DeepSeek hid the pre-release of its powerful new AI model from Nvidia and AMD - Reuters

Chinese AI startup DeepSeek has not given U.S. chipmakers access to pre-release tests of its flagship V4 model, which is expected to launch soon, sources told Reuters. But it has allowed Chinese chipmakers, including Huawei, to test it, the sources said. By doing so, the company has departed from standard industry practice on the eve of a major model refresh, Reuters notes. A year ago, the release of the first DeepSeek R1 model led to a collapse in the shares of U.S. technology companies.

Chinese startup DeepSeek has not given U.S. chipmakers a chance to test its latest model, sources told Reuters. AI developers typically share pre-release versions of major models with leading chipmakers such as Nvidia and Advanced Micro Devices to ensure the software runs efficiently on widely used hardware, the agency explains. DeepSeek has previously worked closely with Nvidia's technical experts, it notes. This time, however, the lab, which is expected to unveil the next major version - V4 - soon, has granted early access to domestic vendors, including Huawei Technologies.

Nvidia and AMD declined to comment to Reuters, while DeepSeek and Huawei did not respond to the agency's requests for comment.

DeepSeek has given Chinese chipmakers a head start of several weeks to optimize software for their processors, Reuters' interlocutors claim.

"The impact on Nvidia and AMD in the general-purpose processing gas pedal segment is minimal - most enterprise customers don't use DeepSeek, which serves more as a benchmark than anything else," Ben Bajarin, CEO of research firm Creative Strategies, told the agency. He added that the new AI programming tools reduce the time it takes to optimize software on specific hardware "from months to weeks."

Bajarin said the move is likely part of a broader strategy by the Chinese government "to keep U.S. equipment and models at a disadvantage" in the PRC market.

A year ago, in January 2025, the release of the first DeepSeek AI model led to the collapse of the US market. It was developed under export restrictions and with minimal investment, which called into question the need for colossal expenditures on AI training in the US.

Since the Chinese company entered the AI market, its models have been downloaded more than 75 million times on the open-source platform Hugging Face, fueling a wave of Chinese open-source models competing with U.S. AI labs. Among models released over the past year, Chinese developments on the platform have surpassed those from any other country in terms of downloads, Reuters writes.

The new V4, according to The Information's sources, has powerful programming capabilities and beat out competitors from Anthropic and OpenAI in internal testing. It could be trained on Nvidia's most advanced chips, Blackwell, in mainland China, which appears to violate the supply ban, Reuters wrote on Feb. 24.

This article was AI-translated and verified by a human editor