017Can Anthropic’s compute keep up with its demand?

What the S-1 says about demand

Customers are spending more, faster than anything we can compare it to. Net revenue retention is about 900%: customers billed in last year’s third quarter paid nine times as much this quarter. Customers paying $100K a year went from about 950 to more than 11,900 in a year, and doubled in the third quarter alone. Over half of them now use both the platform and one of Anthropic’s apps.

Demand follows the models. Over 80% of third-quarter revenue came from models released in the six months before. Revenue growth slowed to 45% in the third quarter, from 144% in the second; part of that was Anthropic cutting off customers who broke its terms, mostly by distilling Claude, which had brought in 5–7% of revenue.

And compute is the limit Anthropic names. The filing says demand for compute “has historically exceeded available capacity” across research, serving and its own use, which it runs as one pool; that it has asked its cloud providers for more compute than they were willing to provide; and that it has changed usage and rate limits. It doesn’t say how much customer demand went unserved.

How far along is that?

About 2% of the way, counted by customers. If every company with $1B of revenue put 1% of it into AI, about 15,000 would spend $10M a year. Anthropic has about 250 at that level, 1.7%. At 5% the field widens to about 72,000 companies with $200M of revenue or more, and Anthropic has 0.35% of them.

The real share is probably lower. The S-1 doesn’t say who the 250 are. Some are likely AI-native companies whose products run on Claude, which aren’t among the big companies counted here. Counting only companies with $1B of revenue, the share paying Anthropic $10M a year is below 1.7%.

Counted in dollars, further along, but not far. Anthropic’s run-rate of about $96B is 10% of the $0.97T a year that 1% of big-company revenue would be, and 1.6% of the $6.0T at 5%. All AI model spend, Anthropic’s and OpenAI’s together, comes to about 0.16% of the revenue of the world’s $1B+ companies today, and not all of it comes from them.

Is 5% a stretch? Worldwide IT spending is about $6.3T this year, by Gartner’s count; 5% of the revenue of every company that could spend $10M on AI is about the same, so 5% would make AI about as big as all of IT. The S-1 puts the work of knowledge workers at more than $40T a year, and 1% of revenue is small next to a payroll. Nothing in the S-1 says where spending settles, but it isn’t near either number yet.

Where the efficiency comes from, and where it goes

Serving got 6× cheaper per dollar of revenue in a year. Cost of revenue was 88¢ per $1 of revenue in the first half of 2025, 51¢ in the second and 15¢ in the first half of 2026, an 85% gross margin Anthropic kept in the third quarter. On OpenRouter, the average list price paid per Claude token rose over the same period, from $3.03 to $3.87 per million as usage moved to Opus. Put the two together and the cost of serving a Claude token fell about 4.7× in a year. The S-1 credits its mix of chips (Amazon’s Trainium, Google’s TPUs, NVIDIA and AMD) and what it calls compute multipliers, improvements in training and serving that it says Claude now helps its researchers find faster.

So far it has kept the savings. The gross margin went from 12% to 85% in a year while the price paid per token rose. That is how revenue grew 12× in a year while compute grew 4.8×. It can’t repeat: at 15¢ of cost per dollar of revenue, there is little margin left to gain.

From here, the savings go to prices. The S-1 says efficiency lets Anthropic cut what customers pay “while preserving our margins,” and October shows it starting. Opus 5.5 does a task for about 40% less than Opus 5 did, and lists at $4 and $20 per million input and output tokens, against $15 and $75 for Opus 4.1 fourteen months ago; Haiku 5.5 lists at a tenth of Haiku 4.5’s price. On OpenRouter, the average price paid per Claude token fell 30% in October from the third quarter. Revenue per dollar of compute serving customers has held near $6.70 since the start of 2026, and we assume it stays there.

Research takes most of the compute. Serving customers took 25.5% of compute and infrastructure in the third quarter, from 43.6% in the first half of 2025; the rest goes to research, training and Anthropic’s own use of Claude. Revenue per $1 of all compute rose from $0.40 in the first quarter of 2025 to $2.08 in the second quarter of 2026, then fell to $1.70 in the third, as compute grew 78% in a quarter, mostly for research.

