Beijing Pushes Chinese AI Firms Off Nvidia and Onto Huawei

The clearer story than any single directive is the trend line. Nvidia's share of China's advanced AI chip market has cratered from roughly 66% in 2024 to about 8% this year, according to Bernstein Research cited by CNBC, while Huawei has climbed to around 50%. Jensen Huang himself told analysts Nvidia has 'largely conceded' China's advanced AI chip market to Huawei, which is a striking admission from a company that had built something close to a monopoly there.

The state pressure behind that shift is now explicit rather than nudged. Beijing has added Huawei and Cambricon accelerators to an official government procurement list, with written directives to public agencies and state-owned enterprises to prioritize local suppliers, and new state-funded data centers must source at least 80% of core technologies from domestic vendors. The Wall Street Journal frames the drumbeat aimed at Chinese AI companies to migrate off Nvidia as the sharpest political signal yet in that self-sufficiency push, and the recent visit by Vice Premier Ding Xuexiang to Huawei's chip lab in Shanghai, broadcast on prime-time state news, was a very public endorsement of the same message.

For the labs actually training frontier models the tension is real. The reason Nvidia hardware was hard to give up is the software stack and the raw performance ceiling for large-scale training. Huawei is trying to close that at scale: Fortune reported Huawei plans roughly 600,000 flagship Ascend 910C units in 2026, with overall Ascend output stretching toward 1.6 million dies, and the 950, 960 and 970 rolling out through 2028. Whether that supply, and the tooling around it, actually matches what Alibaba, ByteDance or DeepSeek need for frontier runs is the practical question that determines how quickly the mandate can be complied with, not just announced.

Take the specifics as reported, not settled. What the coverage doesn't give you is a clean readout on whether frontier training gets quiet exceptions, how existing Nvidia inventory inside Chinese data centers gets treated, or how uniformly the procurement rules are enforced across private labs versus state buyers. The forward-looking point is that the base case for Nvidia's Chinese accelerator business is now zero rather than depressed, and the beneficiaries of that reallocation, Huawei, Cambricon and the state-anchored domestic accelerator startups, are where the actual money and the actual policy weight are pointing.