China's AI Ecosystem Broadens

China’s AI industry is no longer moving on a single U.S.-built stack: DeepSeek’s next flagship model has been delayed as it is rewritten for Huawei chips, even as Chinese open-weight models spread faster worldwide. (usnews.com) (technologyreview.com) Reuters reported on April 3 that DeepSeek’s unreleased V4 model will run on Huawei’s latest Ascend chips and that DeepSeek spent the past few months working with Huawei and Cambricon to rewrite parts of the model’s code. The same report said Alibaba, ByteDance, and Tencent placed bulk orders for Huawei chips totaling hundreds of thousands of units ahead of V4’s launch. (usnews.com) That rewrite is about software plumbing as much as silicon. Nvidia’s CUDA is the dominant toolkit for training and running AI on Nvidia chips, while Huawei’s Ascend line uses its own software stack, so code tuned for one environment does not move over cleanly to the other. (usnews.com) At the same time, Chinese model makers are gaining reach through open-weight releases, which means publishing the trained numerical parameters so developers can download and run models themselves. MIT Technology Review reported in February that a recent Massachusetts Institute of Technology study found Chinese open-source models had passed U.S. models in total downloads, and that Alibaba’s Qwen family had overtaken Meta’s Llama in cumulative Hugging Face downloads. (technologyreview.com) Hugging Face said on January 20 that one year after DeepSeek released R1, China had built “a growing organic open source AI ecosystem,” with new model makers and open releases spreading across 2025. The post said DeepSeek’s R1 became the most liked model on Hugging Face and that the platform’s top-liked models were no longer majority U.S.-developed. (huggingface.co) The performance gap has narrowed at the same time the distribution gap has shifted. Stanford’s 2026 AI Index said U.S. and Chinese models have traded the lead multiple times since early 2025, and that as of March 2026 Anthropic’s top model led the top Chinese model by 2.7%. (hai.stanford.edu) That combination leaves developers with two separate decisions instead of one: which model family to use, and which hardware-and-software stack can run it efficiently. A model can be globally popular because its weights are open, while still being optimized for a regional toolchain that is harder to support outside China. (technologyreview.com) (usnews.com) Chinese firms are not all making the same bet, either. MIT Technology Review said DeepSeek’s January 2025 R1 release pushed more Chinese companies toward open-weight distribution, while Reuters said DeepSeek withheld early optimization access from U.S. chipmakers and gave domestic suppliers that window instead. (technologyreview.com) (usnews.com) The result is a broader Chinese AI market with more models, more chip paths, and fewer assumptions that Nvidia’s software will remain the default everywhere. DeepSeek’s delayed V4 release shows how costly that shift can be in the short term, even as Chinese models keep reaching more developers in the rest of the world. (usnews.com) (huggingface.co)