Go to Groundby Jaidev Karthickeyan
← All notes Markets · November 2026

QuarryIs frontier AI becoming a better business, or only a bigger one?

Anthropic's S-1: revenue per compute dollar rose from $0.40 to $2.08

Anthropic's revenue grew 14 times in a year while its compute and infrastructure cost grew 3.8 times. The AI race is becoming a capital allocation problem: the winners will generate the most economic output from every dollar invested in compute, research and distribution.

What happened

In the first quarter of 2025, every dollar Anthropic spent on compute brought back about 40 cents of revenue. Fifteen months later, the same dollar brought back $2.08. For an industry often accused of burning money to heat data centres, that is a remarkable sentence to be able to write.

The S-1 has the numbers. Revenue went from $1.16bn in the first half of 2025 to $16.29bn in the first half of 2026, about 14 times. Compute and infrastructure costs went from $2.33bn to $8.95bn, about 3.8 times. The second quarter of 2026 was the first in the filing with an operating profit: $974m.

The ratio is not a margin, and it isn't return on capital. The compute line covers serving customers, training models, research and internal use, and it is cost as booked, not cash committed. But I think it is the cleanest public measure yet of how fast a frontier lab's revenue is outrunning its biggest bill.

Update, 3 November 2026. The amended S-1 adds the third quarter, and it cuts the other way. Revenue rose to $16.96bn, but compute and infrastructure cost rose faster, to $9.86bn, so revenue per compute dollar fell to $1.72. Compute went from 48% to 58% of revenue and the operating margin from 8% to −3%. Anthropic had warned that more short-term compute, which costs significantly more than its long-term contracts, would compress margins in the near term. The chart now uses the filing's own quarterly figures.

Going to ground

First order

Most of the compute builds the product, not the service. Software usually costs a lot to write and almost nothing to sell. AI costs a lot twice: once to train the model, and again every time a customer uses it.

The filing splits the two. In the first half of 2026, compute and infrastructure cost $8.95bn, but cost of revenue, which Anthropic says is substantially all compute for customer workloads, was only $2.38bn. That is 14.6% of revenue to serve customers and about 54.9% on compute in total. The rest goes on research, new models and internal use.

So Anthropic is not really an inference business with a research hobby. It is a research engine that pays for itself through usage.

Second order

The price per token is the wrong number. What matters is how much useful work comes out of each unit of compute. A cleverer model may use more compute per answer and still be cheaper per job, if it gets there in fewer attempts with less human tidying up.

The launch of Opus 5.5 shows how slippery that is. Anthropic cut list prices by 20% and estimated that typical workloads would cost about 40% less to run than on Opus 5. An independent benchmark by Artificial Analysis, run at maximum effort, found the opposite: cost per task rose about 21%, because the model used more tokens. Both can be true. That is rather the point.

Unlike a software licence, every extra task carries a real cost. The commercial question is whether cheaper tasks bring in enough new work to raise total revenue and contribution profit. Whoever understands that loop best will price better and build better.

Third order

The money may end up in the system around the model. Capability still sells: model versions released in the six months to 30 September earned over 80% of Anthropic's third-quarter revenue. But a model on its own does nothing in a business. It needs data, permissions, tools, memory, orchestration and somewhere to run.

Those layers don't share the value equally. An enterprise can swap model providers while keeping its applications, data and workflows. The company that owns the workflow and the customer may end up with more pricing power than the clever supplier behind the scenes. The model with the best benchmark score is not necessarily the one with the best return.

Ears up: what to watch

  • Revenue per unit of compute. Does it climb back above $2 as agentic workloads eat more inference?
  • Cost per completed task, including supervision and error correction, not the token price.
  • Contribution margin on the next dollar of usage. Revenue growth alone proves nothing.
  • Capacity planning. Anthropic buys compute 12 to 18 months before it arrives, so it is always betting on demand it can't see yet.
  • Who owns the customer. Workflow, data and the execution environment decide whether pricing power stays with the model maker.

Bone of contention

I would not declare the economics of frontier AI solved yet. The ratio is partly about timing: compute is bought 12 to 18 months ahead and expensed as it comes online, sometimes before the revenue it will support. Anthropic itself expects more short-term compute to squeeze margins. The third quarter has already shown how far the ratio can move, from $2.08 to $1.72 in three months.

The numerator flatters it too. Revenue is booked gross, including sales through cloud marketplaces, which the filing puts at 44% of revenue in the first nine months of 2026. And revenue per compute dollar says nothing about whether research earns its keep: a model is a long investment that pays back over several product cycles.

Then there is value capture. If open-weight models keep pushing down the price of intelligence, model makers may face pricing pressure and the rents may drift to whoever controls data, workflow and distribution. Or frontier models keep doing work cheaper ones can't, and the intelligence layer keeps its pricing power. Both stories are still live.

Either way, the contest is not just between the best models. It is between economic systems for producing and deploying intelligence. The winners will get the most useful work out of every dollar of compute and keep enough of the customer to collect for it. The unit of competition is moving from the model to the finished task.

J. & the researcher

Sources
  1. Anthropic, PBC, Form S-1, filed 26 October 2026, and S-1/A: quarterly revenue and compute (Q1 2025 to Q3 2026), CSP marketplace share, gross recognition; half-year revenue, compute and cost of revenue; Q2 2026 operating income; model-version revenue share; compute procurement timingconfirmed
  2. Commentary on cloud-marketplace revenue and gross recognition: TokenPost and The Data Letterreported
  3. Opus 5.5 pricing and Anthropic's 40% workload estimate, as reported by IT Brief; Artificial Analysis cost per task, as reported by AlphaSignalreported

Data for the chart: anthropic-unit-economics.csv