DeepSeek Blocks US Chipmakers from Its New AI Model

DeepSeek breaks from standard practice, keeping its V4 flagship model from U.S. chipmakers Nvidia and AMD.

In a break from industry protocol, Chinese AI lab DeepSeek has not provided its upcoming flagship model to US chipmakers for optimisation. Two sources familiar with the matter told members of the press that the move comes ahead of a major update for the company whose low-cost model rattled markets last year.

Instead, domestic vendors, notably Huawei Technologies, were given early access to the lab, which is scheduled to release its next significant update, V4, according to the sources.

To make sure their software works well on widely used hardware, AI developers usually share pre-release versions of significant models with top chipmakers like Nvidia and Advanced Micro Devices. Previously, DeepSeek collaborated closely with the technical team at Nvidia.

The sources claimed that DeepSeek gave Chinese chipmakers, including Huawei, a few weeks’ head start to optimise the software for their CPUs for its upcoming model, which was anticipated to be released around the Lunar New Year vacation.

AMD and Nvidia choose not to comment. Requests for responses from DeepSeek and Huawei were not answered. Reuters was unable to ascertain the decision’s rationale right away.

Ben Bajarin, CEO of research firm Creative Strategies, stated, “The impact on 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.” 

New AI coding techniques, he continued, are cutting the time it takes to get software to function properly on hardware “from months to weeks”.

Bajarin said the move likely reflects a Chinese government strategy to keep U.S. hardware and models at a competitive disadvantage in China.

In an apparent violation of U.S. export regulations, a senior Trump administration official told Reuters that DeepSeek’s most recent AI model was trained on Nvidia’s most sophisticated chip, Blackwell, using a cluster in mainland China.

A US official alleges DeepSeek plans to remove technical indicators tying it to American AI chips and falsely assert it used Huawei chips for training.

Since its launch in January 2025, DeepSeek’s models have been downloaded over 75 million times on the open-source platform Hugging Face, contributing to a surge of Chinese open-source models that are competing with US AI laboratories. Chinese models have received more downloads than models from any other nation on the site among those launched in the last 12 months.

The argument over transferring cutting-edge US AI processors to China has heated up in Washington due to the quick rise of Chinese open-source models. While licensing for more sophisticated processors is still restricted, US officials this year permitted the resumption of shipments of AMD’s MI308 and Nvidia’s H20 chips, which are intended for AI inference. Whether DeepSeek had obtained authorisation to buy the US chips remained unknown. 

Inference, or the execution of taught AI models, is the focus of the H20 and MI308 CPUs. The MI308 was in high demand; according to AMD, sales of the chip reached $390 million in its most recent quarter.

Several Chinese AI companies, including DeepSeek, are anticipated to release new models this month.

A view of the important developments in the month of February 2026.

• Favouring Domestic Suppliers: A few weeks before the global release, DeepSeek gave Chinese suppliers, such as Huawei Technologies, early access to optimise the model for their hardware.
• Strategic Disruption: By guaranteeing that domestic chips run more smoothly on their cutting-edge software, this action is perceived as a “retaliatory message” as a tactic to undercut American hardware. 
• Concealment Strategies: According to reports, DeepSeek may try to conceal technical clues that point to the usage of American AI processors and openly attribute the training process to Huawei hardware.
• Market Impact: The announcement has already made Nasdaq stocks tremble, even though the exclusion of American chipmakers from early optimisation may be constrained by contemporary AI coding tools.

The model developed by DeepSeek was allegedly trained on Nvidia’s cutting-edge Blackwell chips in mainland China, which may have violated U.S. export regulations, according to a senior U.S. official.