Glonce

Deepseek plans 160,000-chip Huawei cluster in Inner Mongolia

infrastructureThe Glonce editorsThe Decoder🔥68Glonce rating68Impact68Novelty58Relevance87Surprise47Players72Credibility77
Deepseek plans 160,000-chip Huawei cluster in Inner Mongolia

Deepseek plans to deploy at least 160,000 of Huawei's next-generation Ascend-950DT chips in a data center in Inner Mongolia, according to Bloomberg. The chips would handle only inference, not training; for the much heavier training workloads, Deepseek still relies on Nvidia hardware. If it goes ahead, the buildout would be the largest known Huawei chip cluster and a concrete step toward reducing China's dependence on Nvidia for AI compute. Huawei probably cannot deliver the full order for more than a year, held back by production limits and shortages of the memory chips that AI processors need. That memory bottleneck could ease over time: China's leading memory maker, CXMT, has started producing small batches of HBM3E, the high-speed memory used in many AI chips, for the first time. CXMT still lags Samsung, SK Hynix, and Micron by three to five years, since all three rivals are already mass-producing the newer HBM4 standard. Deepseek's order fits into a broader Chinese government drive to build up a domestic chip industry while keeping pace in AI.

Key facts

  • Deepseek plans to deploy at least 160,000 of Huawei's Ascend-950DT chips in Inner Mongolia, per Bloomberg.
  • The cluster would run inference only; Deepseek still trains its models on Nvidia hardware.
  • Huawei probably cannot deliver the full order for over a year because of production limits and memory chip shortages.
  • China's top memory maker, CXMT, has begun small-batch HBM3E production but remains three to five years behind Samsung, SK Hynix, and Micron, which are already mass-producing HBM4.
  • The order is part of a broader Chinese government push to build a domestic chip industry without falling behind in AI.

Why it matters

A 160,000-chip Huawei cluster would be the largest known deployment of Chinese-made AI accelerators, a concrete marker of how far China's chip industry has come since export controls cut off easy access to Nvidia's top hardware. Deepseek pairing Huawei silicon for inference with Nvidia hardware for training shows the substitution is happening piece by piece rather than all at once, and only where the technology is currently ready.

Who it affects

The plan concerns Deepseek, which would run its inference workloads on the new cluster; Huawei, whose Ascend-950DT chips and production capacity are the bottleneck; and CXMT, whose progress on HBM3E memory determines how fast Huawei can actually build systems like this one. More broadly it affects the Chinese government's chip self-sufficiency push and, by extension, Nvidia's position in the Chinese AI hardware market.

How to use it

There is no product or service here to adopt; this is a hardware deployment plan, not a released tool. Companies watching China's AI supply chain can treat it as a data point on how quickly domestic accelerators are being adopted for production inference workloads, and on how memory supply, not chip design, is currently the binding constraint.

How solid is it

The report comes from Bloomberg via The Decoder, and the piece describes a plan rather than a completed or confirmed deployment. No timeline for construction, no cost figure, no specific facility name, and no named executives are given in the source; the one concrete schedule detail is Huawei's own delivery estimate of over a year for the chip order.

Risks and caveats

The deployment is a plan, not a finished fact, and Huawei's own estimate puts full delivery more than a year out, subject to production limits and memory shortages. CXMT's HBM3E output is described as small batches, well short of the mass production Samsung, SK Hynix, and Micron already have with HBM4, so the memory bottleneck the article flags may persist regardless of how the chip cluster itself progresses.