Anthropic pays SpaceX $1.25 billion a month for compute at the Colossus 1 data center in Memphis, a contract that runs through May 2029 and can be cancelled by either side on 90 days' notice.
That single line, buried in SpaceX's own prospectus, tells you more about the true price of a frontier lab than the number everyone is quoting.
The number everyone is quoting is $965 billion, the post-money mark on Anthropic's $65 billion Series H, closed May 28. Four days later the company confidentially filed a draft S-1 with the SEC, becoming the first frontier lab to walk toward public markets.
Here is the problem. That trillion-dollar-adjacent figure appears nowhere in any primary Anthropic disclosure.
The S-1 is confidential. It sets no share count and no price. The run-rate everyone cites is unaudited. And the closest thing to a live trillion-dollar mark is a set of tokenized Anthropic shares trading on Solana at implied valuations above $1 trillion.
Anthropic identified eight platforms selling those unapproved shares in May and warned investors not to buy them. The tokens promptly fell 34 to 40 percent.
So the market is pricing a frontier lab off numbers the company either didn't disclose or actively disavowed. That is the whole story in one sentence.
So the market is pricing a frontier lab off numbers the company either didn't disclose or actively disavowed. That is the whole story in one sentence.
Consider the revenue figure. Anthropic's run-rate crossed $47 billion in May, up from roughly $9 billion at the end of 2025. A more than five-times jump in five months, against a base that already exceeds what most public software companies report in a year.
But run-rate is a monthly number multiplied by twelve. As one Chinese aggregator report reportedly citing anonymous netizen claims put it bluntly: "the real revenue is much lower."
That distinction is not pedantic. It is the entire gap between a durable business and a moment.
A public prospectus discloses gross margin, customer concentration, model-training capex, contract terms with hyperscalers, and audited GAAP income. Futurum's Nick Patience said the distance between the disclosed run-rate and the audited GAAP income statement "is likely to be the single most-discussed line in the S-1."
Everything bullish about Anthropic lives in the run-rate the company chose to disclose. Everything uncertain lives in the audited gap it hasn't.
Now put the SpaceX contract back on the table, because that is where the real risk lives.
$1.25 billion a month to SpaceX for Colossus 1 capacity. Terminable on 90 days' notice. Meanwhile Anthropic has pledged more than $100 billion to AWS over ten years, with Amazon committing up to $25 billion in equity and up to 5 gigawatts of capacity.
These are not arm's-length suppliers. Amazon is a shareholder. Google reportedly holds roughly 14 percent, though without voting power, per a Chinese aggregator citing anonymous netizen claims. The same firms underwriting Anthropic's capital and compute are the ones whose data centers its revenue depends on.
That circularity — backer as shareholder and supplier at once — is the same structure our open thread on the Oracle-OpenAI overrun keeps circling. The question there is whether lenders keep absorbing the risk or start repricing it. The question here is whether public equity investors will.
Because a public shareholder is going to ask a blunt durability question. Can a lab this dependent on Amazon, Google, and SpaceX for both money and compute be an independent franchise, or is it a feature that hyperscalers eventually internalize?
And here is the turn. That is the exact question Satya Nadella is posing to enterprises.
On CNN's "Fareed Zakaria GPS," Nadella warned that companies relying wholly on proprietary AI labs won't survive. "Any firm that doesn't have this control, I will claim will not remain a firm because you've essentially outsourced your thinking."
His prescription: keep the harness separate from the model, keep context and memory separate too. Then "any one model can go away, and you can still continue to be in control of your own destiny."
He named names. Claude Code and OpenAI's Codex — the built-in coding harnesses — are exactly what he told enterprises to stop leaning on.
Note who is saying this. Microsoft is an investor in both Anthropic and OpenAI, and its cloud business sells the very hedge infrastructure Nadella is recommending. The warning is self-serving. It is also correct.
Look closely at the two warnings and they are the same wager. Nadella is telling enterprises to hedge lab dependence because any one model can go away. Public investors will ask whether Anthropic itself is durable given how dependent it is on its own backers.
The IPO and the multi-model hedge are two faces of one bet on lab durability. One asks whether the customer can survive relying on the lab. The other asks whether the lab can survive relying on its suppliers. Same axis, opposite ends.
The day after Anthropic filed, Microsoft released seven of its own models built for "long-term self-sufficiency." The distribution partner is building the alternative in public while the S-1 sits sealed.
None of this settles in a press release. It settles at a clearing price.
If the public book prints at or above $965 billion, the Series H priced the company correctly and every late-stage AI mark — OpenAI's $852 billion included — is anchored higher.
Print below, and every venture mark from the last twelve months gets reset in spreadsheets across Sand Hill Road within the afternoon.
And there is a stranger third case. Print above $1 trillion, and the speculative crypto tokens Anthropic told you not to buy turn out to have been the closer signal than the round its own lead investors — Altimeter, Dragoneer, Greenoaks, Sequoia — actually led.
For four years the frontier labs were priced privately, behind term sheets seen by a small circle. The numbers that leaked were the ones PR teams chose to leak. A prospectus ends that arrangement.
Wedbush's Dan Ives called the filing "an opening of the floodgates for the IPO market." That is the bull case, delivered before anyone has seen an audited number.
The bear case is the same event four months later, when the disclosures replace the signaling. What does it tell you that the most accurate forward valuation of a frontier lab is currently being set by tokenized shares the company itself warned you against?
The figure that will actually price this sector is not $965 billion. It is the gap between that mark and the GAAP income statement still sealed inside the confidential S-1 — and the risk that gap exposes is that the hyperscalers underwriting Anthropic are the same ones Nadella is telling everyone to hedge against.
Watch October. The first hard verdict on frontier-lab durability isn't a thesis. It's a print.