DeepSeek is planning a cluster of more than 160,000 Huawei accelerators
The one-gigawatt site in Inner Mongolia would be among the largest known deployments of domestic Chinese AI silicon. It is the test China's whole compute strategy depends on.
DeepSeek is planning to deploy more than 160,000 Huawei AI accelerators at a one-gigawatt site in Inner Mongolia, according to Bloomberg. It would be among the largest known clusters of domestic Chinese AI silicon.
This is the experiment that determines whether US export controls have bought a delay or an advantage. Everything else in the Chinese compute story — the smuggling prosecutions, the subsidiary workarounds, the five-year plans — is provisional until somebody demonstrates that frontier-scale training runs on Huawei parts.
What has and has not been proven
Chinese labs have shipped genuinely competitive models. Alibaba refreshed Qwen3.8-Max at 2.4 trillion parameters with a one-million-token context window. Z.ai released GLM-5.3-Flash, 320 billion parameters with 18 billion active, at a tenth of its predecessor's cost. DeepSeek itself has repeatedly produced results that surprised people who assumed compute constraints would show.
What is not established is what hardware those runs used. A large amount of Chinese frontier work has been done on Nvidia parts acquired before restrictions tightened, or through routes the controls were meant to close — Taiwan indicted nine people over 74 smuggled B300 servers, and the New York Times reported Inspur shipping about $3 billion in Blackwell systems to Southeast Asia through its US subsidiary Aivres.
A 160,000-accelerator Huawei cluster is a different claim. It is a commitment to domestic silicon at a scale where the weaknesses — interconnect bandwidth, memory, software maturity, failure rates across a long run — cannot be hidden.
The software problem is the harder one
Raw accelerator performance is the part of this that gets discussed and the part that matters least.
CUDA's advantage is fifteen years of accumulated libraries, kernels, debugging tools, and the fact that essentially every training framework was developed against it. Huawei's CANN stack has to substitute for all of that, and at 160,000 devices the problems that matter are distributed training reliability, communication efficiency and the ability to diagnose a failure in a run that has been going for weeks.
Those problems are solved by operating at scale, which is the point of building the cluster.
The scale in context
One gigawatt puts this in the same class as the largest Western AI campuses. Five gigawatt-scale AI data centres are expected online this year, each run by a different hyperscaler. OpenAI has a partnership to deploy at least 10GW of Nvidia systems, and SB Energy granted it $5.5 billion in warrants tied to a planned 10GW Ohio campus.
China's Ministry of Industry and Information Technology has set a national target of 9,800 EFLOPS by 2030, backed by 3.8 trillion yuan. This cluster is the first serious instalment.
It also has a supply chain of its own to prove. South Korea's August chip exports were $46.65 billion, nearly triple a year earlier and 47.5 percent of the country's total exports — a measure of how much of the world's AI memory comes from two Korean companies. A Chinese programme built on domestic accelerators still needs high-bandwidth memory, and that is a chokepoint the controls also reach.
DeepSeek and Huawei have not confirmed the deployment, its timeline, or what the cluster will be used to train.
Runs the newsroom. Rename this profile in the studio to your own byline.
Related
Every weekday, the AI stories that moved money or shipped code.
No cross-posting, unsubscribe anytime. See all newsletters