Epoch AI: China's six leading AI firms earn about a tenth of OpenAI and Anthropic, and open weights send the money to compute owners

Why it matters

Key points

Epoch AI published a report in the early hours of October 7 Beijing time (October 6 US time) titled "How do Chinese AI companies make money?" Authors Cheryl Wu and Anson Ho break down the revenue of six firms: Alibaba, ByteDance, Z.ai, Moonshot, DeepSeek and MiniMax. The headline finding: as of September 2026, the six together earn about 10% of what OpenAI and Anthropic combined earn in AI-related revenue.

The report identifies five revenue streams: consumer apps, selling model access via API, customized deployments for enterprises and governments, licensing fees, and indirect revenue from complementary businesses such as compute. A few numbers stand out. MiniMax's AI-native apps had a 4.7% gross margin in the first nine months of 2025, versus 69.4% for its open platform and enterprise services; of 27.6 million average monthly users, about 1.77 million paid, roughly 6%. After Z.ai released GLM 5.3 Flash weights on August 26, 26 third-party hosts appeared within 20 days, 8 of them cheaper than Z.ai, and Z.ai's share of the model's OpenRouter tokens fell from 88% to 22%. Z.ai's API gross margin was 18.9% in 2025 and 24.6% in the first half of 2026; for comparison, the report puts Anthropic's Opus 4.8 API-serving margin at list price around 85%.

The outlier is DeepSeek, whose V4 API gross margin is estimated at 70% to 80%, partly because it keeps more than 70% of its compute for training and uses higher prices to hold down inference demand. Yet only about 8% of OpenRouter spending on DeepSeek models goes to DeepSeek itself.

Verifiable

Report: https://epoch.ai/publications/how-do-chinese-ai-companies-make-money , dated October 6, 2026 on the page. Epoch AI's announcement post: https://x.com/EpochAIResearch/status/2107521419903222055 .

Other key figures in the report: Z.ai's share of revenue from on-premise deployments fell from 73.7% in 2025 to 13.5% in the first half of 2026, with the company citing rigid delivery and maintenance costs; Moonshot released Kimi K3 weights under a modified MIT license that requires a separate agreement from model-as-a-service operators with more than $20 million in revenue over 12 months, and the report cites Reuters saying Moonshot is seeking up to a 30% revenue share from major cloud platforms; Alibaba's AI cloud and compute revenue was $7.1 billion in Q2 2026, up 45% year over year, and ByteDance raised its 2026 model-as-a-service target to $4 billion.

A note on methodology: the figures are annualized run rates from different months between June and September, and Alibaba and ByteDance are counted only for model-as-a-service revenue, not cloud overall, so the 10% figure is conservative for those two.

Our take

What this report really shows is not that Chinese models are weaker, but that open weights separate building a good model from making money on it. Once weights are out, the model becomes something anyone can serve, the price is set by the cheapest host, and the money flows to whoever owns compute and distribution. The biggest commercial winners of Chinese open-weight models are therefore compute sellers such as Alibaba Cloud and Volcano Engine, plus the overseas clouds that host these models, not the startups that train them. Alibaba CEO Eddie Wu's public line that a thriving open-source ecosystem drives cloud demand is exactly this logic.

That will change what open source looks like. Kimi K3's revenue-threshold license and Z.ai's revenue-sharing deal with Amazon Bedrock keep the technology open while putting a toll gate on commercial use. We expect more flagship models from Chinese labs to ship with this kind of pay-if-you-are-big license, or with delayed weight releases, rather than full open release on day one.

For AI progress itself, a tenfold revenue gap eventually shows up as a compute gap. DeepSeek choosing to price away inference demand so it can keep compute for training shows that, under compute constraints, the business model is built around the research schedule. Whether Chinese labs keep pace on the next generation depends heavily on whether licensing and revenue-share income becomes real. The moment to watch is when upcoming financial reports and listing documents put concrete numbers on that income for the first time.