Ecosystem vs. Silicon
- Nvidia's CEO warned that if China's DeepSeek shifts from Nvidia's stack to Huawei's, it would be 'a horrible outcome'. - Reports say DeepSeek V4 can run on Huawei's Ascend 950PR chip and use the CANN software stack. - That substitution risk could reshape procurement, developer habits, and long‑term vendor lock‑in ( ).
Nvidia chief executive Jensen Huang said this month that a DeepSeek shift from Nvidia’s tools to Huawei’s would be “a horrible outcome” for the United States. (dwarkesh.com, scmp.com) The warning followed a Reuters report on April 3 that DeepSeek’s next flagship model, V4, is being built to run on Huawei chips, not Nvidia hardware. Reuters said Alibaba, ByteDance, and Tencent had placed orders totaling hundreds of thousands of Huawei chips ahead of the launch. (usnews.com) DeepSeek has also spent months working with Huawei and Cambricon to rewrite parts of the model’s underlying code, according to that Reuters report. Reuters said DeepSeek gave early access to domestic chip suppliers instead of U.S. chipmakers before the V4 release. (usnews.com) AI chips are only part of this fight; the software layer decides how easily developers can build, tune, and deploy models. Nvidia’s software stack is centered on CUDA, while Huawei’s competing stack is called Compute Architecture for Neural Networks, or CANN. (dwarkesh.com, trendforce.com) That is why Huang framed the issue around standards as much as silicon. On the April 15 Dwarkesh Podcast, he argued that keeping Chinese artificial intelligence developers on the American stack serves U.S. interests better than pushing them onto domestic alternatives. (dwarkesh.com, thenextweb.com) Huawei’s side of the story is not just one chip launch. TrendForce, citing EE Times China and other reports, said Huawei launched the Ascend 950PR in the first quarter of 2026 and is planning roughly 600,000 Ascend 910C chips this year, with total Ascend capacity targeted at 1.6 million units. (trendforce.com) DeepSeek’s V4 work comes after U.S. export controls narrowed Nvidia’s China business and pushed Chinese labs toward local substitutes. Reuters reported that DeepSeek’s low-cost V3 and R1 models had already rattled global tech stocks last year by raising new questions about how much compute top-tier models really need. (usnews.com) Neither DeepSeek nor Huawei immediately responded to Reuters requests for comment on the April 3 report. Nvidia’s argument is now plain: if Chinese labs can train and serve major models on Huawei chips with Huawei software, the contest shifts from who sells the fastest processor to who owns the developer ecosystem. (usnews.com, dwarkesh.com)
