How AI Companies Make Money: The Business of AI Explained

AI companies make money in seven ways: selling chips and hardware, renting computing capacity, charging per token or per action, selling subscriptions and seats, charging per completed outcome, selling advertising and commerce, and licensing data. Most of the profit so far has gone to the layer that owns the scarcest input, which in 2023–26 meant the chip makers. Most of the companies closer to the user, including the best-known model labs, still spend more than they earn.

The clearest way to see this is one quarter's results. On 28 October 2026 Microsoft reported that Azure had grown 48% and that its contracted-but-unbilled commercial revenue stood at about $677 billion, much of it from two model labs: OpenAI's ~$250 billion commitment and Anthropic's ~$30 billion (company and reported). Two days earlier Anthropic had filed to go public, showing about $16.3 billion of revenue in the first half of 2026 alongside roughly $265.6 billion of fixed or minimum compute commitments (from coverage of the filing). The labs' promises to pay are the clouds' backlog, and the clouds' spending is Nvidia's revenue. This guide follows that chain from the top to the bottom and explains who keeps the money.

Why is AI not an ordinary software business?

Software became the best business of the late twentieth century because the thousandth customer cost almost nothing to serve. Gross margins of 75–85% were normal. Generative AI looks like software and sells through the same channels, but its economics differ in four ways, which The Intelligence Equation sets out in its first chapter.

Every answer costs compute. A language model produces its output one token (about three-quarters of a word) at a time, and every token is arithmetic on an accelerator someone paid for and is paying to power. Reasoning models and agents can multiply the tokens per task by ten to a hundred times. A heavy user can cost a hundred times as much as a median one.

The core asset wastes fast. A frontier model costs billions to train, is paid for before any revenue arrives, and typically leads its nearest rival for a few months. The best open-weight model catches up within perhaps six to twelve.

A few suppliers sit under everyone. Almost every frontier model of 2023–26 ran on Nvidia chips or custom chips made, like Nvidia's, by TSMC in Taiwan, packaged with high-bandwidth memory from three suppliers and housed in data centers built by a handful of clouds.

Quality is hard to observe. Models give fluent answers whether or not they are right, and benchmark scores are mostly run by the sellers. A New York lawyer was sanctioned in 2023 for filing fake cases that ChatGPT had invented.

None of this makes AI a bad business. It means an AI company's results depend less on how many customers it signs and more on what each customer costs to serve, who it depends on, and how long its lead lasts.

What are the layers of the AI industry?

"AI company" covers businesses as different as a chip designer and a legal-research app. The useful picture is a stack, in which each layer sells to the one above and depends on the one below.

Around the stack sit three things it cannot do without: power, capital and rules. Each layer makes money differently, so the rest of this guide takes them in turn.

How do AI chip makers make money?

Chip makers sell hardware up front, and in this cycle they have earned the most. Nvidia's revenue rose from about $27.0 billion in fiscal 2023 to about $215.9 billion in fiscal 2026, and in the quarter to July 2026 it reported about $96.2 billion at a 75.0% gross margin (company). That is a software-like margin on hardware.

The reason is that Nvidia does not really sell a chip. It sells a platform: the GPU, the memory packaged beside it, the NVLink connections that join 72 GPUs into one rack-scale system, the networking between racks, and the CUDA software that most AI code is written against. CUDA was launched in 2006 and was free; by the time demand exploded, it had about fifteen years of libraries and trained engineers behind it. An H100 was reported at about $25,000–40,000 per GPU in 2023 (Nvidia publishes no list prices). We explain the full story in why Nvidia dominates AI chips.

Nvidia is not the only supplier earning excess margins. TSMC fabricates nearly every leading accelerator and reported a 66.4% gross margin in the third quarter of 2026. SK Hynix, the first memory maker qualified with Nvidia for high-bandwidth memory, reported preliminary third-quarter 2026 figures that imply an operating margin near 80%, for a business that lost money in 2023 (company). Server assemblers, by contrast, earn gross margins of about 5–12%, because the GPUs inside are pass-through.

