China accelerates shift to homegrown AI chips under U.S. pressure

China's biggest tech firms are working quickly to replace U.S. silicon, but the immediate benefits are in running AI, not training it. Export curbs on advanced Nvidia parts have pushed local designers to build "good-enough" chips that keep services online while longer investments in manufacturing mature.

Alibaba is the leading example. Previously a major Nvidia customer, it's now testing a new in-house processor aimed at a broad set of inference tasks, not just narrow, single-purpose jobs. The chip is fabricated at a Chinese foundry rather than TSMC, a shift forced by U.S. restrictions. To smooth adoption, Alibaba kept compatibility with the Nvidia software ecosystem, so teams can reuse existing code.

Rivals are approaching from different perspectives. Shanghai startup MetaX rolled out a GPU with more memory than Nvidia's H20, the most advanced Nvidia model Washington briefly let back into China before Beijing told buyers to hold off, trading higher power draw for capacity on certain workloads. MetaX plans to scale using older process technology and a multi-die approach to bypass capacity limits at domestic fabs. Meanwhile, Cambricon booked about $247 million in quarterly revenue on strong orders for its Siyuan 590 chip and warned investors after a sharp share surge; its market value still sits well above earlier levels.

State support is extending the timeline. Beijing launched an $8.4 billion fund to cut foreign dependence, and Huawei showcased a system that stitches together 384 Ascend chips. Some evaluations say it can top leading U.S. gear on select metrics, but the energy cost is steep. Even so, major public-cloud providers have been cautious about bulk Ascend purchases, in part because they see Huawei as a direct cloud competitor.

Tooling and supply chains remain the bottlenecks. Many engineers still prefer Nvidia's mature software stack; domestic chips can be tougher to integrate, and reports of overheating or system failures during long training runs persist. Chinese fabs, constrained by limited access to cutting-edge equipment, struggle to add the capacity designers want, pushing some vendors to combine smaller dies or lean on older nodes. Alibaba's new chip helps on the compatibility front, but it doesn't solve the training bottleneck.

This divide, smooth inference versus difficult training, defines the current gap. U.S. controls block the most advanced training processors, and Alibaba's latest part focuses on serving pre-trained models rather than creating them. Until local hardware can reliably handle long, hot training cycles at scale, China's major advances will skew toward keeping AI services responsive instead of building bigger foundational models.

Nevertheless, progress continues. DeepSeek hinted that software workarounds plus improving domestic silicon could move training forward, and some investors argue a full "made-in-China" AI stack could reach scale sooner than expected, pressuring Nvidia at home and abroad.