compute price exchange, estimates Anthropic must roughly triple from its $67.8 billion annual pace to carry the payments due through 2030. The $200 billion is Ornn’s own arithmetic, and Ornn sells price data on the market Anthropic buys from.
- Zvi Mowshowitz says Anthropic is months ahead of the AI 2027 forecast. Mowshowitz reads METR’s Anthropic findings as AI already speeding up AI research, and notes that METR says Anthropic redacted most of its best evidence. He backs giving lawmakers the full report, and Anthropic’s cofounder says its own analysis is due within weeks.
Cheaper, Smaller, Closer to the Chip: Hardware Shapes the Models
Four companies are building models and silicon around a specific job and a specific price. The numbers in all four come from the companies themselves.
- Google’s two new AI chips: one for huge inputs, one for secret data. Google says TPU 8i Max puts two chips on one package for roughly twice the silicon in the same rack, while TPU 8tx keeps data and model weights shielded even while in use. The blog gives no prices or ship dates.
- Meta releases a cheap AI model built for the busywork of agents. Meta says Muse Ember scores 35.8 against Claude Haiku 5.5’s 36.2 on one index at $0.136 per task instead of $0.411. It is sold through Meta’s API only, and several of its results are Meta’s internal tests.
- NVIDIA shrinks a 120B model to fit on one GPU, by designing for the chip. NVIDIA researchers cut Nemotron 3.5 Super from 120 billion to 74 billion parameters so it serves on a single B200, at a cost of 2.85 points of average score. It is one data point from NVIDIA’s own models on NVIDIA hardware.
- NVIDIA’s new robot brain learns by imagining what happens next. NVIDIA Research says GR00T 2 reached 95% success on new tasks after five demonstrations, learning partly from videos of people. It all comes from NVIDIA’s own robots and tests, and the rate with no robot-specific training was 40%.
Where the Compute Goes: Power Limits and the Cost of Thinking
Three writers ask who gets the computing power and what extra thinking time actually buys. They disagree about where the gains come from.
- A venture investor says the US will soon use more AI than it can power. Tomasz Tunguz’s basic model has US demand passing home-grown data center capacity between 2028 and 2030, depending on how many planned projects get cancelled. It is one investor’s estimate, and his own model holds computing prices flat.
- AI that keeps improving on problems bigger than it trained on. A Forethought essay argues that some models keep improving with far more thinking time than they trained with, as AlphaGo Zero did. The essay itself says its evidence for commercial models is circumstantial.
- Blogger says AI’s extra thinking time was paid for in training. Alex Wang argues the hard searching happens during training, so one answer at question time usually works without a longer chain. He calls it a theory written mostly for himself, and he tests no commercial model.
Off Switches and Break-ins: Who Can Stop the Machines
One lab describes how it halts its own research jobs, and one hosting company explains how much more got out than it first said.
- OpenAI built a master off switch for its AI research jobs. OpenAI says Control Room can cut a research model’s output within 35 seconds of a stop order, and that every request is denied if the switch itself fails. The post is OpenAI describing its own system, and some jobs were not yet covered.
- Fireworks AI links its breach to a gang that resells stolen AI access. Fireworks says billing records and, for some accounts, partial datasets and unfinished models were exposed, and blames a financially motivated group. Customer dataset uploads stay switched off with no return date.
Today’s Quick Hits
- Anthropic reportedly buys $3.4 billion of gas generators. The Financial Times reports Anthropic took over Caterpillar generator orders worth 1.7 gigawatts, originally placed by OnePWR, which could not find the financing. That works out to about $2 billion per gigawatt of power.
- Apple starts charging $49.99 for extra AI in Logic Pro. The Verge reports Logic Pro now sells a one-time $49.99 Intelligence Usage Pack, tied to one Apple Account, with credit that lapses after 12 months. Apple’s other creative apps are expected to follow this month.
- A Claude user says Opus 5.5 started thinking less on 3 November. Shane Melaugh claims Opus 5.5’s high setting now reasons far less, citing one bug and an analysis of 1,400 of his sessions by another AI. Anthropic has not responded in the thread, so nothing confirms a change.
- ChatGPT Enterprise can now be bought on an AWS bill. OpenAI listed ChatGPT Enterprise in AWS Marketplace, so eligible spend can count toward a company’s Amazon commitments. OpenAI does not say how much of the spend will qualify and points buyers to their account teams.