TL;DR

  • The Catalyst: The Wall Street Journal reviewed Anthropic's confidential draft S-1: $16.3 billion of revenue in the first half of 2026 (up from $1.1 billion a year earlier), an $80.7 billion net loss of which about $72 billion is non-cash, and up to $750 billion in compute commitments. A listing could come as soon as November at a valuation reportedly near $2 trillion.
  • The Mechanism: OpenAI is negotiating a $30 billion private round at a $1.4 trillion valuation, a price it proposed itself, and reportedly wants to close before Anthropic lists. Whoever prices first sets the comparable for every frontier lab that follows.
  • The Outlook: The loss is mostly accounting. The revenue is real and growing faster than almost anything in software history. The risk that matters is structural: Anthropic's three largest compute suppliers (Google, Amazon and Microsoft) are also its competitors, and the filing itself warns that compute could be "curtailed, repriced, or terminated."

For three years, the valuation of frontier AI labs has been a private negotiation between a handful of founders and a few dozen of the world's largest pools of capital. Nobody outside the room saw the numbers.

That ends with an S-1. Even a draft one, reviewed by a newspaper, changes the conversation, because it has to say things a pitch deck never does.

The Numbers, as Reported

Item Figure What it means
H1 2026 revenue $16.3B Up from $1.1B in H1 2025
Q3 2026 revenue (told to investors separately) $16.8B Up 45% from Q2; one quarter now exceeds the whole first half
H1 2026 net loss $80.7B About $72B is non-cash charges tied to financing instruments
H1 2026 compute and infrastructure spend Nearly $9B More than the $7.3B spent in all of 2025
Compute commitments Up to $750B over coming years Much of it payable regardless of usage
Noncancelable cloud contracts Google $111B, AWS $110B, Microsoft $31B Three suppliers, all with competing models
Customer concentration About 20% of revenue from two customers Unusually high for a company this size
Listing timing As soon as November 2026 Public filing expected within weeks
Reported valuation Near $2 trillion Press reporting, not a company disclosure

Do the arithmetic on the revenue line and it becomes the story. If H1 was $16.3 billion and Q3 alone was $16.8 billion, up 45% from Q2, then Q2 was roughly $11.6 billion and Q1 roughly $4.7 billion. That is a company more than tripling its quarterly revenue in two quarters at a scale of tens of billions. I cannot think of a precedent in software.

The Loss Is Mostly Not What It Looks Like

An $80.7 billion loss in six months reads like a catastrophe. About $72 billion of it is non-cash: accounting charges tied to the financial instruments Anthropic issued to raise money. When the value of a company rises very fast, instruments that convert into its shares get marked up, and that markup flows through the income statement as a loss.

Strip that out and the cash picture is still heavy, but it is legible. Nearly $9 billion of compute and infrastructure spending in six months, against $16.3 billion of revenue. The Broadcom earnings call in September gave the other half of the equation: management estimated Anthropic could generate about $30 billion of annual revenue per gigawatt deployed, and Anthropic is on track to take 5 GW of TPUs in 2027 with another 10 GW in line of sight for 2028.

If that revenue-per-gigawatt holds, the business works. If it does not, the commitments are the problem.

The Real Risk: Your Landlords Are Your Rivals

The line from the filing that stayed with me is the one investor Chamath Palihapitiya also flagged on October 23. Anthropic warns that hyperscalers increasingly treat it as a competitor, which makes it harder to secure compute, and that its compute could be "curtailed, repriced, or terminated."

Look at who those hyperscalers are:

Supplier Commitment Competing product
Google $111B Gemini, including Gemini 4 Argon, at the same $2 / $10 workhorse price point as Claude Sonnet 5.5
Amazon Web Services $110B Its own models and Bedrock, which also resells OpenAI's GPT-6 Astra
Microsoft $31B Its own models, plus its long partnership with OpenAI

This is the platform risk every startup learns about in its first year, at a scale nobody has seen. Anthropic sells intelligence. Its three biggest suppliers of the raw material sell competing intelligence, set the price of that raw material, and in Google's case design the chips it runs on.

The mitigation is visible too. Broadcom's XPV financing vehicle with Apollo and Blackstone, a large Blackstone-led debt facility that closed in October, and direct TPU purchases all push Anthropic toward owning more of its compute rather than renting it. An IPO is, among other things, a way to pay for that.

OpenAI's Counter-Move: Price It Privately First

OpenAI is taking the opposite route, for now. It has postponed any IPO until at least next year and is instead negotiating a $30 billion round at a $1.4 trillion valuation, about 1.6 times its March valuation of $852 billion. Unusually, OpenAI proposed the price itself as a fixed pre-money figure. No lead investor has been chosen. MGX and BlackRock are in talks, MGX is assembling a UAE syndicate of up to $10 billion, and University of California endowments, Thrive Capital and Andreessen Horowitz are reportedly in discussions.

The timing matters more than the size. Reuters reported the round aims to close before Anthropic's expected mid-November listing. Once a frontier lab trades publicly, every private round gets compared to a live market price every day. Closing first locks in a number before that comparison exists.

Anthropic OpenAI
Route IPO, as soon as November Private round, IPO postponed to at least 2027
Reported valuation Near $2T $1.4T (proposed by OpenAI)
Last known revenue signal $16.8B in Q3 2026 alone Not disclosed in this round's reporting
Disclosure burden Full public filing Private investor materials
Key exposure Compute from rival hyperscalers Concentration in its own consumer product, 1.2B+ weekly ChatGPT users

What a Public Frontier Lab Changes

  1. Quarterly truth. Revenue, margins and compute spending become public every three months. The entire AI pricing debate (whether $2 / $10 is profitable) gets real data instead of leaks.
  2. Risk factors as policy documents. The filing reportedly warns that a catastrophic AI incident could exceed insurance coverage. That sentence, in a securities filing, will be read by regulators, insurers and plaintiffs. After the agent containment incidents of the past few months, it is not hypothetical language.
  3. Index gravity. A company valued anywhere near $2 trillion eventually lands in index funds. That means ordinary retirement accounts end up holding frontier AI risk, directly.
  4. Comparables for everyone else. Moonshot, DeepSeek and others are preparing listings in Hong Kong and Shanghai. A US frontier lab trading publicly gives every analyst a yardstick.

The Leftover

Somewhere in the draft there is a sentence about compute being "curtailed, repriced, or terminated." Three verbs, chosen by lawyers, describing what could happen if the companies that rent you your factory decide they would rather win.

The revenue line is the headline. That sentence is the part I would read twice.

Sources