Chinese AI models: compute deflation catches up with OpenAI and Anthropic prices in the US

After the price hikes from OpenAI and Anthropic, several American companies are adopting DeepSeek and Qwen: Chinese compute deflation becomes an adjustment variable for US AI budgets.

The fact

According to CNBC (07/07/2026), several American companies are adopting Chinese models - notably DeepSeek and Z.ai - for internal workloads, as the costs of accessing OpenAI and Anthropic increase. The substitution remains focused on non-critical use cases (summarization, internal code) and via on-premise or third-party cloud inferences (fil cloud-gpu-specialized-contracts).

Our analysis

The dynamic confirms the compute deflation scenario: the value shifts from training (dominated by US hyperscalers) to inference (democratized by Chinese open-weight models). The real risk for Anthropic and OpenAI is not sovereignty (data remains US) but price elasticity: each price increase expands the addressable market for open-weight models. In the short term, this also fuels demand for captive infrastructure (cf. TeraWulf lease 19 B$, pub #931): US labs seek to lower their unit cost through direct leases rather than API prices.

To watch

DeepSeek V4 publication mid-July 2026. OpenAI/Anthropic pricing response. Inference usage statistics by platform (Hugging Face, Together AI). CFIUS/Department of Commerce positions on Chinese open-weight models used in US production.