AI is sold by the token, a small chunk of text, so the price of a token and the number of tokens people use are the two numbers behind the whole business. Token Intelligence is our recurring brief on both: what AI costs, who is paying, and whether the economics hold. This first issue pulls together four figures from published sources.
1. 55.9%: businesses paying for AI
In Ramp's October index (September data, published October 14), 55.9% of businesses pay for AI, up 0.62 points on the month. 43.8% pay Anthropic and 40.6% pay OpenAI. Ramp notes that most businesses already pay for AI, so growth may now come from heavier use by existing customers more than from new ones.
2. Usage up, spending roughly flat
Ramp reports that total AI spend reached a new peak in early October, about 1.5% above August's peak. Coverage of the report by ExplainX says token volume keeps rising while token spending has barely grown for two months, and attributes that to falling prices. We didn't find that second point in Ramp's written report, so treat it as reported. Ramp also says 80% of OpenAI and Anthropic spend comes from 1% of firms.
3. About $1.80: revenue per dollar of compute
Anthropic's S-1 shows $16,286 million of revenue and $8,945 million of compute and infrastructure expense for the first half of 2026. That's roughly $1.80 of revenue per dollar of compute, by our arithmetic. Snowflake CEO Sridhar Ramaswamy puts the same comparison another way: compute expense was 55% of sales, against 201% a year earlier. His figure matches ours. We covered the filing in our Anthropic S-1 piece.
4. About 80%: revenue that can move quickly
Per Fortune's reading of the S-1, about $13.1 billion, roughly 80%, of Anthropic's first-half revenue was usage-based enterprise fees, and customers can reduce usage on short notice. About $10.8 billion is locked in under multiyear contracts. Ramaswamy's point is that this is the opposite of the long-term cost commitments on the other side of the ledger.
What this means, in plain English
Falling prices per token are good for people who buy AI and hard on people who sell it. If a seller's price per token drops, it needs more tokens just to earn the same money. The first two numbers say usage is rising and spending isn't keeping pace. The third says Anthropic is currently earning more than it spends on compute. The fourth says that revenue can change quickly. Whether volume keeps outgrowing price cuts is the question we'll be tracking.
What we're watching next
Ramp's next index. Any change in per-token prices from major providers. We're building a price table and will add it once each price is checked against the provider's own page. And whether usage-based revenue shifts toward fixed contracts.
Sources: Ramp AI Index, October 2026, ExplainX on the Ramp index, Anthropic, PBC Form S-1, Fortune, and Sridhar Ramaswamy's thread on X.