Exclusive-DeepSeek withholds latest AI model from US chipmakers including Nvidia, sources say

By Krystal Hu, Stephen Nellis and Fanny Potkin

SAN FRANCISCO/SINGAPORE, Feb 25 (Reuters) – DeepSeek, the Chinese artificial intelligence lab whose low-cost model rattled global markets last year, has not shown U.S. chipmakers its upcoming flagship model for performance optimization, two sources familiar with the matter said, breaking from standard industry practice ahead of a major model update.

Instead, the lab, which is expected to launch its next major update, V4, granted early access to domestic suppliers, including Huawei Technologies, the sources said. 

AI developers typically share pre-release versions of major models with leading chipmakers such as Nvidia and Advanced Micro Devices to ensure their software performs efficiently on widely used hardware. DeepSeek has previously worked closely with Nvidia’s technical staff.

NVIDIA, AMD LEFT OUT

For its forthcoming model, which was expected to be released around the Lunar New Year holiday, DeepSeek did not provide access to Nvidia and AMD and gave Chinese chipmakers, including Huawei, a head start of several weeks to optimize the software for their processors, the sources said.  

Nvidia and AMD declined to comment. DeepSeek and Huawei did not respond to requests for comment.

Reuters could not immediately determine the reason for the decision.

“The impact to Nvidia and AMD for general data accelerators is minimal – most enterprises are not running DeepSeek, which serves as a benchmarking model more than anything else,” said Ben Bajarin, CEO of research firm Creative Strategies. He added that new AI coding tools are reducing the time it takes to make software run well on hardware, “from months to weeks.” 

The move is likely part of a broader strategy by the Chinese government “to try to keep U.S. hardware and models disadvantaged” in China, Bajarin said.

The development comes as a senior Trump administration official told Reuters DeepSeek’s latest AI model was trained on Nvidia’s most advanced chip, Blackwell, using a cluster in mainland China, in a move that appears to violate U.S. export controls.

DeepSeek may seek to remove technical indicators revealing its use of American AI chips, and plans to publicly claim that it used Huawei’s chips to train its model, according to the U.S. official. 

DeepSeek’s models have been downloaded more than 75 million times on the open-source platform Hugging Face since the company burst onto the scene in January 2025, helping fuel a wave of Chinese open-source models competing with U.S. AI labs. Among models released in the past year, downloads for Chinese models have surpassed those from any other country on the platform.

The rapid rise of Chinese open-source models has intensified debate in Washington over exporting advanced U.S. AI chips to China. U.S. authorities last year allowed Nvidia’s H20 and AMD’s MI308 chips — designed for AI inference — to resume shipments to China, even as licenses for more advanced processors remain restricted. It was unclear whether DeepSeek has secured approval to purchase those U.S. chips.

The H20 and MI308 chips are aimed at inference, the process of running trained AI models. Demand for the MI308 was significant, with AMD saying it generated $390 million in sales of the chip in its most recent quarter.

DeepSeek is among several Chinese AI firms expected to unveil new models this month.

(Reporting by Krystal Hu, Stephen Nellis in San Francisco and Fanny Potkin in Singapore; Editing by Kenneth Li, Rod Nickel)