DeepSeek, the Chinese artificial intelligence (AI) lab that roiled markets and Silicon Valley a year ago, is at it again: It reportedly is breaking industry protocol by freezing out U.S. chipmakers from its latest development cycle.
The lab has withheld its forthcoming flagship model, V4, from NVIDIA Corp. and AMD Inc., sources familiar with the matter told Reuters. Instead, DeepSeek has granted exclusive early access to its domestic suppliers, including Huawei Technologies, that gives them a weeks-long head start to optimize the software for Chinese-made hardware.
In the high-stakes world of AI, developers typically share pre-release versions of major models with leading chip manufacturers. This collaboration ensures that new software runs seamlessly on the world’s most prevalent hardware. DeepSeek itself has a history of working closely with NVIDIA’s technical teams, making the arrangement with Huawei a significant departure from standard operating procedure.
If the reports are accurate, they “highlight the complex interests involved in digital sovereignty,” said Mitch Ashley, vice president and practice lead, Software Cycle Engineering, at The Futurum Group. “The pattern is visible and accelerating. I see the larger question as whether AI agent capability, the systems that will autonomously execute software development at scale, is a national security asset. Nations and enterprises that control the AI technologies used for agent development, execution, observability, and governance own a strategic chokepoint that compounds almost daily as the AI innovation cycle rapidly collapses.”
The timing of this technical freeze-out coincides with mounting geopolitical tension. A senior official from the Trump administration recently alleged that DeepSeek’s current models were trained on NVIDIA’s restricted Blackwell chips within mainland China, which is a potential violation of U.S. export controls.
According to the official, DeepSeek may be attempting to scrub technical indicators of its reliance on American silicon. There are further suggestions that the lab plans to publicly credit Huawei’s processors for the training of its V4 model to satisfy domestic political narratives, regardless of the underlying hardware used during development.
Despite the friction, DeepSeek’s influence continues to surge. Since January 2025, the lab’s models have surpassed 75 million downloads on the Hugging Face platform. The explosion in popularity has made Chinese open-source models the most downloaded in the world over the past year, intensifying the debate in Washington, D.C., over whether to further tighten restrictions on AI inference chips like NVIDIA’s H20 and AMD’s MI308.
As DeepSeek and other Chinese firms prepare a battery of new model releases this month, the wall between Eastern software and Western hardware appears to be growing taller.