The pattern across the chip chain is that value follows scarcity, and scarcity moves: advanced packaging in 2023, memory in 2024–25, power and data-center shells in 2025–26. The semiconductors and AI chips study tracks each bottleneck.

How do cloud providers make money from AI?

Clouds rent compute by the hour, by the token or under multi-year reservations. AI has turned the cloud from a software business with buildings attached into a capital business with software attached. The four largest US platform companies spent about $230 billion of capital in 2024 and about $400 billion in 2025, and guided to more than $700 billion for 2026.

The revenue has started to come back. AWS reported about $46.0 billion in the third quarter of 2026, up 39%, at an operating margin of about 38.7%; Google Cloud's operating margin was about 36.2% (company). But spending has run ahead of cash. Amazon's trailing free cash flow was about −$42.4 billion at September 2026, and Alphabet's quarterly free cash flow turned negative on about $51.4 billion of capex (company and reported).

Below the hyperscalers sit the neoclouds, companies such as CoreWeave, Nebius and Lambda that rent the newest Nvidia clusters to labs on contracts of two to five years. CoreWeave reported about $5.13 billion of revenue in 2025 with adjusted EBITDA of about 60%, but a GAAP net loss of about $1.17 billion after depreciation and interest, and about $35.6 billion of debt by 30 June 2026 (company). Our explainer on the CoreWeave business model walks through why a neocloud is closer to a leasing company than a software firm.

How do AI labs like OpenAI and Anthropic make money?

Model labs earn from four sources: API usage billed per million tokens, consumer subscriptions, enterprise contracts and, more recently, advertising and commerce. The two independent labs at the frontier chose different mixes.

OpenAI runs a consumer business. Roughly 5–6% of ChatGPT's weekly users paid anything on the figures reported in February 2026, and every free user is a compute bill. Its answer has been tiers from ChatGPT Go up to Pro at $200 a month, plus advertising, which it claims reached an annualized run-rate of about $1 billion by 31 August 2026. It also reported an operating loss of about $5 billion on 2024's revenue (reported), and its compute commitments, put by Sam Altman at about $1.4 trillion over about eight years in November 2025, were reported in February 2026 to have been reset to about $600 billion by 2030.

Anthropic chose developers and enterprises. Most of its revenue came from companies calling Claude through the API, many of them coding tools, and then from its own coding agent, Claude Code, which reached a company-claimed run-rate above $2.5 billion by February 2026. Coverage of its October 2026 filing showed an operating loss of about $693 million for the first half of 2026 but operating income of about $974 million in the second quarter alone.

Lab income statements need care. Training compute usually sits in research and development, not cost of revenue, so a lab can report a healthy gross margin on serving while spending more on training than it earns. OpenAI's adjusted gross margin was reported at about 33% in 2025; a simple calculation from Anthropic's filing coverage gives about 85% for the first half of 2026, though definitions differ. The question that matters is whether each model earns more in its short commercial life than it cost to train, serve and replace. The foundation model labs study sets out that arithmetic.

How do open-source AI models make money?

Nobody sells open weights directly. The weights are the free sample, and the money comes from around them. Meta gives Llama away to commoditize a layer where rivals charge, protecting its advertising business. Alibaba gives away Qwen so developers rent its cloud. Nvidia releases models because more capable open models mean more inference on more hardware. Hosts such as Together AI sell inference on identical weights, and their prices converge on cost.

The result is a ceiling on what closed labs can charge for anything the open frontier has reached. DeepSeek's R1 reasoning model launched in January 2025 at $0.55 per million input tokens and $2.19 per million output, against $15 and $60 for OpenAI's o1 at list. Openness also shifts with position: Meta launched its new flagship, Muse Spark, with closed weights on 8 April 2026, and DeepSeek raised its API prices for the first time in August 2026 after taking outside money. Our comparison of open vs closed AI models covers the trade-offs for builders and buyers.

How do AI applications make money?

Applications sell subscriptions, seats, usage or outcomes to the people who do the work. Their largest variable cost is the model call, which is why their margins are lower than traditional software. Bessemer reported in August 2025 that fast-growing AI applications averaged gross margins of about 25%, often negative, against about 60% for steadier ones.

