DeepSeek’s new AI model signals China’s shift away from Nvidia chips

DeepSeek’s latest model showed how far China’s AI industry has pushed toward hardware independence. The startup previewed V4 on April 24, adapting it for Huawei chip technology and signaling a break from the Nvidia dependency that had powered its earlier rise. Huawei said its chips were used in part of V4’s training process, a step Beijing has cast as a milestone in building a homegrown AI stack that can survive escalating U.S. export controls.

The technical message was as important as the political one. DeepSeek said V4 Pro outperformed other open-source models on world-knowledge benchmarks, falling behind only Google’s closed-source Gemini-Pro-3.1. Lewis Tunstall of Hugging Face said the model quickly rose to the top spot on the platform and handled very long, complex text tasks well, though it still lacked multimodal support for images and video. That is the core trade-off in China’s self-sufficiency drive: progress is real, but it is not yet full parity with the most capable systems.

The hardware story moved just as quickly. Huawei’s Ascend chips, which He Hui of Omdia called China’s best homegrown alternative to Nvidia, drew fresh demand after the DeepSeek launch. ByteDance, Tencent and Alibaba reached out to Huawei about new chip orders, and cloud-computing and GPU-rental firms also scrambled to secure supply. Huawei said its Ascend supernode infrastructure built on the Ascend 950 series would fully support DeepSeek V4 for inference, with the full Ascend SuperNode product line adapted for V4 inference as well.

That is where self-sufficiency becomes practical rather than rhetorical. Reuters reported that the Ascend 950PR outperformed Nvidia’s H20, but still trailed Nvidia’s H200, which had not yet shipped to China because Beijing and Washington were still at odds over sale conditions. In other words, China’s domestic alternative is advancing, but the ceiling remains visible. Performance gaps, supply constraints and regulatory limbo still shape what Chinese firms can actually deploy at scale.

The stakes extend beyond a single model or chip line. The DeepSeek-Huawei pairing landed ahead of President Donald Trump’s planned visit to Beijing for a summit with Xi Jinping, in a tense period when the White House accused China of industrial-scale theft of U.S. AI intellectual property. For Beijing, the message is clear: the goal is no longer simply to replace Nvidia in theory, but to prove that China’s best AI models can run on domestic silicon in enough volume to matter.