A gigawatt bet on domestic silicon
DeepSeek plans to build a large intelligent-compute centre in Ulanqab, Inner Mongolia, deploying at least 160,000 of Huawei’s next-generation Ascend 950DT chips. The order is worth about USD 2.56 billion. If completed it would rank among the largest domestic AI chip clusters ever assembled.
The centre is planned at gigawatt scale. The 160,000 chips cover only part of that capacity, and because Huawei’s supply is constrained, fulfilment will take at least eighteen months. For context, one renminbi converts to roughly 0.14 US dollars, so the cluster alone represents the better part of RMB 18 billion in committed spend.
Training on Nvidia, inference on Huawei
DeepSeek still trains its large models mostly on Nvidia GPUs. The Huawei Ascend chips are earmarked for inference. The two companies are extending the partnership into system design: they are jointly defining an Ascend super-node aimed at lowering latency and raising throughput on long-context inference while controlling cost, with a path to scale toward 10,000-card clusters.
The arrangement is less about a one-off purchase and more about co-developing the compute architecture. DeepSeek brings the model and the workload. Huawei brings the silicon and the system integration. The explicit goal is to make long-context inference cheap enough to run at scale inside China.
Why a Chinese AI lab is building its own foundry
The move is the clearest sign yet that US export controls have pushed China’s leading labs to lock in domestic compute at scale. For Beijing, a gigawatt data centre in Inner Mongolia is not just an AI project. It is real estate with a power allocation, and power has become the new bottleneck on the mainland’s AI build-out.
The strategic read is straightforward. The more of DeepSeek’s inference fleet that runs on Ascend, the less exposed the lab is to a single foreign supplier, and the more credible China’s claim to a sovereign AI stack becomes. That is why the centre is sized in gigawatts rather than in chips alone.
Editor’s note: This is an adapted translation of the original Leiphone report. It has been trimmed and restructured for readability for an international business audience.