Cursor is the clearest case. Its annualized revenue went from about $100 million in January 2025 to about $2 billion in February 2026 (company and press-reported run-rates), yet its gross margin was reported at about −23% in early 2025 (The Information, reported). A flat $20 plan bought access to models whose cost rose with use. Cursor moved heavy users to usage billing in June 2025, trained its own models, and reached "slight" gross-margin profitability by April 2026 (TechCrunch, reported). SpaceX then completed a purchase of the company valued at about $60 billion in August 2026.

Applications also face their suppliers as competitors. OpenAI, Anthropic and Google all launched coding agents in 2025 that compete with their own API customers. Whether an app has a defensible position or is a thin layer on someone else's model is the subject of our guide to AI wrapper startup moats.

How do consumer AI products make money?

Consumer AI has the fastest adoption of any product category on record and some of the hardest economics. Google's Gemini app reported about 950 million monthly users in July 2026 (company claim), and ChatGPT claimed about 900 million weekly users in early 2026. Unlike search or social media, every use costs real money, so the free tier has only three ways to be paid for: subscriptions from a minority, advertising to the rest, or a larger company that funds it as part of something else.

A stylized calculation in the book's first chapter shows the stakes. If a free assistant has 100 million weekly users sending 20 messages a week, at one cent per message it needs about 4.5% of users on a $20 plan to cover the serving bill; at five cents it needs more than a fifth. Engagement can make things worse: a companion app whose users chat for hours a day can lose more as engagement rises. Character.AI, one of the most-used AI apps by time spent in 2024, filed for Chapter 11 in October 2026.

What are the main AI business models?

Almost every AI business runs on one or more of seven revenue engines. Each allocates the risk of compute costs differently, which is the core of pricing in AI.

The choice of unit matters more than the price level. ChatGPT Pro launched at $200 a month in December 2024, and within weeks Altman said OpenAI was losing money on it, because the customers who choose a $200 plan are the ones who use it most. Salesforce, whose revenue came from seats, moved its Agentforce agents from $2 per conversation to $0.10 per action in May 2025, because automation that replaces people also shrinks seat counts. Our practitioner guide on how to price an AI product works through the trade-offs.

How much does AI cost to run?

The price of a fixed level of capability has fallen faster than almost any price in economic history. GPT-4 cost $30 per million input tokens and $60 per million output at launch in March 2023; GPT-4o mini, which scored higher on the MMLU benchmark, cost $0.15 and $0.60 in July 2024 (company price lists). Epoch AI found in March 2025 that the price of reaching a fixed performance level fell by a median of about 50 times a year.

Yet bills often rise. Reasoning models write hidden chains of thought billed as output, agents re-send their growing context at every step, and users move to the newest, dearest model. In a stylized example in the book, the same document-review task became about 170 times cheaper in a year on a fixed model, then about 30 times more expensive than the 2023 baseline when handed to a 40-step reasoning agent. Anthropic added weekly limits to Claude Code in August 2025 because some subscribers ran it "continuously in the background, 24/7" (company statement, reported). Our explainer on AI inference costs covers the mechanics and the margins.

Where does the money in AI actually end up?

A stylized calculation in Chapter 3 of the book follows one dollar of a law firm's AI subscription down the stack. The proportions are round assumptions, not any real company's figures.

Three lessons follow. Gross profit is not value captured: the app must pay its sales force and engineers from its half, and the lab must pay for the next model from its 20 cents. Power is the smallest bill, just over one cent per dollar, yet the hardest constraint, because nothing runs without it. And one layer's margin is the next layer's cost: if the app switches to a reasoning model that uses twice the tokens at the same price, its gross profit falls from 50 cents to 10 while the lab's and cloud's revenue double.

Who is paying for the gap between AI spending and AI revenue?

The chain would be simple if every dollar spent at the bottom had been earned at the top. It has not been. In 2026 the four largest US platforms guided to more than $700 billion of capex, while the revenue of all model labs combined was a small fraction of that. Four sources fill the gap.

