Open weights win users but cost Chinese AI labs their pricing power

Epoch AI took a close look at how six Chinese AI firms actually earn money, and the picture is less flattering than their benchmark scores suggest. As of September 2026, the six together bring in roughly 10% of the combined AI revenue of OpenAI and Anthropic. Epoch splits them into conglomerates with other businesses to lean on (Alibaba, ByteDance) and model-first startups (Z.ai, Moonshot, DeepSeek, MiniMax). ByteDance leads the group at about $4 billion in annualized model-as-a-service revenue, with DeepSeek and Moonshot near $1 billion each.

The most useful finding is about open weights. Releasing a model openly drives adoption but erodes pricing power, because third-party hosts then compete on price and capture much of the spending. DeepSeek, Epoch estimates, captures only about 8% of OpenRouter spending on its own models, even at a healthy 70 to 80% API gross margin. Z.ai's GLM 5.3 Flash saw its OpenRouter share fall from 88% to 22% within about 20 days of its weights going public. Consumer apps tell a similar story: MiniMax's Talkie ran a 4.7% gross margin over the first nine months of 2025, and only about 6% of its monthly users paid anything.

So the profitable paths are narrow. Enterprise and government deployments earn good margins but are hard to scale because each one needs heavy customization and maintenance, which is why Z.ai is retreating from on-premises work. The conglomerates are the comfortable exception: Alibaba's AI cloud revenue hit $7.1 billion in Q2 2026, up 45% year over year, because it can host rival open models next to its own Qwen and still sell the compute underneath. Epoch's full breakdown is careful to flag that many of these figures are estimates stitched together from different reporting periods.

Why it matters

If you depend on cheap open-weight Chinese models, this explains why the labs behind them struggle to profit and why those prices may not hold, so it is worth watching whether licensing terms tighten or releases slow.