DeepSeek Orders 160,000 Huawei Chips for Inference Cluster
Based on reporting by the-decoder.com. Analysis and framing are Seedwire's own.
DeepSeek is planning to deploy at least 160,000 of Huawei's next-generation Ascend-950DT chips in a data center in Inner Mongolia, according to Bloomberg reporting first covered by the-decoder.com. If completed, it would be the largest known Huawei chip cluster built to date.
The chips are earmarked for inference only, not training. DeepSeek will keep using Nvidia hardware for the heavier training workloads that build its models in the first place. Huawei likely can't fulfill the full order for over a year, the report says, because of production limits and shortages of memory chips needed to build the processors. Nvidia offers additional context on this topic.
The memory bottleneck
That memory shortage is not fully closed but it is starting to crack. CXMT, China's leading memory chipmaker, has begun producing small batches of HBM3E, the high-speed memory type that powers many AI processors. That's a real milestone for a domestic supplier. But CXMT remains three to five years behind Samsung, SK Hynix, and Micron, which are already mass-producing the newer HBM4 standard. DeepSeek's order lands inside a broader Chinese government push to build out a domestic chip industry without falling behind in the AI race. DeepSeek offers additional context on this topic.
The split between inference and training in this deal is the detail worth sitting with. It tells you where Huawei's Ascend chips are considered ready today, and where they are not. Inference, running an already-trained model to answer queries, is a more forgiving workload than training, which needs enormous, sustained, tightly synchronized compute across a cluster for weeks or months. By keeping Nvidia for training and routing Huawei chips toward inference, DeepSeek is effectively grading Huawei's hardware as good enough for the easier half of the job, not yet the harder half.
That's still meaningful progress for China's chip independence effort. Inference is where the ongoing, recurring compute demand sits once a model ships and users start hitting it. A company that can run its production inference load on domestic silicon reduces its exposure to US export controls on Nvidia chips, even if it still needs Nvidia to build the next model. Given how central Nvidia access has become to Chinese AI policy debates, any credible reduction in that dependency at scale is the kind of proof point Beijing wants to point to.
The timeline caveat matters just as much as the ambition. A cluster Huawei can't fully deliver for over a year, gated by memory chip shortages that a domestic supplier is only now starting to address in small batches, is a plan, not a deployment. CXMT's HBM3E milestone shows the supply chain moving, but a three-to-five-year gap behind Samsung, SK Hynix, and Micron on the newer HBM4 standard means China's AI chip buildout will likely keep depending on components its rivals shipped years earlier. Whether DeepSeek's 160,000-chip target actually materializes on schedule will say a lot about how fast that gap can close, and how much slack Huawei has in its own production beyond the memory bottleneck.
Worth watching is whether other Chinese AI labs follow DeepSeek's lead with similar inference-on-Huawei, training-on-Nvidia splits. If that becomes the standard pattern across the industry, it suggests a durable division of labor is forming rather than a one-off hedge, and it would mark a slower but steadier path toward chip self-sufficiency than a full swap to domestic hardware.