Anthropic's compute bill swings it to a loss, and schools lose Search with AI

Anthropic's amended S-1, filed Monday, adds results for the quarter ending September 30. Revenue reached $17 billion for the quarter, but the company swung from an operating profit of $1 billion in the June quarter to an operating loss of $567 million. The filing blames $5 billion of compute bought under short-term deals, which it says costs significantly more than capacity under its multi-year contracts, plus about $2 billion of non-cash charges tied to a donation-matching program.

Compute and infrastructure rose to 58 percent of revenue, up from 48 percent the quarter before. Anthropic now reports commitments worth about $895 billion, mostly for compute, and is shifting toward leasing chips directly through a program it calls Anthropic Direct. TechCrunch notes this is the opposite of the margin trend an IPO candidate would want to show, though with revenue still climbing fast, the cause looks like a scramble for supply rather than weak demand.

Anthropic separately published research on "provable inference," a cryptographic proof that a response came from the model a provider claims, without revealing that model's weights. Its system, FlashProve, targets cases where you cannot simply trust the label: a routing service quietly swapping in a cheaper model, an attacker tampering with the token stream, or an outside evaluator needing to know it tested the model actually being served.

On Llama 4 Maverick, a single H100 GPU produces proofs at about 110 tokens per second, which Anthropic puts at 70 to 95 times the cost of running the model itself. Anthropic says Claude built many of the speedups, so the team backed the key security claims with formal Lean proofs. The code is open source under Apache 2.0, and Anthropic calls this intermediate progress with no plan to deploy it yet.

Google says Gemini, Gemini Notebook, Gemini in Classroom and Google Search will be entirely off for K-12 student accounts until a school administrator turns them on, with details coming in the months before the change rolls out. The catch is Search: Google's education chief, Chris Phillips, told The New York Times that search and AI are now so deeply integrated that a school blocking AI will block Search too, leaving students with a message that it is unavailable.

Alongside the change, Google made its Guided Learning tutor, which coaches students through problems instead of handing over answers, available worldwide in Google Classroom, and will pilot it inside Search at select US schools later this month.

At a press event on Monday, Adam Fry, who leads consumer product for ChatGPT, said the volume of searches run through ChatGPT now equals 20 percent of Google's. OpenAI counts a search as any query that triggers a web lookup, not simple definitions or how-to answers, and Fry did not say where the Google figure came from, so treat the comparison as OpenAI's own.

Fry also said ads have not hurt usage and that OpenAI may lean further into them; last month it began showing images of sponsored products in ChatGPT answers. The Verge notes that Microsoft's Satya Nadella testified in 2023 that Bing held about 3 percent of search after more than a decade of trying.

Apple is adding a "Dating and Relationships" declaration to its App Store age ratings, covering dating features and any virtual companion that engages users in sexual or romantic content. Any app that declares it gets a minimum rating of 18+. Developers can start declaring on November 16, and every app must complete the questionnaire by January 12, 2027.

The rule matters most for AI companion apps such as Character.AI and Replika, which often draw younger users and mix companionship with other features. A second new declaration covers apps that share a user's real-world location with other people.

Fal raised $1.3 billion in a round led by Bessemer Venture Partners, with Nvidia and T. Rowe Price joining, Bloomberg reported. That values the five-year-old company at $16 billion, double its valuation in April. Fal confirmed the round as its Series F and disclosed a Series E led by Sequoia alongside it.

Fal hosts image, video and audio models, including its own, for studios, ad agencies and game developers, and Bloomberg reported it passed $800 million in annualized revenue this year. TechCrunch notes that inference providers such as Baseten and Fireworks AI are raising at similar pace, though their margins stay thin because compute is expensive.

Some AI data center builders are generating their own power with engines pulled from retired aircraft and rebuilt to burn natural gas, The New York Times reported. The draw is speed: a Rystad Energy analyst said these units can have a data center running in about 24 months, while the International Energy Agency puts a 100-megawatt grid hookup in Northern Virginia at about seven years.

The Times reported, citing state permits and satellite imagery, that xAI has installed eight converted units in Mississippi to help power its Colossus 2 data center. Crusoe, which is building OpenAI's Stargate campus in Abilene, Texas, ordered 13 in April. The engines are less efficient and pollute more than turbines built for gas plants, which is why some experts expect them to be a stopgap.