DeepSeek is moving training and inference workloads from Nvidia GPUs to Huawei's Ascend AI chips, be...
The AMW Read
DeepSeek testing whether Ascend silicon can sustain frontier training, not just inference, meaningfully updates the export-control-bottleneck debate with cross-segment compute and geopolitical implications.
DeepSeek is moving training and inference workloads from Nvidia GPUs to Huawei's Ascend AI chips, betting on domestic silicon to keep scaling despite US export controls.
According to The Information, DeepSeek is increasingly running both training and inference on Huawei Ascend hardware rather than restricted Nvidia GPUs. The shift extends beyond inference — where Chinese labs have already leaned on domestic chips — into training, the workload where Nvidia's CUDA stack and interconnect performance have been hardest to replace.
The move tests a live question for the whole China AI segment: whether Ascend-class silicon can sustain frontier-scale training runs, not just serve pre-trained models. If DeepSeek succeeds, it weakens the assumption that US export controls durably cap Chinese frontier-model progress by starving labs of high-end GPUs, and reinforces the sovereign-silicon substitution track domestic vendors have been racing to prove out. Per the AI Market Watch index, DeepSeek has generated 105 tracked news items in the past 90 days versus 94 in the prior period (name-matched, pipeline-ingested sources only) — sustained volume reflecting how closely the market is watching every infrastructure and product move from the company.
For builders and investors, the timing matters: DeepSeek has been hiring roughly 150 senior backend engineers to overhaul its infrastructure as it moves from research lab to service platform, and is reportedly preparing a STAR Market IPO partly to fund compute. A durable move to Ascend would reshape that cost base and supply chain outside Nvidia's ecosystem, but it also raises reproducibility questions for benchmark claims — echoing recent independent findings that some of DeepSeek's reported V4 Pro gains traced to post-training rather than architecture. Enterprises evaluating DeepSeek models should treat chip provenance as a due-diligence line item alongside pricing and benchmark claims.
