DeepSeek는 서구 가속화 공급업체 의존도를 낮추기 위해 신규 계획 데이터센터에 Huawei AI 칩을 대량 도입하는 주요 주문을 진행 중이다.

DeepSeek plans to deploy at least 160,000 of Huawei Technologies Co.’s top accelerators at a massive data center it is building in Inner Mongolia. The project could create one of the largest known clusters of Huawei artificial intelligence (AI) chips and accelerate China’s efforts to replace Nvidia Corp.

According to people familiar with the matter, the AI pioneer intends to use Huawei’s next-generation Ascend 950DT chips to run its models, though installation timelines will hinge on Huawei’s production capacity. One source noted that DeepSeek does not currently plan to use the 950DT chips for training, despite Huawei designing and marketing them for that more demanding workload. While DeepSeek attempted to train earlier models on domestic hardware, it has relied primarily on Nvidia accelerators for this critical step.

DeepSeek seeks to purchase additional units of Huawei’s leading chips, but the company lacks the immediate production capacity to fulfill the demand. Shortages of critical components, including high-end memory, will limit Huawei’s 950DT output to the low hundreds of thousands this year, according to one source. Balancing orders from other clients while attempting to export limited volumes overseas means fulfilling DeepSeek’s request could take over a year, the sources said, speaking anonymously regarding a confidential agreement.

If completed as planned, the Inner Mongolia facility will be central to DeepSeek’s AI development. The startup is currently in discussions to raise billions of dollars to expand its infrastructure. Once operational, the installation would surpass all other publicly documented Huawei Ascend processor clusters—a significant milestone as China works to reduce its reliance on Nvidia components. According to one individual familiar with Huawei’s manufacturing capabilities, China may host several similarly sized clusters by the time this facility goes live.

DeepSeek did not respond to requests for comment, and a Huawei spokesperson had no immediate remark.

Deploying over 100,000 chips in a single facility places DeepSeek’s ambitions on par with data centers routinely constructed by Western firms, albeit using less capable processors. Nvidia accelerators remain the global standard for training and deploying AI models, yet their sale to China is heavily restricted without U.S. approval.

Huawei remains one of a select group of companies pivotal to Beijing’s objective of achieving AI dominance despite U.S. sanctions. The Shenzhen-based telecommunications giant has become China’s de facto national semiconductor leader, while Hangzhou-headquartered DeepSeek has established a blueprint for disruptive AI model launches that competitors like Moonshot have since emulated.

China’s capacity to manufacture—and eventually export—domestic AI chips stands as a defining element of its technological rivalry with the United States. This capability is closely monitored by U.S. policymakers, who have imposed years of export controls on firms including Huawei to impede its advancement.

These restrictions have significantly constrained Huawei’s production volume and performance. Bloomberg News reported last year that Huawei aims to produce approximately 1.6 million Ascend dies in 2026, providing enough base components for roughly 600,000 of its flagship 910C chips alongside smaller batches of other models. Unveiled roughly a year ago, the 950DT represents an upgraded, higher-performance variant widely considered comparable in capability to Nvidia’s prior Hopper architecture.

Nevertheless, the prospect of a 160,000-chip cluster demonstrates tangible progress. Six months ago, an official publication from the technology hub Shenzhen announced the launch of China’s first 10,000-chip intelligent computing cluster utilizing Huawei’s Ascend series—a scale dwarfed by U.S. facilities. In comparison, Elon Musk’s xAI deployed a cluster ten times larger in 2024.

In June, Chinese food-delivery giant Meituan announced it trained an AI model using 50,000 domestic chips, declining to provide further specifics. Meanwhile, Z.AI Co. completed construction of a 1-gigawatt data center in July, intending to equip it exclusively with Chinese-made silicon. Its current installed base remains modest, estimated at around 10,000 accelerator cards.

San Jose-based Nvidia has argued that China’s pivot toward domestically produced silicon is a direct result of prolonged U.S. export controls. Washington initially implemented these restrictions in 2022, citing national security concerns that advanced American AI chips could enhance China’s military capabilities.

Beijing is actively encouraging Chinese AI laboratories to transition to domestic hardware, though adoption rates vary across the sector.

Despite government directives, China’s private sector continues to favor Nvidia accelerators, as demonstrated by substantial procurement deals from Alibaba Group Holding Ltd. and Tencent Holdings Ltd. following recent easing of U.S. restrictions under the Trump administration. Furthermore, Chinese firms can still access vast quantities of Nvidia’s most advanced chips by leasing overseas computing capacity, a practice that typically complies with U.S. regulations. Washington has also alleged that companies including Moonshot and DeepSeek have utilized smuggled Nvidia hardware to accelerate their AI development.

However, Beijing maintains a cautious stance on imports. Although the Trump administration signaled willingness to permit up to one million Nvidia H200 chips—the flagship model from a prior generation—Chinese authorities have approved only a fraction of that volume for entry.

DeepSeek is among several Chinese firms that have optimized their models to operate on domestic accelerators. According to one source, the company has sought Beijing’s intervention to persuade Huawei to increase chip allocations to DeepSeek and expedite deliveries.

China’s Ministry of Industry and Information Technology, which regulates national technology policy, did not respond to a faxed request for comment.

DeepSeek’s AI infrastructure expansion in Inner Mongolia targets a gigawatt-scale footprint. At full utilization, the facility would consume sufficient electricity to power approximately 750,000 households. The planned deployment of 160,000 Ascend 950DT chips represents only a portion of the site’s total computational capacity.

The specific types of accelerators intended for the remaining capacity remain undisclosed. Officials in the Trump administration have alleged that DeepSeek acquired Nvidia’s latest Blackwell chips and installed them at the Inner Mongolia site, a claim Bloomberg has not independently verified.