U.s.: China AI Watch: The Stack Is Splitting

“America is still defining the frontier. China is trying to industrialize it faster.”

Alert: material inflection. The evidence over the past two weeks materially strengthens our core thesis: the United States still controls the AI capability frontier, but China is increasingly challenging the economics, infrastructure stack and speed of diffusion of that frontier. The competition is shifting from who has the best model to who can turn intelligence into the cheapest, most scalable and most industrially embedded factor of production.

The most consequential development is the DeepSeek–Huawei convergence announced September 30. DeepSeek is open-sourcing compute and communication libraries for Huawei Ascend processors and extending TileLang to Huawei's Ascend 950 architecture, including systems built around 128-chip supernodes. This matters because the real strategic moat around Nvidia has never been silicon alone—it has been CUDA and the surrounding developer ecosystem. China is now explicitly attacking that software layer.

1. 01 China is building a vertically integrated alternative AI stack

   DeepSeek models + Huawei Ascend silicon + CANN + TileLang + open-source libraries increasingly constitute an indigenous model-to-chip pathway. Huawei also says Ascend is now supported as a PyTorch accelerator backend and is opening access to 10,000-NPU-scale compute resources. This does not mean China has displaced Nvidia. It means U.S. export controls are increasingly producing a second-order effect: accelerating the construction of a parallel Chinese compute ecosystem. signal: Export controls are catalyzing a parallel Chinese compute ecosystem.

2. 02 The U.S. is responding to Chinese cost pressure by collapsing inference prices

   OpenAI's September 22 GPT-6 Sol and Luna releases cut API prices by roughly 50% versus their GPT-5.6 equivalents. GPT-6 Sol is now $2/$10 per million input/output tokens and Luna just $0.10/$0.50, while GPT-6 Astra remains OpenAI's highest-capability model. The U.S. therefore retains the frontier while increasingly matching China on commoditized intelligence economics. signal: America keeps the frontier while matching China on commoditized economics.

3. 03 China's competitive advantage is moving from models to deployment density

   Huawei says its Industry AI Foundry now supports more than 1,000 deployed projects, its Agentic Infra serves over 3,500 customers, and its newest AI Cluster Service delivers 20% more token throughput than its prior generation. These are company-reported figures, but the direction is important: China is attempting to industrialize AI through standardized infrastructure, agents and sector-specific platforms rather than waiting for model leadership. signal: China industrializes through standardized infrastructure, not model leadership.

4. 04 Industrial AI is becoming China's transmission mechanism

   Official Chinese reporting says more than 30% of large manufacturing enterprises were using AI by mid-2026, while China had passed 700 million generative-AI users by June. Huawei separately says Chinese inference consumption has reached roughly 500 trillion tokens per day. These figures point to something larger than consumer adoption: an economy increasingly organized around cheap, abundant inference. signal: An economy increasingly organized around cheap, abundant inference.

5. 05 Washington increasingly recognizes that the competition has moved beyond chips

   The Trump administration has created a new high-level AI task force under intelligence chief Jay Clayton, while Treasury Secretary Scott Bessent has proposed an emergency U.S.–China AI notification mechanism following discussions with Chinese officials. The strategic debate is therefore shifting simultaneously toward frontier capability, infrastructure security and the systemic consequences of increasingly powerful open models. signal: The strategic debate has moved beyond chips to systems.

Key Takeaways

the asymmetric contest

• China is moving aggressively toward Phase IV industrial diffusion.
• The United States remains ahead in Phase I's most valuable technological frontier.
• America is still defining the frontier; China is trying to industrialize it faster.
• The Great Split is no longer between American and Chinese models — but between two production systems for intelligence.
• The U.S. is optimizing for maximum capability; China for maximum diffusion.