Goldman Sachs’ Ronald Keung team released a roughly 50-page in-depth report on Chinese AI models in early July.

The report says Chinese models bring token costs down to 1/4 to 1/8 of the US SOTA by using MoE sparse architectures, and they first lose money to seize the market.
Chinese models have already captured overwhelming shares in two major token-heavy scenarios: agent and programming, and they have also driven the whole industry’s mixed token prices to peak and then fall.
Specific details are as follows:
1. The mixed token prices of China’s top programming models (GLM5.2, Qwen3.7 Max) are about $1 per million tokens, while the US SOTA models’ prices for outputs in the same tier are $4–$8.
2. Sold at a loss. GS estimates that the agentic models focused on value for money currently have an EBIT profit margin of -30%, and the programming models are -39%. They are able to absorb the losses by hard-carrying through a cash-rich balance sheet; based on GS’s model, by 2030 they will turn positive to +14% and +22%, respectively.
3. The reason they can sell so cheaply is the architecture—across the full line, the proportion of activated parameters per token is under 8%. DeepSeek V4 Pro activates only 49B out of 1.6T total parameters, GLM5.2 activates only 40B out of 744B, with fewer FLOPs, so prices have a structural lower bound.
4. Adoption is already reflected on OpenRouter: by task, Chinese models account for 5–16% of spend, but they account for 85% of agent tokens and 89% of code tokens. In any scenario where “single-task cost” becomes a key metric due to runtime and scale, they are winning.
5. Result: mixed token prices also peak and then fall. SDLLMTK peaked around 2.07 in early June, and is now 1.67.