DeepSeek intends to install at least 160,000 of Huawei’s Ascend 950DT accelerators at a hyperscale data center it is building in Inner Mongolia, according to Bloomberg. It would be among the largest known clusters of Chinese-made AI chips, and it is meant to serve models rather than train them.
That last part is the odd detail, and it fits a pattern. Huawei designed and marketed the 950DT for training, the more demanding job, and DeepSeek is reported to have no current plans to use it that way. What it wants the chips for is inference, running models it has already built. Chinese accelerators have been good enough at inference for a while and have struggled with training, which is why it counted as a milestone in June when a team including Huawei completed full-parameter post-training of DeepSeek’s 1.6-trillion-parameter V4-Pro on a cluster of at least a thousand Ascend 910C chips.
The scale is unusual by the standards of domestic Chinese silicon. China’s first publicly reported ten-thousand-chip Huawei cluster came online roughly six months ago. The Inner Mongolia site is being built at gigawatt scale, drawing at full utilization about what 750,000 households use, and the 160,000 chips would occupy only part of it.
Supply is the constraint rather than demand. DeepSeek would take more chips than Huawei can currently produce, and filling the order could take more than a year while Huawei balances competing customers. Ascend production capacity, not design, is what limits how fast Chinese labs can substitute away from Nvidia.
DeepSeek is one of several Chinese labs building on domestic accelerators. Z.AI brought up a one-gigawatt data center running entirely on Chinese chips in July. Huawei has been pushing the same hardware abroad, offering Ascend and Atlas systems in South Korea since the summer.
An unresolved question hangs over the same site. Trump administration officials have said DeepSeek acquired Nvidia’s top-end Blackwell chips and installed them in Inner Mongolia. Bloomberg reports it has not independently verified that claim, and it is not clear what mix of hardware the company is running there now. Nvidia’s authorized channel into China opened only in July, when the first H200s reached Chinese buyers under newly approved export licenses.
The commercial pressure behind the buildout is visible in DeepSeek’s pricing. The company began charging by the clock in August, halving inference rates during off-peak hours, which is what a lab does when serving capacity is the binding cost and it wants to flatten demand across the day. A 160,000-chip inference cluster is the other way to solve that problem.
Sources: The News, South China Morning Post
–
By the Control Plane Editorial Team