Bloomberg reported on Thursday, citing unnamed sources, that DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT accelerators at its new data center near Ulanqab in Inner Mongolia. At roughly 111,000 yuan per chip, the order is worth about $2.5 billion – one of the largest ever placed for Chinese-made silicon. If the plans materialize, it would also be the biggest known AI cluster built outside the Nvidia ecosystem.
Inference on Huawei, training on Nvidia
The most important detail of the report is not the numbers but the division of labor. Those 160,000 chips are meant to handle inference only – serving models and answering user queries. Training future model generations stays on Nvidia hardware. That is a deliberate choice: a training attempt on Huawei silicon failed a year ago, and the 950DT is designed for decoding – it packs 144 GB of Huawei’s own HiZQ 2.0 memory at 4 TB/s plus a 2 TB/s interconnect, which in practice means fast generation of long answers. In other words, the most expensive, most prestigious part of the AI stack remains on the American side, while what API customers actually see – millions of daily queries – moves onto domestic chips.
A scale nobody has in the West
160,000 chips is a number worth comparing with the largest Western clusters. xAI’s Colossus supercomputer in Memphis started with 100,000 H100s and later grew to 150,000 H100s plus 50,000 H200s. Google’s TPU fabric (Virgo) links roughly 134,000 chips. DeepSeek’s order, if fulfilled in full, beats both. The whole Ulanqab facility is planned at gigawatt scale, and the 160,000 chips cover only part of its final capacity. The location is no accident: electricity in Inner Mongolia costs a fraction of the price in Beijing or Shanghai (Alibaba pays about 0.33 yuan per kWh there), and the cold climate allows free-air cooling for most of the year. Alibaba, Baidu and ByteDance – which is reportedly negotiating another 5-6 GW – already run facilities in the region.
The catch: Huawei cannot build enough
The most interesting part of the story, however, is on the supply side. The Ascend 950DT is only entering the market – the official launch of the 950 series was planned for Q4 2026 – and production capacity is constrained by component shortages, including advanced memory. Bloomberg estimates this year’s output at just 200,000-300,000 units, which Huawei must split among all its customers. Hence the reports that DeepSeek is lobbying Beijing to increase and accelerate its allocation, and hence the realistic scenario in which full delivery stretches over 12-18 months. This is also part of a bigger picture: Huawei has already published a SuperCluster configuration linking more than 520,000 950DT chips, so 160,000 is not an architecture limit – it is a factory limit.
What it means for the market
For Nvidia, this is another nail in the coffin in China. Jensen Huang has admitted that Nvidia’s share of China’s AI accelerator market has dropped to zero, and TrendForce data says domestic chips now control about 90% of the country’s high-end market, with Ascend alone expected to hold roughly 62%. For developers working with DeepSeek’s open weights, another thread matters more: the company itself says the 4-Pro API throughput is limited by the shortage of high-end cards, and once Ascend 950 supernodes enter mass production, Pro prices are expected to drop significantly. If Chinese APIs stop being the bottleneck, pricing pressure across the entire model market will intensify further. The Bloomberg report is unconfirmed – neither DeepSeek nor Huawei commented – but the mere fact that such a plan leaked says a lot about the direction of China’s AI infrastructure.
This content was generated using artificial intelligence and reviewed by a human editor.
This article is also available in Polish (PL).
Sources
- Bloomberg: DeepSeek Plans Big Huawei AI Chip Order to Power New Data Center (04.09.2026)
- XenoSpectrum: DeepSeek Plans 160,000 Huawei AI Chips for Inference, Still Trains on NVIDIA (05.09.2026)
- China Biz Insider: DeepSeek’s $2.56B Huawei Chip Order Exposes China’s AI Infrastructure Inflection Point
- TechTimes: DeepSeek’s 160,000-Chip Huawei Order (05.09.2026)
- Huawei Central: DeepSeek to order 160,000 Huawei AI chips over Nvidia
- TFTC: DeepSeek Orders 160,000 Huawei Chips, Proving U.S. Sanctions Backfired
Leave a Reply
You must be logged in to post a comment.