DeepSeek Orders 160,000 Huawei Ascend 950DT Chips, Moving Inference Under Chinese Jurisdiction

DeepSeek is betting its commercial inference business on domestic Chinese silicon at an unprecedented scale. According to a Tech Times report on September 5, which recaps a Bloomberg story from September 4, the Hangzhou AI lab has ordered 160,000 Huawei Ascend 950DT chips for a massive data center under construction in Ulanqab, Inner Mongolia. At roughly 111,000 yuan per chip, the order carries a face value of about 17.76 billion yuan (approximately $2.64 billion), and if fully delivered it would create one of the largest clusters of domestically manufactured Chinese AI chips in history.

The 950DT is engineered specifically for AI inference — the revenue-generating stage where a trained model answers user queries — and DeepSeek plans to deploy the chips to run its models in production at the Ulanqab facility. According to people familiar with the matter cited by Bloomberg, the company does not intend to use the 950DT for training new models, a task that continues to depend on Nvidia hardware. The choice of the 950DT over the already-available 950PR also signals an optimization target: the 950PR is tuned for the prefill phase of inference, while the 950DT targets high-throughput token generation during decode.

Inference and training place very different demands on hardware. Training requires sustained parallel floating-point throughput, while production-scale LLM inference is memory-bandwidth-limited: each token demands loading model weights from memory before any computation. The 950DT addresses exactly that bottleneck, integrating Huawei's proprietary HiZQ 2.0 high-bandwidth memory with 144 GB at 4.0 TB/s. Huawei's Atlas 950 SuperPod system can pack 8,192 of these chips behind its UnifiedBus 2.0 interconnect, delivering what Huawei claims is 8 exaflops of peak compute — though the report notes every published specification comes from Huawei itself, with no independent benchmark from a Western auditor.

Scale will take time. Bloomberg's reporting indicates DeepSeek targets partial operation of at least part of the Ulanqab facility by late 2027 or early 2028. The site, roughly 350 kilometers northwest of Beijing, was chosen for its climate — an average annual temperature near 4°C cuts cooling needs — and for low power costs that could run a 1-gigawatt facility well below coastal-city rates. Supply is the other constraint: advanced HBM stacks have challenging yields, sources estimate 2026 production of the 950DT at only hundreds of thousands of units with Huawei serving other customers too, and DeepSeek has reportedly asked Beijing to pressure Huawei into prioritizing its allocation.

The order has a software dimension as well. Running DeepSeek's models on Ascend took months of co-engineering, producing the V4 model — the first major DeepSeek release built specifically for Ascend, running on Huawei's CANN framework, the company's answer to Nvidia's CUDA. SemiAnalysis confirmed the V4-Ascend work was a ground-up co-design rather than an after-the-fact port. DeepSeek founder Liang Wenfeng offered the most candid benchmark in a leaked July investor-call transcript: everything a GB300 can do, the Huawei supernode can do with the same latency, at the cost of roughly four Huawei GPUs equaling one Nvidia GPU, about two years behind.

The deal sits inside a rapidly shifting market picture. Bernstein Research forecasts Huawei capturing roughly 50 percent of China's AI chip market by the end of 2026, with Nvidia's share sliding from about 40 percent in 2025 to roughly 8 percent; Nvidia CEO Jensen Huang told CNBC in May that the company has "largely conceded the China market to Huawei." The report argues US export controls have produced a bounded delay while accelerating China's drive for chip self-sufficiency, leaving two parallel hardware and software ecosystems that increasingly cannot run each other's code.

The compliance angle is what makes this more than a hardware story. Once online, every inference query routed through the Ulanqab facility — every prompt from a developer calling DeepSeek's API, every enterprise workflow processing customer data — will be handled by infrastructure operating under Chinese jurisdiction. The report cites Article 7 of China's National Intelligence Law, which obligates all organizations and citizens to support and cooperate with national intelligence efforts, alongside data-localization and government-access provisions in the Data Security Law and Cybersecurity Law. For enterprises routing sensitive customer data through DeepSeek's API, the governing legal framework includes those statutes, and the report argues such obligations cannot be negotiated away by contract; it also notes DeepSeek's privacy policy acknowledges storing user data on servers in China, plus existing government-device bans in Texas, New York, and Virginia and earlier GDPR action in Italy and a ban in the Czech Republic.

The decisive question is whether DeepSeek — or any Chinese frontier lab — will eventually train its next-generation models on domestic silicon, which would mark true decoupling. As of this reporting, that has not happened: the Ulanqab cluster is inference-only, and training continues on Nvidia hardware stockpiled before export rules tightened. Reuters reported in July that DeepSeek is developing its own proprietary inference chip, which could eventually reduce its dependence on both vendors. Whether the 160,000-chip order is fulfilled on schedule, and whether Beijing treats it as a strategic priority, will be the practical test of the argument that cluster scale can close the gap with Nvidia.