Our projection. We expect the tokens Anthropic serves per dollar to keep doubling each year, 45× the first half of 2025 by the end of 2029 (1.5× to 3× a year gives 17× to 175×), slower than the last year’s 4.7×. And with the margin near its ceiling, we expect Anthropic to pass nearly all of it to customers, as it began to in October, so revenue per gigawatt rises only as fast as what a gigawatt costs, which we take as about 5% a year. Revenue per $1 of compute then stays at $1.70 if research keeps its share, or rises to $2.67 if Anthropic shifts compute to serving until serving takes 40%. Moving compute from research is the biggest lever it has: each point of compute moved to serving adds about 4% to the revenue it can serve.

Short by the end of 2027, on two of three paths

On the reported trend, Anthropic needs about 13 GW at the end of 2027, a third more than the roughly 10 GW it has lined up, and 34 GW at the end of 2029 against about 25. At today’s split between research and serving, its compute can serve about $231B a year of revenue at the end of 2027, 25% short of the trend’s $309B. Our tracker’s faster path, $425B, would need 18 GW.

Both paths start above what Anthropic expects. The reported trend ends this year at $113B and our tracker at $126B. At its investor presentation the company expected about $110B (Reuters), close to the $110B we estimate its compute can serve on December 31; the FT reported annualized revenue was expected to stay below $100B, without giving a date. Slower demand, a cautious forecast or compute that is already binding would each explain the gap, and the filing doesn’t say which.

Only Anthropic’s plan fits. $110B at the end of this year and about $195B of revenue in 2028 would leave its compute 44% spare at the end of 2027. On reported plans, Anthropic is lining up compute for more demand than the plan behind its IPO valuation shows.

It can close much of the gap by moving compute from research to customers. Serving takes only about a quarter of its compute. At 40% the same gigawatts serve about $363B at the end of 2027, more than the reported trend needs, at the cost of slower model progress, which the S-1 itself lists as a risk. The other valve is price: if Anthropic held prices instead of passing the next round of efficiency on, each gigawatt would earn about twice as much a year later. Either way a shortage shows up as rate limits, rationing and prices that stop falling, and demand bends down toward the compute line.

The answer leans on one estimate: how much compute Anthropic ran in the third quarter. We take about 3.1 GW, on the way from about 1.4 GW at the end of 2025 to 5 GW at the end of 2026, which makes a gigawatt worth about $22B of revenue a year. It is the softest number here: on that path, compute spending per gigawatt rose from $9.8B a year in the second quarter to $12.7B in the third, 29% in one quarter. At the second quarter’s cost, Q3 ran on about 4.0 GW, and the 10 GW lined up would serve about $179B at the end of 2027: 42% short of the reported trend, and only 12% spare on the plan. Starting from the 2 GW 22V Research estimated for early 2026 gives $203B. Our assumption that Q3’s serving compute was fully used cuts the other way: if it had room to spare, the same gigawatts serve more.

The question for the $1T+ raise is compute, not demand

The market’s worry runs the other way. The hyperscalers, neoclouds and labs need about $1.5T of outside money through 2030 (pack 005), and the fear is that the demand won’t be there to pay it back. Sized against company revenue, demand at 1% would need about 44 GW, research and training included, a quarter of the 175 GW of AI capacity the world has by the end of 2029. At 2% it needs 97 GW, more than half, with consumer apps and advertising still to serve. At 5% it needs 272 GW, more than all of it.

The S-1 points up the range, not down. Customers billed a year ago spend nine times as much; the number spending $100K a year grew 12.5×; and Anthropic, which knows its pipeline, is lining up compute, on reported plans, well beyond what its own IPO plan needs. Pack 010 has the whole market about 19% short at the end of this year and about level at the end of 2027 before demand pulls ahead again, on demand that doesn’t assume any particular share of revenue.

What would make us wrong. Growth slowed to 45% in the third quarter, and some of the cushion was distillation revenue Anthropic chose to give up. Five customers brought in about a quarter of revenue in the nine months to September, and about a fifth in the third quarter; the S-1 doesn’t name them. And demand is price-elastic: the S-1 credits lower prices per task for driving usage, so if a shortage keeps prices up, demand grows more slowly. A shortage would look like that: slower growth at high prices rather than a glut.