How do Chinese AI companies make money?
A · Strong evidence Epoch AI report by Anson Ho with GovAI and Yale economist Cheryl Wu, built from company filings, OpenRouter data and Chinese-language sources; strong evidence, though a topical analysis rather than an Epoch flagship.
Summary based on the full text·Original: about 25 minutes to read
In short
Cheryl Wu (GovAI, Yale) and Anson Ho (Epoch AI) analyse revenue sources for six Chinese AI firms: Alibaba, ByteDance, Z.ai, Moonshot, DeepSeek and MiniMax. Together they earn about a tenth of OpenAI and Anthropic's combined AI revenue. They identify five revenue streams and show that releasing open weights increases adoption but erodes pricing power, while conglomerates earn more indirectly through cloud and advertising.
Key numbers
- ~10%AI-related revenue of China's six leading AI firms relative to OpenAI and Anthropic combined, September 2026
- 88% to 22%Z.ai's share of GLM 5.3 Flash tokens on OpenRouter within 20 days of releasing the weights
- 4.7%gross margin of MiniMax's AI-native consumer apps, first nine months of 2025 (vs 69.4% for its enterprise platform)
- 73.7% to 13.5%share of Z.ai revenue from on-premise deployments, 2025 to H1 2026
As reported by the source.
English·Web page·opens epoch.ai
What it reports
- Five streams: consumer apps, model access (API and flat-rate plans), enterprise and government customisation, licensing fees, and indirect revenue from complementary products.
- MiniMax apps: 27.6 million MAU and 1.77 million paid users in the first nine months of 2025; ads in Talkie were 20.9% of company revenue.
- After Z.ai released GLM 5.3 Flash weights on 26 August, it faced 26 rival hosts by 15 September, 8 cheaper; its token share fell from 88% to 22% and daily volume dropped 70%.
- Z.ai API gross margin: 18.9% in 2025, 24.6% in H1 2026. DeepSeek V4 API margin is estimated at 70-80%, yet DeepSeek captures only about 8% of OpenRouter spending on its own models.
- Z.ai on-premise gross margin went from 66.0% (2024) to 48.8% (2025) to 37.9% (H1 2026).
- Moonshot's Kimi K3 licence requires a separate agreement for Model-as-a-Service businesses above $20 million revenue over 12 months; Moonshot reportedly seeks up to 30% of K3 revenue on major clouds.
- Alibaba AI cloud and compute revenue was $7.1 billion in Q2 2026, up 45% year over year; ByteDance's Volcano Engine handled 49.5% of China's public-cloud model token calls in 2025.
- About 128,000 micro-dramas launched in Q1 2026, over 95% AI-generated, versus about 33,000 in all of 2025; illustrative estimate of $1.6 billion a year in related Douyin ad spend.
How they got it
Annual and interim reports, IPO documents, OpenRouter token and pricing data sampled across eight days, press reports (The Information, Reuters, SemiAnalysis, Chinese outlets) and executive statements. Revenue run rates are from June to September 2026; translations are the authors'.
Limits
- Revenue figures mix run rates, annualised interim results and estimates from different months.
- OpenRouter is one channel and may not represent enterprise demand; licensing revenue is largely unknown.
Why it matters for companies
Open-weight Chinese models are cheap partly because their makers cannot hold prices once weights are out, and that pressure is likely to persist. Buyers can get these models from many hosts, but should check licence terms such as Kimi K3's revenue threshold. Vendors moving away from on-premise customisation signal that bespoke deployments are expensive for providers too.
Summary based on the full text. Read the original for the full argument and data.Read the original on epoch.ai