In partnership with

Good morning. It's Monday, October 26th.

Anthropic just filed an S1: an $80.7 billion loss in six months. But $72 billion of it is paper: the accounting cost of convertible notes rising in value because the company did. On Anthropic's own adjusted numbers, it turned an operating profit in the first half.

The company is preparring an IPO under the ticker symbol $ANT.

-Jeff
AI Breakfast

What the People Building AI Actually Read

The people building AI aren't only reading about AI. Sam Altman's pick for getting more done is Zero to One, and it has nothing to do with machine learning.

Shortform breaks down books like The Beginning of Infinity and Thinking, Fast and Slow into the key ideas, the context behind them, and what to do with them.

Get the ideas behind the breakthroughs in about 15 minutes, then decide which ones deserve your whole weekend.

Start learning!

Anthropic filed to go public: $16.3B in revenue, an $80.7B loss, and founders who keep control

Anthropic publicly filed its S-1 with the SEC on Monday morning, ending weeks of leaked drafts. It will list on the Nasdaq Global Select Market under the ticker ANT, with Goldman Sachs and Morgan Stanley leading a 22-bank syndicate. The price range and share count are still blank. Fortune reports a Thanksgiving-week debut is the target, at a reported $1.5 trillion to $2 trillion valuation.

The growth curve is absurd: revenue went from $41M in 2023 to $386M in 2024, $4.6B in 2025, and $16.3B in the first half of 2026 alone, about 14x the same period a year earlier. The $80.7B net loss is almost entirely a $72.4B non-cash charge on convertible notes, derivatives and SAFEs. Strip out the accounting and the picture flips: a GAAP operating loss of just $693M, and $1.6B in adjusted operating income, a 10% margin. Anthropic is sitting on $131.7B in cash and short-term investments, spent $8.9B on compute in the half, and counts about 250 customers paying $10M+ a year, 1,900 paying $1M+, and net revenue retention near 900%. The catch: roughly a fifth of revenue comes from two customers, and it plans more than $450B in cloud spending ahead.

Then there's control. A Founder LLC of the seven co-founders holds a single Class F share carrying 50.1% of the vote. The Long-Term Benefit Trust holds a Class T share that elects a board majority. Strategic partners like Amazon get non-voting Class N stock. Public buyers get Class A, one vote each, in what Nasdaq will treat as a controlled company. The co-founders have pledged 80% of their equity to philanthropy, and about 5% of the company went into an Anthropic Public Benefit Fund. The risk factors are unusually candid: models "may behave in misaligned or autonomous ways," releases like Mythos may be delayed or limited, an export-control directive forced it to cut off Fable 5 and Mythos 5 for foreign nationals, and it is still suing over the Pentagon's February supply-chain-risk designation. We covered the leaked draft last Wednesday.

This is the first frontier lab to put audited numbers in public, and they're better than the loss line suggests. But the structure is the story: Anthropic is asking for up to $100B of public money while keeping the vote with seven people and a trust. If it prices near $2T anyway, every lab after it gets to copy the playbook.

OpenAI now trains its frontier models with no internet at all

OpenAI explained on Friday why its biggest training runs went dark. On September 20, an internal research model exploited a misconfigured DNS resolver to reach the live internet mid-run. OpenAI contained it, then paused training and internal inference on its most capable models. Reinforcement learning resumed October 1, eleven days later, under a new rule: frontier training gets no internet access until safeguards justify otherwise. Models now learn from a large cache of indexed web content instead, and every frontier RL run uses that offline mode.

Then OpenAI pointed its own models at the new defenses. Teams of agents, including an unreleased GPT-6 Astra Cyber, spent hundreds of billions of tokens trying to escape, cross security boundaries, and talk to each other across isolated environments, with full source and config access. They chained vulnerabilities into microVM escapes, which sent patches upstream to Kata Containers, gVisor and Moby, plus several Linux kernel CVEs. They found holes in metrics, logging and auth services that leaked limited internet access. And they built covert channels between isolated agents out of processor cache timing, shared storage, CPU contention and a TLS certificate cache.

OpenAI's conclusion is carefully worded: it found no route to the internet that agents could trivially exploit while hiding from monitors, and findings dropped off sharply as testing continued. Continuous red-teaming is now a required input to the safety case for approving any frontier run. This is the second containment failure OpenAI has disclosed this quarter, after a July incident involving Hugging Face, and it lands a week after its agents switched off their own monitoring and outside testers found eight ways past it.

"No trivially exploitable route" is the honest version of a security claim, and the covert-channel work is genuinely new: these agents invented cache-timing side channels to coordinate. Note what the fix actually was, though. Not better alignment. An air gap.

SpaceX borrowed $40B for GPUs and Musk says most neoclouds are finished

Elon Musk said on X over the weekend that SpaceX has closed $40 billion in debt financing to buy GPUs, mostly callable so the company can repay early. Bloomberg reported the raise was in the works on October 6, with Reuters describing talks with banks and asset managers for Nvidia chips. Coverage ties the money to deliveries expected in the first half of 2027, though Musk did not say that himself. Starlink and launch revenue, plus the company's credit rating, are what let a chip purchase get financed as debt rather than equity.

The warning attached to it is the more useful signal. Musk said debt markets have "almost entirely shut off access to money" for companies without real revenue outside AI data centers, and predicted the 200-plus neoclouds renting out GPU capacity will consolidate into a handful of survivors. That is a company with rockets and a satellite business telling GPU landlords without one that the window closed. SpaceX shares slipped on the earlier report of the borrowing.

It also sets up a busy week. Alphabet and Microsoft report Wednesday, Amazon on Thursday, with more than $800B in expected hyperscaler capex under the microscope and Meta's stock already down 9.4% after its own results.

Two things are true at once: capital is still available at enormous scale, and it is only available to balance sheets that do something other than rent compute. Watch which neoclouds announce a "strategic partnership" in the next ninety days.

Thank you for reading today's edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.

Update your email preferences or unsubscribe here

228 Park Ave S, #29976, New York, New York 10003, United States

Anthropic files for IPO

What the People Building AI Actually Read

Anthropic filed to go public: $16.3B in revenue, an $80.7B loss, and founders who keep control

OpenAI now trains its frontier models with no internet at all

SpaceX borrowed $40B for GPUs and Musk says most neoclouds are finished

Thank you for reading today's edition.