How Chinese AI companies make money: about a tenth of what OpenAI and Anthropic earn

How do Chinese AI companies make money, and how much? Epoch AI's October 6, 2026 report answers with a number: the six leading firms (Alibaba, ByteDance, Z.ai, Moonshot, DeepSeek and MiniMax) earn roughly 10% of what OpenAI and Anthropic earn together. The gap is in revenue, in margin and in pricing power, and it exists even though Chinese open-weight models trail American frontier models by only months on capability.

This is a money story for a race publication because revenue is what pays for the next training run. A lab that cannot fund compute cannot stay near the frontier for long.

The numbers in Epoch's table

Epoch gives annualized revenue figures. Treat every one as an estimate, because most of these companies do not report AI revenue separately.

• Anthropic: $65 billion
• OpenAI: $40 billion
• ByteDance (model-as-a-service): $4.0 billion
• Alibaba (model-as-a-service): $2.4 billion
• Z.ai: $1.8 billion
• Moonshot: $1.0 billion
• DeepSeek: $1.0 billion
• MiniMax: $0.8 billion

Add the six Chinese figures and you reach about $11 billion. The two American labs total about $105 billion. That is where the 10% comes from. Note that the ByteDance and Alibaba lines count only their model-selling businesses, so their wider AI-driven income is excluded by design.

Five ways the money is made

Epoch sorts the revenue into five channels, and each has a catch.

Consumer apps. They draw large audiences and thin profits. MiniMax's AI-native apps earned a 4.7% gross margin in the first nine months of 2025, while its enterprise services earned 69.4%. Epoch attributes the weak consumer result to low payment conversion, about 6% of users, and modest spending by those who pay.

Model access. Selling API access is the clearest parallel to the American labs, but weights get in the way. When Z.ai released the weights of GLM 5.3 Flash on OpenRouter, its market share fell from 88% to 22% within 20 days as competitors offered cheaper hosting. Z.ai's API margin was 24.6% in the first half of 2026, against the 85% margin Epoch reports for Anthropic.

Enterprise and government work. Customized deployments pay well but do not scale. Z.ai earned 73.7% of its 2025 revenue from on-premises deployment at a 48.8% gross margin, then moved toward standard cloud services; on-premises revenue fell to 13.5% by the first half of 2026 because maintenance costs limited the economies of scale.

Licensing. Moonshot released Kimi K3 under a modified MIT license that requires revenue sharing from providers earning more than $20 million a year. Epoch notes enforcement is unclear, given legal uncertainty about copyright in model weights.

Indirect income. For the conglomerates, the model sells other products. ByteDance's Volcano Engine handled 49.5% of China's public-cloud model tokens in 2025, and Epoch reports that AI-generated short dramas allegedly drove about $1.6 billion a year in advertising spend on Douyin.

The consumer side shows why scale alone does not fix the economics. MiniMax's apps reached large audiences, yet only about 6% of users paid, and those who paid spent modestly, which is how a business ends up with a gross margin under 5% on its consumer apps while its enterprise services earn nearly 70%. The enterprise side has the opposite problem: high margins that did not scale, as Z.ai's retreat from on-premises work shows.

Why open weights cost so much

Epoch's reading is that Chinese firms are younger, have less capable models and have lower pricing power, and that the open-weight strategy follows from those conditions as much as from any ideology. Open releases buy diffusion and goodwill. They also remove the thing a seller needs, which is a product no one else can host.

The price pressure is not unique to China. Separately, Epoch's September 22 analysis, "The plunging price of thought," found that the cost of reaching a given level of AI performance has fallen about 47% per quarter since 2023, or 13 times per year. A lab that releases weights hands that deflation to every competitor on day one. Mistral's new Large 4 will test the same trade in Europe once its weights ship.

Why the comparison needs care

Three cautions apply before you quote the 10% figure. First, the periods differ across the evidence: MiniMax's margins cover the first nine months of 2025, Z.ai's API margin covers the first half of 2026, and Volcano Engine's token share is a 2025 figure. Second, the American numbers are also estimates, and Epoch's 85% margin for Anthropic is a reported figure, not an audited one. Third, the Douyin advertising number is described by Epoch as allegedly driven by AI-generated short dramas, so it is an attribution, not a measurement. None of this changes the order of magnitude, which is the point of the report, but it does mean the exact ratio could move by a few points in either direction as better data arrives.

What this means for the race

The comparison with American labs supports a more careful claim than the usual one. Chinese labs are close on capability, as our coverage of GLM 5.3's cyber results showed, but they are far behind on the revenue that funds compute. Anthropic's estimated figure alone is nearly six times the six Chinese firms combined.

It also explains behavior that otherwise looks odd, such as heavy reliance on grey-market access to American models, which we reported in the Claude API transfer-station story.

The caveats are real. These are estimates from public filings, platform data and press reports. The report is an analysis, not a set of interviews, and we did not interview anyone. Revenue definitions differ across companies, and two of the six are conglomerates whose AI income is partly indirect.

Our position: stop treating benchmark parity as commercial parity. The Chinese labs have shown they can build near-frontier models. They have not yet shown they can sell them at American margins. If that holds, the open-weight strategy is as much a necessity as a choice, and the thing to watch is whether any of the six finds a way to charge for access once the weights are public.