How AI companies make money, and which model is failing

How do AI companies actually make money? Five business models are in play, and they have wildly different economics. Two of them work today, and the most visible one is under the heaviest pressure.

The five models AI companies make money with

Model	Revenue	Margin pressure
Sell compute	Hardware, cloud capacity	Low, supply is scarce
Sell tokens	API access per request	High and rising
Sell seats	Per-user subscriptions	Moderate, depends on conversion
Sell outcomes	Software that does a job	Low, if the job has value
Sell implementation	Consulting and integration	Low, labour-bound

Selling compute is the best business

Everyone else needs it, and nobody can substitute it. Nvidia demonstrated the strength of that position by arranging up to $500 billion of financing for its own customers, guaranteeing the collateral value of its chips.

Jensen Huang told CNBC his chips are an “investable asset”. Suppliers who can turn their product into collateral aren’t fighting on price.

Selling tokens is the worst

Charging per API request looks like a clean software business until you notice the cost of goods sold. Every request consumes compute you paid for, so gross margin depends on inference efficiency rather than on code.

An API business has a marginal cost. That single fact separates it from every software business people are pattern-matching it to.

Worse, prices keep falling. Grok 4.6 matched GPT-5.6 Sol Max at 61 on the Artificial Analysis index while charging $6 per million output tokens against $30, and published figures put input roughly 60% cheaper too.

That gap matters because buyers can act on it. Model choice is one of the few decisions in a software stack that can be changed with a configuration line, so any provider charging a premium has to justify it continuously rather than once.

Then there’s a free competitor. Meta’s 30B open-weight release under Apache 2.0 sets the ceiling at zero for anyone able to self-host, and CNBC read it as aimed squarely at the paid providers.

Selling seats depends entirely on conversion

Per-user subscriptions are the familiar software model, and they work when people actually pay. Distribution alone doesn’t get you there.

Microsoft has enormous distribution and still restructured, merging its Copilot apps and moving Deep Research behind a paywall. TechTimes read the consolidation as evidence of a paid-adoption problem.

That’s the seat model’s weakness. If the free tier is good enough, you’ve built a large user base that costs you inference money and pays you nothing.

Which is why free features keep getting withdrawn rather than improved. Removing a popular free capability is a blunt way to force conversion, and it works precisely because the alternative is subsidising non-payers indefinitely.

Selling outcomes is where the margin hides

The strongest position is charging for a job completed rather than for access to a model, because the price anchors to the value of the work instead of to the cost of tokens.

Databricks is closest to this shape among the big private names, raising $5 billion at a $190 billion valuation on run-rate revenue past $7 billion with growth above 80%.

It sells a platform that does data work, not access to a model, so falling token prices help its costs rather than hurting its pricing. That asymmetry is the whole argument for this model.

The catch is that outcome pricing demands you actually deliver the outcome. Selling access shifts the risk of a poor result onto the customer, while selling a job done keeps it, which is a harder product to build and a much better one to own.

Selling implementation is unglamorous and defensible

The fifth model barely gets discussed and may be the safest of all. Enterprises can’t deploy this technology alone, and the bottleneck is integration and compliance rather than capability.

IBM is betting on exactly that, folding OpenAI models into its consulting platform and building a practice of thousands of consultants, per its own announcement.

Consulting margins are capped by headcount, so it never becomes a software business. But it’s indifferent to which model wins, which is a valuable place to stand while the model layer commoditises.

Which AI companies make money today

Honestly, nobody outside the companies knows, and that should temper every confident claim including this one.

Most interesting AI companies are private and disclose selectively. Public ones bury AI inside large segments, so a revenue run-rate in a funding announcement is a company’s own unaudited number on its own timing.

What can be said is structural. Compute suppliers earn on scarcity, implementation earns on labour, and outcome-based products earn on value delivered. Token sellers are squeezed between falling prices and a free alternative.

Beware the revenue figures that circulate, too. Annualised run-rate takes a recent month and multiplies it, which flatters any fast-growing business and says nothing about retention or whether the revenue recurs at all.

The counter-case for token businesses

The pessimism above deserves a challenge, and there’s a decent one.

Falling prices expand the market. Workloads uneconomic at $30 per million output tokens become viable at $6, so volume can grow faster than price declines. Commodity businesses at enormous scale are still good businesses.

Efficiency also cuts costs continuously. Epoch AI reports pre-training compute efficiency improving roughly 3x per year, and serving efficiency improves alongside it, so margins can hold even as list prices drop.

And switching is less frictionless than it looks. Prompts, evaluations and tuning are model-specific, which creates real stickiness once a product is in production.

What would settle it

Gross margin disclosure from any pure token seller would answer the central question immediately, and the absence of it is itself informative.

Watch whether outcome-priced products spread beyond data platforms into ordinary software, since that would confirm the pricing model rather than the company.

And watch whether open weights keep closing the capability gap. If free models stay one generation behind indefinitely, paid providers hold a premium. If they catch up, the token business becomes a utility, and utilities are priced accordingly.