The Physical Stack of AI · Sovereign AI compute
You can explain why China's sovereign-compute strategy is fundamentally different from every other country's, what its official domestic-stack evidence does and does not prove, and how to avoid overreading model-release claims.
Every other country in this unit takes the global chip supply chain as fixed and tries to negotiate good access to it. China has been forced — by Washington, mostly — to try to build the entire stack itself, from lithography upward. That is a different shape of problem with a different shape of cost. The US Commerce Department's October 2022 BIS rule was the trigger; BIS's 2025 GP10 guidance kept PRC advanced-computing ICs, including Huawei Ascend, legally sensitive for US exporters and US persons.
By mid-2026 the official-source picture is genuinely mixed. Huawei Cloud markets Ascend AI Cloud Service for large-model training, inference and migration. Huawei's 2025 keynote describes Atlas 900 A3 SuperPoD and CloudMatrix384, a cluster-scale domestic infrastructure story. Z.ai and other Chinese labs publish internationally visible model services. Those sources establish a serious domestic-stack push. They do not, by themselves, prove that any particular model was pretrained end to end on domestic chips with no Nvidia participation.
That distinction is the lesson. In a chip-controls debate, "Huawei sells Ascend cloud," "Huawei has a CloudMatrix cluster story," "Z.ai ships a capable model," and "this exact model was fully pretrained on Ascend" are different claims with different evidence requirements. A good reader keeps them apart.
Chapter contains 3 lessons.