Skip to content
HomeBriefFor youTrackersTopicsCompaniesGlossaryNewsletterSecurityHealth & medicineFinance & marketsRoboticsModels & researchPolicy & regulationEnterprise & businessDeveloper tools
Explainer · beginner · 4 min read

How AI companies make money: the six revenue lines behind a frontier lab, and why the costs arrive before the cash

Anthropic's listing puts a lab's income statement in public for the first time. Here is where the money comes from (API tokens, subscriptions, enterprise contracts, cloud resale, licensing, and investor-funded compute), what each line costs to deliver, and which ones scale.

Updated 2026-11-19 · For: Beginners Finance Executives Developers

An AI lab looks like a software company from the outside and like a utility from the inside. The product is a model, but the business is renting out the computers that run it, and almost every line of revenue is a different way of metering that rental. Anthropic's prospectus, which became public in November 2026, is the first full set of accounts for a frontier lab, and it makes the structure visible: about $17 billion of revenue in a single quarter, an operating loss in the same quarter, and $895 billion of compute commitments stretching years ahead. The sections below walk through where the money comes from and why the costs are front-loaded.

1. API tokens: usage-metered revenue

The core line is the application programming interface. Developers send text to a model and are billed per million tokens, with separate rates for input, output and cached input. It is the purest form of the business: revenue scales with usage, and the cost of serving it, inference compute, scales almost in step. Margins depend on three levers: how efficiently the model runs on the hardware, how much customers use caching and batch tiers (which are cheaper for the lab to serve), and how fast the lab can cut prices without losing share. Prices for a flagship model have fallen roughly 20% per generation this year, so API revenue growth has to come from volume, which is why labs publish usage figures so eagerly.

2. Consumer subscriptions: fixed-price, variable-cost

Chat apps sold at $20 to $200 a month are the opposite shape. Revenue per user is fixed; cost per user depends on how much they use, which is why every vendor rations heavy users with usage caps, slower models at peak times and 'fast mode' upsells. Subscriptions are valued by investors for their predictability and their distribution: a consumer app is the funnel that brings developers and enterprises to the API.

3. Enterprise contracts: the committed-spend layer

Large customers sign annual or multi-year agreements that commit to a minimum spend in exchange for discounts, dedicated capacity, higher rate limits and contractual terms on data use. These contracts are what let a lab sign its own commitments to cloud providers: a book of committed customer spend is collateral for a book of committed compute purchases. Read the concentration note in any filing. If a handful of customers make up a large share of revenue, and some of them are also investors or suppliers, the growth is less diversified than it looks.

4. Cloud resale: the same model, sold by someone else

Amazon Bedrock, Google Vertex and Microsoft Azure sell lab models to their own customers and pay the lab a share. For the lab this is lower-margin revenue but it reaches buyers who will never sign a direct contract, and it comes bundled with the cloud credits that fund training. For the cloud it is a reason for customers to stay. The circularity is the point: the investor, the supplier and the distributor are often the same company.

5. Licensing and special-purpose models

Labs increasingly sell restricted models to vetted customers at premium prices: cyber-defence models, life-sciences models, government-only tiers. Volumes are small but the price per token can be several times the public rate, and the contracts carry services and support revenue that is closer to a traditional enterprise-software margin.

6. The line that is not revenue: investor-funded compute

The largest cash flows in a frontier lab are not sales but purchases. Training a frontier model and reserving the capacity to serve it means committing tens of billions of dollars years before the revenue that will pay for it exists. Those commitments are funded by equity raises and by strategic investors who are also the suppliers, which is why a lab can report an operating loss and a growing cash pile in the same quarter. Compute spent as a share of revenue is the single most useful efficiency metric: Anthropic disclosed a figure around half of revenue in its filing, and the direction of that ratio over time tells you whether the business is converging on a software margin or staying a utility.

Why the costs arrive first

Every revenue line above depends on a model that has already been trained and on capacity that has already been reserved. The spend is a step function, the revenue a curve that follows it by quarters. A lab that stops investing shows a beautiful quarter and then falls behind; one that keeps investing shows losses and a lead. Public investors now get to vote on that trade-off every ninety days, which is the real change that a listing brings.

What to look for

For an investor: the split of revenue by line, the concentration of customers, compute cost as a share of revenue, and the gap between committed customer spend and committed compute purchases. For a buyer: which line you sit in, because API customers get price cuts first and enterprise customers get capacity first. For a builder: the cached and batch tiers, which are the parts of the price list the lab most wants you to use.

Common questions

Are AI labs profitable?

Mostly not on an operating basis, and deliberately so. Revenue has grown fast enough that some have posted individual profitable quarters, but each new model generation requires a step up in training and serving spend that is booked before the revenue arrives. The metric to watch is compute cost as a share of revenue, not the quarterly profit line.

Why do prices keep falling if the companies are losing money?

Because inference gets cheaper per token with each hardware generation and each efficiency gain, and because share matters more than margin while the market is forming. A price cut that fills reserved capacity which would otherwise sit idle costs the lab little and locks in developers.

Who are the biggest customers of an AI lab?

Filings show a mix of large enterprises, software companies that resell AI features, and the cloud providers that distribute the models. In several cases the largest customers or distributors are also the largest investors and the largest suppliers of compute, which is why analysts read the related-party notes closely.

Get the brief in your inbox

One email every morning: the stories that mattered in AI, each with a one-line "why it matters" for your role. Free, no ads, unsubscribe any time.

Published daily by AI Briefing. We never share your address.
Like this? Get tomorrow's brief by email, free.