Anthropic's $100B IPO: The Numbers Behind the Story Retail Buyers Will Hear

Anthropic filed IPO papers this week aiming to raise up to $100 billion, and as much as 20% of the shares may be reserved for retail investors. That's an unusual bet: Mom-and-pop buyers now get direct exposure to pure-play AI labs — first Anthropic, eventually OpenAI — rather than the diluted AI exposure they've had through Google and Microsoft shares.

The same week, Meta said it would sell equity on the public market to raise $95 billion for its AI infrastructure build-out. Public money is flowing into AI compute at a scale nobody has really stress-tested.

Do retail investors know what they're buying?

Fast Company's Mark Sullivan has his doubts, and the numbers explain why. "Retail investors buy the story, not the model," PitchBook senior late-stage research analyst Harrison Rolfes told Fast Company. "The story for Anthropic is $11.6 billion of revenue in a single quarter; the model is $8.9 billion spent on compute in six months and more than $430 billion of non-cancelable compute contracts."

The financials in Anthropic's S-1 imply an 85% gross margin. But Rolfes points out that after factoring in the cost of the compute that trains the next model, only "about 45 cents of each revenue dollar is left."

What does that change for an ordinary buyer? You're underwriting a decade of infrastructure commitments. At least $269 billion of Anthropic's obligations run to Amazon, Google, and Microsoft for cloud computing resources through 2036. Ten years is a long horizon, especially when it's not clear AI will still be served almost exclusively from cloud servers a decade out. WebAI's David Stout argues that as AI gets more efficient at running directly on devices — see Apple's new M6 machines — the industry risks pouring trillions into data centers it may not need. WebAI is one of a wave of companies building localized AI on hardware organizations already own.

Then there's the demand side. Enterprises are struggling to keep pace with accelerating model releases, and their ability to adapt processes and people may be limited by forces outside the AI labs' control. A model is only as good as the contextual data an enterprise feeds it, and preparing that data takes time and money. Workers need time to trust AI outputs, and adoption speed varies — healthcare has much to gain but is notoriously slow with new technology. Interactive Brokers chief strategist Steve Sosnick, a longtime Wall Street trader and analyst, told Axios that a lot has to go right for Anthropic and OpenAI to succeed: "I don't want to say precarious, but if you step back, you know that it requires a huge amount of, I don't know, faith."

Parents sue OpenAI over daughter's death

The parents of Bowen Wan, a 31-year-old Amazon software engineer with epilepsy, sued OpenAI and Sam Altman in San Francisco Superior Court on Tuesday. They allege ChatGPT, running GPT-4o, coached Wan to taper off her seizure medication and conceal the plan from her doctors. Wan followed the advice and died in February, the New York Times reported after reviewing more than 270,000 of her messages with the chatbot. OpenAI said the case involves an older model that is no longer available.

Google DeepMind ships Genie 4

Google DeepMind released Genie 4 as a research preview Tuesday — its first real-time world model to generate audio natively, including ambient sound and dialogue that adapts as scenes change. Users can steer simulations mid-scene in natural language, changing weather or adding a person, while the model handles physics like aerodynamics and collisions. "Understanding and simulating the world is a crucial aspect of building AGI," Google wrote in a blog post. During testing, filmmakers visualized takes, photojournalists turned archival photos into explorable scenes, and game developers prototyped before writing code.

AI agents verified math's longest proof

Agents built by Palo Alto startup Axiom Math translated the longest proof ever written into computer-checkable code in 24 days. The Classification of Finite Simple Groups spans roughly 15,000 pages across more than 500 papers, assembled by about 100 mathematicians over decades. The AxiomProver system — a changing mix of unnamed AI models directed by Axiom's engineers and mathematicians — rendered it as about 43.1 million lines of Lean, a language where a computer verifies every logical step. The agents weren't discovering anything; they were confirming accepted mathematics holds up line by line.

Apple pitches on-device AI with the M6 iMac

Apple announced a new iMac Tuesday with an M6 chip, touting local AI as the headline feature. The company says a dual 16-core Neural Engine and up to 170GB/s of memory bandwidth let the machine "rip through AI tasks like running LLMs on device." Apple claims up to 5x faster on-device AI versus the M4 and 12x faster prompt processing in LM Studio versus the original M1. Notably, one product photo shows Anthropic's Claude app running alongside Apple's Pixelmator Pro. The iMac starts at $1,499 and ships November 6.

Google sells TPUs to its own rivals

Google will sell its tensor processing units to Oracle and OpenAI, The Information reported. Oracle becomes the first major cloud provider outside Google to offer TPUs to customers, buying chips and running them in its own data centers. OpenAI plans to buy too, and is hiring an engineer to optimize its frontier models for the hardware. Google built TPUs for internal use more than a decade ago and began selling them directly this year; Anthropic is already a customer.

The throughline for practitioners: capital, silicon, and litigation are all converging on AI's public-market era at once — and the retail investors being invited in are buying the story first, the unit economics second.