The most direct way an AI lab earns money is by selling access to a model through an API, and the price is quoted per million tokens. A token is the chunk of text, often a word or a piece of one, that a model reads or writes at each step, and a bill adds up two kinds. Input tokens are what you send, the prompt and any documents attached to it. Output tokens are what the model writes back, and they cost more each. The prices here are what each company’s own page said when we read it, and prices change. Anthropic’s pricing page listed Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, with a read from the prompt cache at $0.10. Google’s page, last updated Oct 26, 2026, listed Gemini 3.5 Flash-Lite at $0.30 for input and $2.50 for output, and says that output includes the model’s thinking tokens. OpenAI’s page listed gpt-6.1-sol at $2.00 for input, $0.10 for cached input and $10.00 for output, for prompts of up to 272,000 tokens. Mistral’s page gave Mistral Large as $0.5 in and $1.5 out, and said batch processing cuts the price by half and cached input by up to 90 percent. Within one price list the spread can be a hundredfold: Anthropic’s page lists Claude Haiku 5.5 at $0.10 for input on prompts up to 100,000 tokens and Claude Fable 5.1 at $10. Two cautions apply to every number here. Google’s page says its output price includes thinking tokens, so a model that reasons at length bills for the reasoning as output. And a token is not a fixed unit: Anthropic’s page says its models from Claude 4.7 on use a newer tokenizer that produces approximately 30 percent more tokens for the same text, so a price per million tokens can hold still while the price of a document rises.
The second line of business is the subscription, which turns a per-token bill into a flat monthly price with a cap on use. Anthropic’s plans page listed Claude Pro at $20 a month billed monthly, Max from $100, and team and enterprise plans priced by the seat; the enterprise plan was $20 per seat per month, billed annually with a 20-seat minimum, plus usage at API rates. The page says usage on paid plans resets on a rolling five-hour window and carries weekly limits as well. Mistral’s page listed its Le Chat Pro plan at $14.99 a month. With a flat price, the cap and not the meter limits what a subscriber can use. The enterprise arrangement shows the two models can be combined: a fee for the seat and a metered charge for what the seat consumes.
A third line runs through the clouds. Anthropic’s pricing page notes that Claude models are available through Amazon Bedrock and on Google Cloud, where the cloud provider invoices the customer, so for those customers the bill comes from the cloud company and not from the lab. The deepest version is Microsoft and OpenAI. In a post dated Oct 28, 2025, Microsoft said it held roughly 27 percent of OpenAI Group PBC on an as-converted diluted basis, a stake valued at about $135 billion, and that OpenAI had agreed to buy an incremental $250 billion of Azure services. Microsoft’s annual report for the year to Jun 30, 2026 says the partnership was extended in October 2025 and again in April 2026, so those terms are a snapshot with a date, not a fixed arrangement.
That annual report is also where the money becomes visible, and the lesson on reading a 10-Q applies to it. Microsoft reports three segments, and its Intelligent Cloud segment, which its Item 1 describes as server products and cloud services, had revenue of $137,791 million in the year to Jun 30, 2026 against $106,265 million a year earlier, and operating income of $56,972 million against $44,589 million. The report says Azure and other cloud services revenue rose 41 percent. The segment and product-line tables show no separate line for AI. The report describes the Azure line as including cloud and AI consumption-based services, along with GitHub and other cloud services, so an AI business can be growing inside a number that only names the cloud. NVIDIA, which sells the chips, reports in its 10-Q for the quarter ended Jul 26, 2026 a Compute & Networking segment, which includes its data center accelerated computing and networking platforms, of $88,299 million against $41,331 million a year earlier, out of total revenue of $96,221 million against $46,743 million. The other place a lab appears is as an investment. The same Microsoft report says net gains from investments in OpenAI raised net income by $5.0 billion, or $0.67 a share, in the year to Jun 30, 2026, while net losses lowered it by $3.6 billion, or $0.49 a share, the year before. Those figures describe Microsoft’s accounting for its stake. They are not OpenAI’s revenue.
What a private lab owes the public is much less. A company becomes an Exchange Act reporting company, and must file annual reports on Form 10-K, quarterly reports on Form 10-Q and current reports on Form 8-K, when it lists securities on a US exchange or when it has assets above $10 million and a class of equity held by 2,000 or more holders of record, or by 500 or more holders who are not accredited investors. A company raising money privately from accredited investors under Rule 506(b) is not required to give them specified disclosure documents; it files a notice, Form D, with the SEC within fifteen days of the first sale. So the figures that circulate about private labs, annualized revenue, valuations, user counts, are statements by the company or by the press, not filings, until the company registers. That boundary is moving for at least one lab. A company preparing a US listing can ask the SEC staff to review a draft registration statement non-publicly, and in return it commits to filing the registration statement and the earlier drafts publicly at least 15 days before its roadshow starts. EDGAR lists a first draft registration statement from Anthropic, PBC dated Jun 1, 2026 and a public Form S-1 dated Oct 26, 2026. When a lab’s numbers do arrive in a filing, read the filing, not a summary of it: the questions to put to it are the ones this path has been building, which are where the revenue comes from, how much of it is usage and how much is subscription, what the compute costs, and who the partners are. Why the compute costs what it does is the subject of the last lesson.