DeepSeek’s V4 Model Signals China’s AI Chip Independence DeepSeek, the Chinese AI startup known for its cost-efficient models, has unveiled a preview of its V4 model, marking a strategic shift toward domestic semiconductor infrastructure. According to Reuters, the V4 model runs on Huawei’s Ascend 950PR chips, making it the first frontier-class AI model trained entirely on Chinese-designed hardware. This development underscores China’s accelerating effort to reduce reliance on U.S. Semiconductor technology amid ongoing export controls. The collaboration between DeepSeek and Huawei represents a significant departure from the startup’s prior dependence on Nvidia’s GPUs. Even as DeepSeek previously gained global attention for its V3 model trained on Nvidia hardware, the V4 launch highlights a deliberate pivot to strengthen China’s AI supply chain. Lushbinary notes that DeepSeek provided Huawei with early optimization access to the V4 model while withholding similar access from Nvidia and AMD, signaling a clear alignment with domestic chipmakers. Despite the hardware shift, the V4 model maintains broad compatibility. Developers can still run the open weights on Nvidia GPUs, and the application programming interface (API) functions identically regardless of the underlying hardware. This approach allows for flexibility in deployment while advancing China’s goal of building a self-sufficient AI ecosystem. Industry reports indicate that the Ascend 950PR, Huawei’s latest AI accelerator, is designed to compete with Nvidia’s H100 and H200 chips. Though full specifications remain unpublished, the chip is believed to feature Huawei’s Da Vinci architecture, HBM3 memory, and support for the CANN (Compute Architecture for Neural Networks) software framework. DeepSeek’s achievement in training a 1.6-trillion-parameter model on Ascend chips—achieving competitive results on benchmarks like LiveCodeBench (93.5) and Codeforces (3206)—demonstrates that Chinese AI accelerators are now capable of frontier-level model training. The V4 launch has already triggered significant demand within China’s tech sector. The Information reported that major firms including Alibaba, ByteDance, and Tencent have placed bulk orders for Huawei’s upcoming chips, totaling hundreds of thousands of units. This level of procurement reflects growing confidence in domestic alternatives and suggests a broader industry shift toward homegrown semiconductor stacks. As the global AI hardware landscape fragments into competing ecosystems, DeepSeek’s V4 model serves as both a technical milestone and a geopolitical indicator. It demonstrates that high-performance AI training is no longer exclusively dependent on U.S.-made chips, reinforcing China’s push for technological sovereignty in artificial intelligence.
19