3 Sources
[1]
DeepSeek plans large Huawei chip order for new Inner Mongolia data centre
The reported one-gigawatt site at Ulanqab would run on Huawei Ascend silicon rather than Nvidia, extending a shift Jensen Huang has called a "horrible outcome" for America DeepSeek is preparing a large order of Huawei AI chips for a new data centre, according to Bloomberg. Neither the size of the
[2]
DeepSeek plans big Huawei AI chip order to power new data centre
The startup doesn't currently plan to use the 950DT chips for training, one of the people said, even though Huawei designed and marketed these processors to serve that more demanding function. While DeepSeek tried to train its previous models on Huawei chips, it has so far relied on Nvidia
[3]
NVIDIA Is Reportedly Giving Up On China As A Lost Cause Amid An Unassailable Lead From Huawei, As DeepSeek Snubs H20 GPUs For 160,000 Units Of Huawei's Ascend 950DT Chips
It was only in July that DeepSeek's CEO created seismic waves by boldly declaring that NVIDIA was digging its own grave in China. And now, merely weeks later, DeepSeek is apparently helping that prognostication come to fruition by placing a massive order for Huawei's Ascend 950DT AI chips, further
Share
Copy Link
DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT chips at its new 1-gigawatt data centre in Ulanqab, Inner Mongolia. The $2.56 billion order marks a significant shift toward domestic AI hardware as US export restrictions limit access to Nvidia's accelerators, advancing China's AI infrastructure independence.
DeepSeek is preparing to deploy at least 160,000 Huawei Ascend 950DT chips at a new data centre in Ulanqab, Inner Mongolia, according to Bloomberg
1
2
. The order, valued at approximately $2.56 billion based on current street prices of around $16,000 per chip, represents one of the largest known clusters of Huawei AI chips and signals a decisive move toward domestic AI hardware3
. The AI pioneer intends to use these Huawei Ascend 950DT chips primarily for inference workloads rather than training, despite the processors being designed for both functions2
.The reported one-gigawatt site at Ulanqab could begin partial operations in late 2027 or early 2028
1
. Ulanqab is part of China's Eastern Data, Western Compute programme, which moves energy-intensive computing infrastructure toward western regions where electricity is cheaper1
. Inner Mongolia has signed around $27 billion in AI computing projects as part of this strategic push1
. The region offers significant cost advantages, with power costs ranging between 0.32 and 0.35 yuan per kilowatt-hour, equating to just $0.05 per kWh3
.Huawei faces significant production constraints that could delay fulfillment of DeepSeek's order by more than a year
2
. Shortages of components such as top-end memory will cap Huawei's output of the 950DT at the low hundreds of thousands this year2
. Huawei is balancing demand from other clients and attempting to export small volumes overseas, further complicating the delivery schedule2
. Bloomberg News reported that Huawei plans to make around 1.6 million Ascend dies in 2026, enough base components for around 600,000 of the company's marquee 910C chips plus other models2
.The choice of Huawei AI chips over Nvidia accelerators reflects the impact of US export restrictions on China's AI infrastructure development. Nvidia's most capable AI accelerators cannot legally be sold in China without US permission, leaving the Ascend series as one of the main domestic options for Chinese AI labs building computing capacity at scale
1
. Jensen Huang described DeepSeek running on Huawei chips as a "horrible outcome" for America, arguing that export controls risk pushing China's AI software ecosystem toward Chinese hardware1
. DeepSeek is choosing Huawei's Ascend 950DT chips despite a roughly 4x performance hit compared to Nvidia's H20 GPUs and not-too-hefty price differential3
.
Source: Wccftech
The Ascend 950DT chip is built on Huawei's self-developed third-generation DaVinci v3 architecture and leverages SMIC's N+3 process technology, an advanced 7-nanometer class process
3
. Each chip comes equipped with 144 GB of specialized High Bandwidth Memory (HiZQ 2.0 HBM), achieving a 4.0 TB/s memory bandwidth and a 2.0 TB/s GPU-to-GPU interconnect speed3
. The chip natively supports low-precision formats such as FP8, FP4, and Huawei's proprietary HiF83
.Related Stories
Every model optimized for the Ascend series creates more software and tooling that does not depend on Nvidia's CUDA ecosystem, which has been one of the company's biggest advantages and a major barrier to competing hardware
1
. Steven Glinert, CEO of Sphere Semi, suggested that Nvidia seems to have given up on China as a lost cause3
. DeepSeek has been moving in this direction for months, with its V4 model tuned to run on Huawei's Ascend 950 when it launched in April1
.
Source: The Next Web
The installation would dwarf other publicly known clusters of Huawei's Ascend processors, representing meaningful progress as China tries to wean itself off Nvidia components
2
. Just six months ago, Shenzhen began operations of China's first ever 10,000-chip intelligent computing cluster using the Ascend series, tiny compared to data centres in the US2
. DeepSeek is in talks to raise billions of dollars to build out its infrastructure, with artificial general intelligence described as the company's goal by its founder1
.DeepSeek has separately been reported to be developing its own inference chip with SMIC, which would make a large Huawei order more of a bridge than a permanent destination if that project succeeds
1
. Export controls have encouraged domestic developers to move from general-purpose GPUs toward custom ASICs, which can be cheaper to produce on the manufacturing processes available inside China1
. The economics inside China are different as DeepSeek has been able to price aggressively in part because it does not have to buy Western hardware at Western prices, while a domestic chip supply chain helps preserve that advantage1
.Summarized by
Navi
10 Jul 2025•Technology

05 Feb 2025•Technology

21 Apr 2025•Technology

1
Policy and Regulation

2
Technology

3
Technology
