Anthropic disclosed around $517 billion in compute commitments, the same week it tightened the usage limits of its agentic coding tool. The two facts do not contradict each other: they describe the same constraint seen from two sides.
We have been documenting for months the gap between this sector's revenue and its commitments. This is the highest figure we have had to deal with.
Putting the order of magnitude in context
$517 billion is a hard amount to picture. A few comparisons drawn from what we have covered.
It is more than China's five-year plan to quadruple national compute capacity, which we reported at around $532 billion over four years — and that is a state programme, not a corporate commitment.
It is roughly forty-five times the quarterly revenue announced by the same company, around $11.5 billion.
And it is fifty-seven times the twenty-year lease on 191 megawatts that we described in August as a considerable commitment.
What a compute commitment covers
Precision matters here, because this is not an immediate expense.
A commitment of this kind is spread over years, sometimes decades. It corresponds to capacity purchase contracts, data centre leases, power supply agreements. The company will not pay out $517 billion tomorrow; it undertakes to pay it over time.
This is precisely what we called off-balance-sheet commitments in our guide to reading the numbers: they appear neither in gross margin nor in operating profit, and yet they determine the soundness of the whole.
This disclosure comes in the week the company tightens the quotas on its coding tool, a change we covered on Monday. The two facts tell the same story: serving agents that loop costs a great deal, that cost grows faster than revenue, and you must simultaneously secure future capacity and curb present consumption. It is not inconsistent. It is the portrait of a company running to keep up with its own demand.
The two readings
The favourable reading. Securing compute today costs less than seeking it in five years. We have documented the strain on components, the constraints on grid connections and local opposition to new sites. A company that locks in its capacity is making a rational bet on scarcity.
The cautious reading. A commitment of this scale assumes that revenue growth holds for its entire duration. We reported an announced gross margin above 80%, which is excellent at product level. A high margin on revenue that stopped growing would not cover commitments calibrated on continuous growth.
Why this figure is coming out now
The timing is probably not neutral. A company preparing for a stock market listing has to disclose its commitments, and it is better to do so before the official document does it for you.
This is, moreover, exactly what we predicted: a prospectus would force the publication of what appears in no standard metric. Early disclosure is a way of preparing the ground.
What to take away
This figure does not say a company is in trouble, and it does not say it is sound. It says that this sector's business model rests on an assumption of long-term growth, and that this assumption is now quantified.
It is useful information because it is verifiable over time. We will know, quarter after quarter, whether revenue follows the trajectory these commitments assume.
And that is probably the right way to follow this story: not by debating bubble or soundness, but by comparing two curves, one of which is now public.