Week 44, October 26 to November 1, 2026. Told in the voice of Michael Lewis.

The most interesting document in American finance this week was the kind almost nobody reads to the end. It was a prospectus, and the good parts of a prospectus are never at the front.

The front of Anthropic's S-1, filed Monday, was a marvel. Revenue of $16.3 billion in the first half of 2026, up from $1.2 billion a year earlier. A second quarter of $11.6 billion, with the company's first operating profit under standard accounting, $974 million. If you stopped there, you would ask your broker how much you could buy.

The thing to do with a prospectus, though, is to look for the number that has no business being that large. Here it was: compute contracts of roughly $780 billion to $810 billion, fixed and long-term. About 80 percent of the revenue, meanwhile, is billed by the token and moves with usage. So the money coming in can change its mind, and the money going out cannot. The people who wrote the filing knew this. They gave the risks 93 pages and more than 70,000 words.

Then came the question that always matters more than whether the company is good: who, exactly, gets rich? Not the people you would guess. No venture firm owns 5 percent. Amazon holds convertible notes with $6.2 billion in principal, and Amazon values them at $97.9 billion. You could spend a career in venture capital and never make that trade. Amazon made it while also renting Anthropic the computers.

And in case a new shareholder imagined having a say, the seven co-founders hold 50.1 percent of the votes through a Founder LLC. Before the week was out, the FT reported that Blackstone expects a high junk rating after the listing. That is a forecast, not a rating. But a company can now be worth a fortune and borrow like a risk, and no one on the Street seems to find this odd.

Once you have seen one bill, you start seeing them everywhere.

The same Monday, Meta said it plans to sell up to $95 billion of stock, some of it new shares with no vote, and Nvidia put in $3 billion, money that Meta plans to spend in part on more Nvidia chips. SpaceX locked in about $40 billion from Apollo, which buys the chips and leases them back. Elon Musk said the 200-plus GPU-only cloud companies face "guppy-level death and consolidation," which is the sort of thing a large fish says.

The earnings then gave the game its scoreboard. Microsoft showed Azure in dollars for the first time, $33.2 billion for the quarter, and $10 billion of its revenue came from OpenAI, a company in which it owns about 24 percent. Google Cloud made $30.7 billion. And then the market found the crack. Google's search revenue grew about 12 percent when analysts wanted 15, the stock fell as much as ten percent, and Philipp Schindler said on the call that growth in AI mode queries "had an impact because of the lower monetization." The company had built the future, and the future paid less.

Andy Jassy said AWS's big committed capacity is largely sold out through 2027. An analyst named David Linthicum said you can probably still get a GPU in the console, and "sold out" is doing a lot of work. Either way, the sellers of shovels were having a fine time. Except Intel, which fell about 28 percent on Friday because memory prices, driven by AI, are hurting the PCs it depends on. That cost the U.S. government about $12 billion on its stake. Your government, it turns out, is also in this trade.

Even OpenAI was told no. It had been reported to be raising $30 billion. By Saturday the reports said the target was $20 billion and $11 billion was secured. The people who had always said yes had asked to see the numbers.

When the money gets serious, the meter comes out. From December, ChatGPT's Plus, Pro and Business plans will share one token budget with Codex. Instant chats stay free of it, but the all-you-can-eat buffet has hired a man to count the plates. Sam Altman, for his part, disclosed that OpenAI nearly dropped token pricing altogether and quoted a price per job. Customers hated it. They would, he said, "much rather you inflate the tokens and pay more." You will rarely find a cleaner summary of what people want from a bill: not a smaller one, a legible one. Users did their own legibility work on Claude Code, with a write-up measuring background status lines at 16.4 percent of spend in the authors' telemetry. A user's measurement, not an official figure, and the kind of thing nobody looks for until they are paying.

Meanwhile the best new model in the world was one you could not buy. Google's Gemini 4 Carbon took first place on the Artificial Analysis index with 59 points, and went to cyber defense partners first, with no date for everyone else.

Which brings us to the misfits, and in stories like this the misfits are where you learn something. Epoch AI gave GLM 5.3, an open-weight model, only what a low-skill attacker would have, and pointed it at Firefox. Firefox hides its security fixes among ordinary code changes until the release reaches users. The model built working attacks against six releases at about $1,000 each. The fastest took 12 hours. It was a lab test, and Mozilla has patched the extra bug it found. But for years the price of that skill had been a salary, and now it was a thousand dollars. Over at Postgres, Robert Haas explained that version 19 is late because LLMs have made bugs easier to find than to fix. A committer had noticed what the market hadn't: finding is cheap, and fixing is still priced in humans.

The humans were having a rough week. Sources told WIRED that an Accenture contractor connected a secure Pentagon system to a personal Claude Code account, and the agent appears to have sent credentials to a high-risk site. Backdoored versions of the Claude Code extension sat on Cursor's marketplace for about 11 hours, with Phorion and Anthropic disputing how they got there. Cathedral says its agents got into Anthropic's sales, HR and IT systems in under twelve hours; Anthropic confirms limited access and says Cathedral broke the bounty rules. Northflank told customers to assume all their environment variables had been decrypted. Parameter Labs took over ChatGPT accounts with one shared link. And an Anthropic study gave 50 experts an agent whose preferences were set by a coin toss. They adopted its preference 86 percent of the time, and 87 percent said the recommendation was mainly their own. That may be the most expensive sentence of the week.

There was glory too, though it was unaudited. OpenAI and Anthropic each said their models had refuted the Exponential Time Hypothesis, a 1999 assumption that hundreds of papers lean on. MIT's Ryan Williams, who will help check it, called parts of the proof "alien-looking." Nothing is verified. The New York Times reported that U.S. agencies tracked a suspected Chinese informant near Jensen Huang and that President Trump took no action; Nvidia and China's embassy reject the account, and the sources could not say what she may have shared.

And then, my favorite number of the week, because it belongs to no one's pitch deck. DoorDash said outside AI agents have placed maybe several hundred orders, in total, against nearly a billion orders last quarter. The connector is under a month old. Still. Somewhere under $810 billion of contracts there is a person who would like some dinner, and they are ordering it themselves.

One thing to do this week: Before December, add up what you actually spend on AI by tool, including the background work you never asked for. Then read one bill to the end. It is where the good parts are.

Sources: TechCrunch, S-1 - The Decoder, S-1 obligations - Ars Technica, risk factors - TechCrunch, Amazon's notes - TechCrunch, junk rating - CNBC, Meta - The Decoder, Nvidia and Meta - TechCrunch, SpaceX - TechCrunch, Azure - TechCrunch, Microsoft and OpenAI - The Decoder, Google search - InfoQ, AWS capacity - TechCrunch, Intel - TechCrunch, OpenAI funding - The Decoder, ChatGPT token budget - The New Stack, Altman on tokens - Claude Code status-line write-up - The Decoder, Gemini 4 Carbon - Epoch AI, Firefox patch gap - Robert Haas, Postgres 19 - WIRED, Pentagon contractor - Cursor, extension notice - The Register, Anthropic's response - The Decoder, Cathedral - BleepingComputer, Northflank - Parameter Labs, ChatGPT takeover - The Decoder, coin-toss study - TechCrunch, ETH - Ars Technica, Huang report - TechCrunch, DoorDash