DeepSeek weighs 160,000 Huawei Ascend chips for a 1GW inference site in Ulanqab

Divide roughly one gigawatt by at least 160,000 chips and the site allows about 6,250 watts per chip ^[13]. The gigawatt is what the facility is expected to draw ^[4]. Against the roughly 1,000 Ascend 910C chips DeepSeek used for full-parameter post-training of its V4 Pro model in June 2026 ^[7], the fleet under consideration is 160 times larger ^[14].

The chips at Ulanqab would be running inference ^[5]. DeepSeek still relies on Nvidia for large-scale frontier pre-training ^[10]. So the domestic stack, on this evidence, covers post-training and serving, and the layer where the biggest clusters sit has not moved.

US restrictions on exports of advanced AI chips to Chinese companies began in late 2022 and tightened in subsequent rounds ^[12]. They bind hardest on the largest training runs. The deployment described is in the serving layer, and serving is the part of the workload that scales with users rather than with model generations. The 910C run produced software as well as a result: DeepSeek has tuned optimisations such as CANN adaptations for the domestic hardware ^[11].

The delivery window Liang gave, around Q4 2026 or Q1 2027 ^[2], sits on top of Huawei's new date for its next-generation Ascend 960DT training chip, Q1 2027, pulled in by roughly three quarters on unusually high demand from Chinese AI developers ^[8]. Wide adoption of the 950DT among those same developers is expected to start in 2027 ^[9]. Three claims on the same fab output inside two quarters. Crypto Briefing, which reported the plan, read the early 960DT launch as either confidence in Huawei's manufacturing pipeline or competitive pressure too high to hold the original schedule ^[15].

The report describes a deployment DeepSeek is contemplating and does not say an order has been placed or at what price ^[16]. That leaves two ways this goes differently. If the 950DT parts arrive on time and in volume, Huawei has a gigawatt-scale reference customer and, in Crypto Briefing's reading, the guaranteed lab demand that justifies further domestic fab spending ^[17]. If deliveries slip or land at a fraction of 160,000, DeepSeek's Nvidia exposure extends through another model generation.

I would take the first as the base case, largely because the 1,000-chip run and the reported stable 950DT test results ^[6] both came before the larger commitment, and because Huawei pulled the date in ^[8]. The claim is narrow, though: the Hangzhou lab ^[1] is domesticating inference while keeping Nvidia at the frontier. It breaks if Huawei's 960DT shipments end up carrying DeepSeek's next pre-training run, because then the substitution is total and the export controls stop reaching the workload that matters most.