Metered intelligence: AI sold by the token

How it makes money

Start with the meter. A developer sends a model a request. The request and the answer are both counted in tokens, small pieces of text, and the bill is the count times the price of the model chosen. Anthropic's S-1 puts the whole business model in one sentence: customers "pay for their usage of our intelligence service, which is metered in tokens and priced according to the model selected."

Walk it through the choices that make up any business model and the shape is unusual. The payer is usually a company building its own product on the model, so the seller's customer has customers of its own. The unit sold is work done, not a seat or a license. The price follows the model, so a more capable model costs more per token. The work is done by software running on computing power, which makes compute the cost line that decides the margin. And the meter itself holds no one: the next request can go to another model.

One filing, read closely

The clearest public record of the pattern is Anthropic's S-1, filed with the SEC on October 26, 2026. Pay-for-intelligence, the company writes, "is our primary revenue model and currently represents the substantial majority of our total revenue." Revenue for the first half of 2026 was $16.3 billion, up from $1.2 billion a year earlier.

Serving the tokens is cheap. In the S-1's table of costs by function, cost of revenue for the half was $2.4 billion, almost all of it the computing power that answers customers' requests. That is a gross margin of about 85%. A year earlier it was about 12%.

Building the next model is not cheap at all. All compute and infrastructure came to $8.9 billion in the same half, 55% of revenue. At least $6.6 billion of it sits outside cost of revenue, most of it in research and development: the computing power spent on models that earn nothing yet.

A meter runs both ways

Metered revenue moves with use, in both directions. Total customer spending fell about 20% in the two weeks to January 2, 2026, the S-1 reports, and many of the largest customers can cut what they spend at any time. That is why the company keeps its Claude Pro and Max subscribers out of its customer-expansion measure: their "fixed-price subscriptions do not reflect the expansion and contraction behavior" the measure is meant to capture.

And the unit keeps getting cheaper. Model versions released in the prior six months made up over 80% of revenue in the third quarter of 2026, and the S-1 says its newest flagship cut customers' cost per task by about 40% against the one before, through a lower price and better efficiency. A metered seller grows by selling far more work at a falling price per task.

The mortgage underneath

The cost side runs on a different clock. Anthropic buys most of its computing power under multi-year commitments, ordered 12 to 18 months before it arrives, and warns that margins may compress when new capacity comes online ahead of the revenue to fill it. So far it has grown into its commitments fast: compute and infrastructure fell from 251% of revenue in the first quarter of 2025 to 48% in the second quarter of 2026.

The meter also runs through other people's stores. In the first half of 2026, 48% of revenue came through the marketplaces of the big cloud providers, and the S-1 names Amazon, Google and Microsoft as investors, customers, distribution partners and competitors at once. Those clouds sell their own capacity the other way, on contract: Microsoft alone reported $625 billion of commercial contracts not yet delivered at December 31, 2025.

How it fails

The S-1 is plain about the ways this goes wrong. The company says it has badly underestimated its compute needs before, which left it short of capacity. It warns that it may end up tied to older hardware while competitors get newer systems, or paying premium prices for resources that have since lost their value. Each is the same failure seen from a different side: a meter that cannot carry the mortgage, or a mortgage too small for the meter.

Questions

What is metered intelligence?

A business model in which an AI company charges by use: customers pay for the tokens their requests consume, at a price set by the model they pick. Anthropic calls its version pay-for-intelligence.

How is token pricing different from an AI subscription?

A subscription charges a fixed fee for access, however much it is used. Token pricing charges for the work done, so the bill rises and falls with use. Anthropic sells both, and its S-1 says the metered side is the substantial majority of its revenue.

Is selling AI by the token profitable?

On the tokens themselves, it can be: Anthropic's S-1 shows a gross margin of about 85% for the first half of 2026. Counting all of its computing power, including the models still in development, compute came to 55% of revenue.

Who sells AI by the token?

Model makers sell access to their models directly and through the marketplaces of the big cloud providers. Anthropic's filing describes the economics in the most detail; in the first half of 2026, 48% of its revenue came through those marketplaces.