Every argument about AI money starts with a very large number and ends with a feeling. This one follows the invoices instead: what companies actually spend on tokens, what the sellers earn on them, and who is quietly paying for the machinery underneath.
The short version: prices are collapsing, demand is concentrated in a few thousand firms, one lab has started earning, and five companies are carrying the capital cost of the whole thing.
By the numbers
$11.38 — what the median US business spends per employee, per month
725× — how much the price of a fixed answer fell in 18 months
$730B — what five hyperscalers will spend on capex in 2026
The price of an answer fell 725-fold
The cleanest measure of AI deflation does not come from price lists. It comes from asking what it costs to reach a fixed score on a hard exam.
Epoch AI estimates that the cost of a given level of AI performance has fallen about 47% per quarter since 2023 — roughly 13× a year, faster than DNA sequencing, Moore's-law compute, or electricity in the century before 1973. In January 2025, OpenAI's o3 could score 75% on GPQA Diamond, a set of PhD-level science questions, for about 30 cents a question. Eighteen months later, GPT-5.6 Luna reached the same score for $0.0004.
The list prices tell the same story in slower motion. GPT-4 cost $60 per million output tokens in March 2023; GPT-6 Luna costs $0.50 today, from the same vendor.
Most companies spend almost nothing
Public companies describe the sellers. For the buyers, the cleanest dataset is Ramp's AI Index, built from the card and bill-pay records of more than 70,000 US businesses — invoices, not survey answers.
- The median business spends $11.38 per employee per month: roughly one subscription seat.
- The top 10% spend $611.
- The top 1% spend $9,568 — and account for 80% of all spending on OpenAI and Anthropic.
Which means "companies are adopting AI" is a weak sentence. A few thousand firms are running serious inference; everyone else is buying a seat. It also punctures the chip-industry talking point that engineers should carry token budgets worth half their salary — even the top percentile is nowhere near it.
There are signs the curve is flattening. Through August and September, token volumes kept rising while token spend went flat, and per-employee spend for the heaviest users fell 3.1% from its July peak. Deflation is outrunning consumption, for now.
One lab earns, one burns
Anthropic told prospective investors in October that third-quarter revenue reached $16.8 billion, up from $11.6 billion in Q2 and $4.7 billion in Q1 — with an 85% gross margin, a positive adjusted operating margin and positive free cash flow, ahead of an IPO.
It also takes 60.1% of token dollars on Ramp's data while being used by 43.8% of businesses. Its share of the money is far larger than its share of the customers: the biggest spenders pay Claude.
OpenAI's first quarter was $5.7 billion of revenue against $3.7 billion burned. The FT reported a company presentation projecting $278 billion of negative cash flow between 2026 and 2030, against roughly $600 billion of compute commitments through the end of the decade.
That is not a scandal; it is a bet that capacity signed today will be worth more than it costs. But it is the opposite of what the word "currency" implies. One lab now has the economics of a software business. The other has the economics of a construction project.
The $730 billion nobody invoices
The capital is not in anyone's token bill. It sits on five balance sheets: Microsoft, Alphabet, Amazon, Meta and Oracle.
Consensus for their 2026 capital spending rose from about $485 billion in January to roughly $730 billion in July, according to Reuters. By 2027, their combined capex is expected to exceed their combined free cash flow — $1.57 of extra investment for every extra dollar of cash.
Microsoft shows the shape of it: $331 billion of revenue in fiscal 2026, $214 billion from Microsoft Cloud, Azure past $100 billion, and $41 billion of capex in the fourth quarter alone.
Two consequences follow. The people who bought the factories need volume to fill them, which is exactly why per-token prices keep falling. And when a supplier is also an investor in its customer, the price of a token stops being a clean market signal.
Where the value lands instead
- Silicon. The scarce input is accelerators; the reliable profit is in selling them, not in reselling inference.
- Cloud capacity. Microsoft's cloud line is now larger than most software companies, and its growth is tied to AI demand.
- Routing. Stripe agreed in August to acquire OpenRouter, which routes requests across more than 400 models. If tokens are a commodity, the margin is in choosing which one to buy, per request.
- Accounting. The Linux Foundation's new Tokenomics Foundation — 30 members, including JPMorganChase, Oracle and SAP — is standardising how AI cost and value get measured.
The exception: discounted credits
One corner behaves more like money than the rest: prepaid credits. Researchers at Vectoral documented brokers reselling unused inference credits at 30–80% below list, through relays and credit marketplaces.
It is tempting to read that as a currency emerging. It is the opposite. Discounted credits are what happens to a prepaid liability — gift cards, airline miles, expiring subscriptions — when buyers over-commit and cannot consume what they bought. Money does not trade at 60 cents on the dollar because someone overbought. That is inventory.
What this means if you are paying
Measure cost per completed task. The price per million tokens is the one number you cannot control and the one that matters least. A cheaper model that needs three retries is not cheaper.
Expect the bill to rise as you add agency. Routing a task to an agentic model costs a provider at least five times a chatbot turn, and agentic models need 5–30× more tokens per task.
Cache and route before you negotiate. Repeated context is the largest controllable cost in most agent workloads.
Watch margins, not launches. A vendor losing money on inference has a reason to change pricing, packaging or terms — usually at your next renewal.
The honest summary
Demand is real and concentrated. Prices are collapsing faster than usage grows. One lab has started to earn, the largest is still burning, and the capex is carried by five companies whose free cash flow is about to be consumed by it.
The token was never a currency. It is a line item — and in 2026 it is the line item that decides which of these companies becomes a utility and which becomes a cautionary tale.
The full report
I wrote this up as an eight-page report with the charts, the source ledger, the caveats and a method section: every figure traceable, every unaudited number labelled.
The full eight-page report collects every figure, chart and source in one place, with the caveats and method notes attached.
Sources
- Epoch AI: The plunging price of thought
- Gartner: inference costs per agentic workflow
- Gartner: inference costs to 2030
- Google I/O 2026 keynote
- Ramp AI Index and the October 2026 update
- Reuters: OpenAI burned $3.7bn in Q1 2026
- Reuters: OpenAI forecasts $278bn cash burn
- Reuters: OpenAI compute spend to 2030
- Bloomberg: Anthropic's $16.8bn quarter
- Reuters: hyperscaler capex and free cash flow
- Microsoft FY26 Q4 earnings call
- OpenAI API pricing and Anthropic model pricing
- Andreessen Horowitz: LLMflation
- Stripe: agreement to acquire OpenRouter
- Linux Foundation: Tokenomics Foundation
- Vectoral: how token brokers sell discounted AI access
No comments yet