The first is the cash flow of the platforms: Office, search advertising, retail and social advertising. By late 2026 that no longer covered spending in every quarter, and Alphabet, Meta, Amazon and Oracle issued large bonds in 2025–26.

The second is venture and growth capital. AI start-ups took about $212 billion of the roughly $425 billion of global venture funding in 2025 (Crunchbase, reported). OpenAI raised $122 billion at $852 billion post-money on 31 March 2026, and Anthropic $65 billion at $965 billion in May 2026 (reported).

The third is supplier capital. Nvidia agreed an equity investment of about $30 billion in OpenAI's 2026 round, AMD gave OpenAI warrants for up to 160 million shares alongside a 6-gigawatt supply deal, and Amazon agreed to invest up to $50 billion in OpenAI while selling it AWS capacity (reported). When a supplier funds its customer and the customer spends the money on the supplier, some revenue at the bottom of the chain is funded by the bottom of the chain.

The fourth is debt, secured on chips, data centers and lab contracts. CoreWeave's creditors are, in the end, relying on the labs in its backlog. The AI investing, valuations and capital cycle study tracks how these flows connect.

Why do so many AI startups fail?

Most failures in this cycle came from treating one strength as proof of another. Inflection AI raised about $1.5 billion and built a model it said came close to GPT-4, but it had no revenue and no distribution, and in March 2024 Microsoft hired most of its staff and paid about $650 million for a licence and a waiver (reported). Humane spent about $230 million on the Ai Pin and shipped about 10,000 units by August 2024 against a goal of 100,000 (reported). Builder.ai was valued at about $1.5 billion before a review found that 2024 revenue presented as about $220 million was nearer $55 million (reported); it entered insolvency in May 2025.

Each case is told in full elsewhere on this blog: what happened to Inflection AI, why the Humane Ai Pin failed, and our list of the biggest AI startup failures and what each teaches.

How do you judge whether an AI company will make money?

The book's answer is that an AI company creates durable value only when four things are true at once, and the factors multiply rather than add. If any is zero, the result is zero.

1. Capability that survives real work. Does the system do a job reliably enough, on evidence a buyer can check, that someone changes how they work and pays for it?
2. Unit economics that hold at real usage. Does each token, query, seat or resolved task earn more than the compute it consumes when real customers, not pilots, use it?
3. A position the next layer cannot route around. Does the company control something scarce, such as a chip ecosystem, capacity, a frontier model, distribution, proprietary data or a workflow?
4. The capital and permission to last the cycle. Can it fund compute, chips and power, and keep its legal and political licence to operate, long enough to get paid?

Two forces sit around the equation like weather: the capital cycle and the rules, from copyright rulings to export controls. Neither decides whether a product works, but both decide how long a company can survive if it does not.

The four factors map neatly onto the stack. Nvidia's story turned on the third factor, a position nobody could route around. The neoclouds and the labs turn on the fourth, capital and staying power. Applications and consumer products turn on the second, because their heaviest users are their most expensive. And the failures of earlier AI booms, from 1960s machine translation to IBM's Watson Health, turned on the first. Readers who want definitions for the terms used here can start with the AI glossary.

A practical checklist follows from the factors. Find the model bill and ask whether it rises or falls as customers use more. Check whether the price unit tracks the cost unit. Ask who could route around the company if its supplier shipped the same product tomorrow. Read run-rates and adjusted margins carefully, since private companies annualize single months and definitions differ. And read compute commitments as a balance sheet: non-cancellable capacity is debt in all but name.

The takeaway

AI companies make money by selling hardware, capacity, tokens, seats, outcomes, attention or data, but so far most of the profit has settled with whoever owns the scarcest input, which in 2023–26 was the chip layer. The layers closer to the user create great value and often keep little of it, because every answer costs compute and their suppliers can become competitors. The gap between spending and revenue is being filled by platform cash flow, venture money, supplier financing and debt, all betting that the top of the stack will one day pay for the bottom. The best way to read any AI business is to test all four factors at once, because a brilliant product with one factor at zero still yields zero. Chapter 3 of the book, Following the Money, traces that chain in full, and The Intelligence Equation applies it to every layer.