"$1.25 billion a month." That is the number reportedly buried in SpaceX's own filing, per the QbitAI report, describing what Anthropic has committed to pay for compute under a single long-term contract. Fully executed, it runs to roughly $3.75 billion a quarter — about a third of the $11.5 billion revenue figure Anthropic just showed prospective investors.
Two facts landed in the same week, and they refuse to sit comfortably beside each other.
The first: Anthropic told investors it crossed $11.5 billion in a single quarter, up from $787 million a year earlier and $4.73 billion in Q1, with what QbitAI reported was its first-ever positive adjusted operating income.
The second: its own CEO, Dario Amodei, publicly conceded that AI faces "a crisis of trust" — and that "by far the most accurate criticism of AI companies including Anthropic is that we haven't yet delivered on our big promises to benefit the world. That is totally on us."
By far the most accurate criticism of AI companies including Anthropic is that we haven't yet delivered on our big promises to benefit the world.
The question worth asking is which of those two facts a $2 trillion IPO valuation is actually pricing. Because it cannot be pricing both.
Start with the curve, because it is genuinely intoxicating. Annualized revenue was, per QbitAI's estimates, around $100 million in 2023, roughly $1 billion at the end of 2024, about $9 billion at the end of 2025, over $30 billion by April 2026, and past $47 billion by May. That is close to vertical.
At $2 trillion, Anthropic would, per the QbitAI report citing the Financial Times, eclipse SpaceX to become the highest-valued IPO in history. Investors are reportedly betting on it this fall.
But look closely at the word doing the heaviest lifting in the headline: adjusted. Positive adjusted operating income is not net profit, and it is not positive cash flow. It is a number after Anthropic has stripped out some set of costs it has not disclosed.
Until an S-1 says which costs were adjusted out, "first profit" is a claim about a definition, not about the business. Reddit's first instinct — why report revenue and not profit? — is the correct one.
Then there is that SpaceX compute contract. One supplier, one contract, eats roughly a third of quarterly revenue. And that figure does not yet include what Anthropic pays AWS and Google for the rest of its inference and training.
The direction of travel on margins is not reassuring either. The Information previously reported, per QbitAI, that Anthropic cut its 2025 gross margin expectation from 50% to 40%, because serving models to paying customers cost about 23% more than projected.
So the real capital-markets question is not how fast revenue grows. It is how much of every incoming dollar routes straight back out to the people who own the chips.
Now the recognition problem, which is subtler and arguably more important. When a customer buys $10 of Claude through AWS, AWS keeps a cut — say $3 — and Anthropic nets $7.
Under gross accounting, Anthropic can book $10 in revenue and record $3 as channel cost. Under net accounting, it books only the $7 it actually keeps.
Same money, same $7 in Anthropic's pocket — but the topline differs by the channel cut. If Anthropic and OpenAI recognize revenue differently, then the crowd-pleasing line that Anthropic has overtaken OpenAI ($47 billion run-rate in May versus OpenAI's reported $40 billion-plus, per QbitAI) may be comparing two things that are not the same thing.
Nobody outside the company can resolve this without an S-1. That is the whole point.
Layer a demand-quality risk on top of the recognition risk. Q2 coincided with the peak of "Tokenmaxxing" — the stretch when AI FOMO pushed companies to encourage employees to burn tokens, some even treating consumption as a proxy for adoption.
The reversal came fast. Some firms reportedly hit three times their annual token budget by April or May, per QbitAI; Microsoft was reported to be tightening internal Claude Code usage, partly on cost.
If a meaningful slice of Q2 was demand pulled forward rather than durable enterprise consumption, then annualizing a single hot month — which is exactly what "run-rate" does, multiplying the latest month by twelve — flatters the future. A vertical line drawn from a spike is not a forecast.
Now put a price on all of it. At $47 billion run-rate, $2 trillion is roughly 43 times sales.
Even if Anthropic hits the aggressive $100–120 billion run-rate some investors reportedly project, per QbitAI, by the end of 2026, that is still 17 to 20 times sales. Buyers at this level are not paying for current revenue. They are pre-paying for years of compounding growth that has not happened yet.
Which brings us back to Amodei's concession, and to why it is the tell.
Gary Marcus has been hammering the same point: if the revenue and the mission are real, why aren't these companies more transparent about their finances? The counterargument raised in the same thread is disarmingly literal — how can a company be more transparent than going public and filing the disclosures a listing requires?
That is precisely the mechanism at stake. Private firms disclose when they need to, and not before. Going public is the event that forces gross-versus-net clarity, a real profit definition, and an honest look at compute as a share of cost.
So the structure is doing something quietly elegant for Anthropic. The IPO is being marketed on preliminary figures — Bloomberg notes they "could still change" — while early investor meetings led by CFO Krishna Rao have, per CNBC, reportedly stayed high-level, without specific financials or a valuation.
Amodei can concede the criticism is fair — we haven't delivered — precisely because the deal's timing lets him defer the proof. The admission costs nothing when the numbers that would settle the argument are still sealed.
Read charitably, this is disarming candor. Read as a capital-markets maneuver, it is the sound of a seller pricing the skeptics' transparency demand into the deal before that demand can be independently tested.
None of this means the demand is fake. $11.5 billion in a quarter is not conjured from nothing, and enterprise coding traction is real.
But the gap between "very good at collecting money" and "has escaped the gravity of the burn" is exactly the gap an S-1 exists to close.
So here is the actual decision facing investors, stripped of the vertical-curve romance. The accounting will become legible — that is what a listing does. The only question is whether you buy before the numbers are readable or after.
Buying before means trusting the adjusted figure, the run-rate extrapolation, and the recognition method sight unseen. Buying after means letting the disclosures Marcus is demanding arrive first, then deciding.
The $2 trillion number is a bet that enough people will choose "before." And Amodei's crisis-of-trust confession, however sincere, is what that bet sounds like when it's spoken out loud.