DeepSeek plans to install at least 160,000 Huawei Ascend 950DT accelerators at a data center it is building in Ulanqab, Inner Mongolia, a cluster that would rank among the largest known deployments of Chinese AI chips, Bloomberg reported.
The chips are earmarked for inference, running models to answer users, not for training. That split says a lot about where Chinese hardware stands in 2026. Huawei’s newest accelerators can carry a large share of deployment work, while DeepSeek’s training pipeline still leans on other equipment. The startup currently has no plans to use the 950DT for training, according to people familiar with the buildout.
Timing is the weak point. The installation schedule depends on Huawei’s production capacity, and the chipmaker has struggled to lift output while juggling competing customers. Deliveries could stretch more than a year. DeepSeek is also negotiating financing for the expansion, with talks reportedly in the billions of dollars, and the hardware mix could still change as supply catches up.
A gigawatt site with room to grow
The Ulanqab facility targets power consumption on the scale of one gigawatt, enough to supply roughly 750,000 households. The 160,000 Ascend chips would cover only part of that capacity, and DeepSeek has not said what fills the rest. At full build, the site would put DeepSeek’s infrastructure in the same neighborhood as Western hyperscale campuses, though populated with comparatively lower-end hardware.
Inner Mongolia has become a default location for these builds. Power is cheap, land is plentiful, and the cool climate trims cooling costs, which is why several of China’s largest computing projects cluster there. The tradeoff is distance: the region sits far from the coastal cities where most Chinese AI users live, so operators lean on high-capacity fiber to move results east.
Even unfinished, the project would set a marker for China’s domestic compute push. Roughly six months ago, a Shenzhen city newspaper reported the country’s first 10,000-chip computing cluster built on Ascend hardware. In June, Meituan said it trained a model with 50,000 domestically produced chips. In July, Z.AI completed a one-gigawatt data center it plans to fill exclusively with Chinese silicon, though only about 10,000 cards are installed so far.
| Project | Scale | Chip origin |
|---|---|---|
| DeepSeek, Ulanqab | 160,000+ planned, 1 GW site | Huawei Ascend 950DT |
| Meituan training run | 50,000 chips | Domestic |
| Z.AI data center | 1 GW site, ~10,000 installed | Chinese only |
| First Ascend cluster | 10,000 chips | Huawei Ascend |
One person familiar with Huawei’s manufacturing expectations told Bloomberg that several clusters of DeepSeek’s planned size could exist in China by the time the Ulanqab site switches on.
Nvidia still anchors the ecosystem
The pivot to Huawei does not mean Chinese labs have given up on Nvidia. Private companies in China still prefer the US chips, as seen in large orders from Alibaba and Tencent placed after Washington loosened export rules. Beijing has approved only a fraction of the up to 1 million H200 units the Trump administration was prepared to allow, and renting overseas compute remains a legal path to top-end accelerators.
The relationship is strained in other ways. US officials have accused DeepSeek and Moonshot of relying on smuggled Nvidia hardware, and administration officials said DeepSeek acquired Nvidia’s leading Blackwell chips and put them to work in Inner Mongolia. Bloomberg could not independently verify that claim. Beijing, for its part, is urging domestic labs to switch to local hardware, though enthusiasm varies across companies.
Why inference first
Huawei has positioned the 950DT as a training chip, so DeepSeek using it for inference is something of a detour. It suggests demand for domestic accelerators has outrun supply of the configurations labs actually want, and that running models locally is now urgent enough to take whatever capacity becomes available.
The inference load is also growing fastest. Every chatbot answer, every API call and every agent action consumes compute at serving time, and DeepSeek’s price cuts earlier this year pushed usage up sharply. Serving that demand on rented foreign hardware is neither cheap nor politically comfortable in Beijing’s current climate, which makes an owned domestic fleet worth the premium.
If the buildout lands as planned, the immediate effect is on inference economics inside China: more local capacity to serve users without touching restricted US gear. The longer-term question, whether Chinese accelerators can handle frontier training at scale, remains open, and nothing in the Ulanqab plan answers it.
discussion