Model 1: Model and inference APIs

Last updated October 16, 2026

Model 1

Model and inference APIs

Revenue grows with your customers' success, and so does your concentration risk.

Model APIs sell intelligence by the token to companies that build it into their own products. It's the purest usage-based model in AI: revenue grows when a customer's product grows, with no seats and, until an account gets big, no sales call. It isn't only for the big labs either. Sacra, a research firm that tracks private tech companies, estimates that Together AI, which runs open-source models for developers, reached about $1 billion in annualized revenue by February 2026⁸ and Fireworks AI, a similar company, about $800 million by May 2026⁹. The catch is that prices fall fast, customers shop around, and revenue piles up in a handful of accounts.

Revenue rides on someone else's growth

An API provider doesn't sell to end users. It sells to developers, whose usage rises and falls with their own customers. OpenAI says more than 9,000 organizations have processed over 10 billion tokens through its API, and nearly 200 have passed 1 trillion each¹. Pricing is built for that ramp. On the OpenAI API, an account moves up spend tiers automatically as its total purchases cross set thresholds, and each tier unlocks higher rate limits³. The largest customers move to committed contracts with reserved capacity and negotiated prices, which is how Fireworks AI expands its biggest accounts⁹.

Big accounts end up in a sales negotiation

Self-serve tiers carry a customer a long way, but at some point a sales call happens. One reason is cost. A company spending heavily wants a discount, and OpenAI offers discounts for annual commitments on Scale Tier, its reserved-capacity plan¹³. The other reasons go beyond price. Scale Tier comes with a 99.9% uptime guarantee and prioritized compute, and only Enterprise customers can buy it through OpenAI's sales team¹³. Data terms matter too. OpenAI says it doesn't train on API data by default, but zero data retention is only for "qualifying organizations," data residency is only for eligible customers, and healthcare customers need a signed agreement to meet HIPAA rules¹⁴. So the move from pay as you go to an enterprise agreement is usually a negotiation over price, capacity, reliability and data, not just a bigger bill.

Who starts that conversation isn't public. My read is that it goes both ways, which is inferred, not documented. The provider sees every account's spend in real time, so it can spot accounts worth a call, and Anthropic hires enterprise sellers to prospect outbound and drive "early expansion across multiple use cases"¹⁵. Customers reach out when they need something self-serve doesn't offer, such as guaranteed capacity, zero data retention or a contract their security team will sign.

A few customers carry the revenue

Ara Kharazian, an economist at the corporate card company Ramp, found that the top 1% of customers generate 80% of enterprise revenue at both OpenAI and Anthropic, which he called "a level of concentration risk unseen in any other software category we track"¹. In 2025, the tech news site VentureBeat reported that two customers, Cursor and GitHub Copilot, drove roughly $1.2 billion of Anthropic's revenue². When your biggest customers are AI products themselves, their growth, their pricing changes and their choice of model all move your numbers.

Prices fall 10x a year, and customers shop around

Guido Appenzeller, an investor at Andreessen Horowitz (a16z), found that the cost of an LLM at a given level of performance falls by about 10x every year⁴. Providers also reward efficient use. The Claude API charges 10% of the base input price for cached prompt reads on most models⁵, and Together AI prices its batch API 50% below real-time inference⁸. Customers notice. In a 2025 a16z survey of 100 enterprise CIOs, 37% used five or more models, and one leader said that for most tasks "all the models perform well enough now," so pricing has become "a much more important factor." The same survey found switching is getting harder as companies build agent workflows tuned to one vendor's model⁶. Market share still moves fast. Menlo Ventures, an Anthropic investor, estimates that Anthropic went from 12% of enterprise LLM API spend in 2023 to 40% in 2025, while OpenAI fell from 50% to 27%⁷.

Startups win by serving or specializing

The big labs dominate frontier models, but startups pursue this business model in two ways. Inference clouds serve open models quickly and cheaply. Only 30–40% of Together AI's revenue comes from per-token API calls, with the rest from renting GPUs, and Sacra estimates its gross margin at about 45%⁸. Fireworks AI runs at an estimated 50% gross margin because GPU costs sit in cost of goods sold⁹. Specialized model companies sell one capability by API. Cartesia sells voice generation at one credit per character and has raised $191 million¹⁰. TypeSafe AI launched Jev, a model that returns structured decisions to code, at $0.042 per million input tokens, backed by a $40 million seed round¹¹.

How it flows

Developers and companies arrive through model launches, docs and benchmarks, developer communities, cloud marketplaces and startup credits. Anthropic offers startups up to $100,000 in Claude API credits¹².
They create an API key and make a first call. Many tinker and never ship.
They prototype and run evals against other providers' models. Some pick another model.
The winners ship to production, and usage now follows the customer's own users.
Paying customers move up spend tiers as usage grows³. The biggest negotiate enterprise agreements covering discounts, reserved capacity, uptime and data terms¹³^,¹⁴.
As their users grow, token volume grows. Most of this expansion needs no sales call.
Each new model release pulls in new developers and sends existing customers back to their evals, where they can be won or lost.
Customers who switch to a cheaper or better model can be won back at the next launch.

Metrics that matter

Metric	Definition	Why it matters here
Time to first call	Median time from creating an API key to the first successful request.	The activation moment. Developers who can't get a call working quickly move on to the next provider.
Production conversion	Share of new accounts running a sustained production workload within 90 days.	Prototypes use few tokens, often paid for with a small prepay or startup credits¹². Production is where revenue starts, because usage grows with the customer's own users.
Tokens per customer	Monthly tokens processed per active paying customer.	The core usage unit. Most growth here needs no sales call, until an account is big enough that discounts, capacity and data terms come into play¹³^,¹⁴.
Net revenue retention	This period's revenue from last period's customers, divided by what those customers paid last period.	One number that captures both expansion and customers switching to another provider.
Revenue concentration	Share of revenue from the top 1% (or top 10) of customers.	At OpenAI and Anthropic, the top 1% of customers generate 80% of enterprise revenue¹.
Effective price per million tokens	Revenue divided by tokens served, after caching, batch and contract discounts.	Prices for equivalent performance fall about 10x a year⁴, so revenue per token shrinks unless volume grows faster.
Gross margin per token	Revenue minus the GPU cost of serving it, as a share of revenue.	Inference clouds run at an estimated 45–50%, well below the 70%+ typical of subscription software⁸^,⁹.

Examples

OpenAI APISelf-serve spend tiers that raise rate limits automatically³. More than 9,000 organizations have processed over 10 billion tokens each¹.

Claude API (Anthropic)An estimated 40% of enterprise LLM API spend in 2025⁷, with heavy concentration in coding customers like Cursor and GitHub Copilot².

Together AIServes open models per token and rents GPUs. About $1 billion in annualized revenue by February 2026, per Sacra⁸.

Fireworks AIPer-token serverless inference for self-serve developers, dedicated GPUs and reserved capacity for large accounts. Customers include Cursor, Notion and Uber⁹.

Cartesia and JevStartups selling a single capability by API: voice for Cartesia¹⁰, structured decisions for code with TypeSafe AI's Jev¹¹.

Sources

The PPTX is built from editable shapes. It also opens in Google Slides (File › Import slides).