WHY THIS MATTERS
According to people familiar with the matter cited by Bloomberg on September 4, 2026, DeepSeek plans to install at least 160,000 Huawei Ascend 950DT accelerators at a new data center in Inner Mongolia, reserved for inference only. Neither DeepSeek nor Huawei has confirmed this publicly, but the report matches months of DeepSeek's failed attempts to train models on Huawei silicon and a $7 billion raise in June earmarked for its own infrastructure.In this article
RUMOR: the biggest known bet yet on Chinese silicon to run AI models
Editorial label: RUMOR. According to people familiar with the matter cited by Bloomberg on September 4, 2026, DeepSeek plans to install at least 160,000 Huawei Ascend 950DT chips at a new gigawatt scale data center in Ulanqab, Inner Mongolia, 350 kilometers from Beijing. If confirmed, the order would create one of the largest known clusters of Chinese accelerators ever assembled to run artificial intelligence. Neither DeepSeek nor Huawei has publicly commented on the report so far.
Why the key detail isn't the number, it's the function
The part that changes how this story reads isn't the volume of chips, it's what they would be used for. According to one of Bloomberg's sources, DeepSeek does not plan to use the 950DT chips for training its models, even though Huawei designed and marketed this chip specifically for that heavier task. The plan reportedly is to use Chinese hardware only for inference, meaning running already trained models to answer users, while training would keep relying on whatever Nvidia accelerators the company can still obtain. That split explains why China has been separating the two stages: training is like studying for an exam for years, inference is taking the exam itself, and Chinese hardware still hasn't convinced anyone it can handle the first part.
The track record that gives the rumor weight
This isn't a rumor floating in a vacuum. DeepSeek reportedly already tried training its previous model generation on Huawei silicon and backed off, returning to Nvidia cards for training, according to earlier reporting. In June 2026 the company raised more than $7 billion specifically to fund its own physical infrastructure, after years of relying on rented compute capacity from third parties. The chosen site also makes economic sense: Ulanqab has cheap wind and solar power and an average temperature of 4.3 degrees Celsius, cutting cooling costs for a site that, at full load, would consume power equivalent to 750,000 homes.
The bottleneck that could delay everything: Huawei itself
Huawei plans to commercially launch the Ascend 950DT only in the fourth quarter of 2026, with its own in house high bandwidth memory. Shortages of that component are expected to cap output at a few hundred thousand units this year, meaning that even if DeepSeek's order is real and finalized, fully supplying 160,000 chips could take well over a year. The data center itself is only expected to partially come online by late 2027 or early 2028, a horizon long enough that any projection today should be treated as intent, not installed capacity.
What would confirm or weaken the rumor
Confirmation would look like an official announcement from DeepSeek or Huawei about the order, a contract surfacing in regulatory filings, or additional reporting with named primary sources. The most realistic clock to watch is the Ascend 950DT's commercial launch in the fourth quarter of this year: any delay in Huawei's production proportionally delays any delivery to DeepSeek, making the installation timeline the first concrete signal of whether the plan is actually moving forward.
What changes in MaxAssistant's read
The rumor has a traceable, credible source, Bloomberg itself, but neither company has confirmed it. That's enough to treat it as a relevant signal that the training frontier remains tied to Nvidia chips even among the most advanced Chinese labs, while inference becomes the first territory where domestic Chinese silicon can prove value at scale. For anyone tracking AI infrastructure outside China, the practical takeaway is that splitting training from inference has stopped being just a technical nuance and become a declared survival strategy under sanctions.
Sources
Bloomberg, DeepSeek Plans Big Huawei AI Chip Order to Power New Data Center: https://www.bloomberg.com/news/articles/2026-09-04/deepseek-plans-big-huawei-ai-chip-order-to-power-new-data-center | tech-ish, DeepSeek turns to Huawei for 160,000 AI chips as Nvidia stays locked out of China: https://tech-ish.com/2026/09/04/deepseek-turns-to-huawei-for-160000-ai-chips-as-nvidia-stays-locked-out-of-china