• DeepSeek withholds pre-release access to its V4 model from US chipmakers
• Move highlights growing China-US friction over AI technology and export controls
What happened: Restricted access to V4 model
DeepSeek—a Chinese artificial intelligence company known for its cost-efficient large language models—has withheld early access to its upcoming flagship AI model, dubbed V4, from major US chipmakers including Nvidia and AMD, two sources familiar with the matter told Reuters. Normally, developers share pre-release model versions with hardware partners so they can optimize performance on popular chips, but DeepSeek granted early access to domestic suppliers, including Huawei Technologies, instead.
This departure from standard practice comes as DeepSeek prepares to launch the new model around the Lunar New Year holiday period. Nvidia and AMD—both of which usually receive early datasets and model builds to tailor software libraries and drivers—were excluded from this process, potentially placing them at a performance disadvantage when the model is publicly released.
Reuters could not immediately determine DeepSeek’s rationale for this decision, and both DeepSeek and Huawei declined to comment. Nvidia and AMD also declined to respond to requests for comment.
Industry observers note that DeepSeek’s models have been widely downloaded on open-source platforms, garnering significant attention, and this latest development is likely part of a broader strategy amid intensifying technological competition between China and the United States.
Also read: DeepSeek’s role in shaping telecom AI remains uncertain
Also read: US and APAC allies ban DeepSeek from government devices
Why it’s important:
DeepSeek’s move underscores the broader geopolitical tensions around AI technology and hardware. The China-US technology rivalry has already seen export controls on advanced semiconductors, with the United States seeking to limit Beijing’s access to cutting-edge chips. In a related development, a senior US administration official said DeepSeek’s latest model had been trained on Nvidia’s most advanced Blackwell chips in mainland China—a potential point of contention under export restrictions.
The exclusion of Nvidia and AMD from early access complicates the usual optimization pipeline that ensures AI models run efficiently on widely adopted hardware. Even though the V4 model will likely be open-source upon release, the lost head start in optimization could mean that hardware from Western chipmakers may perform less efficiently initially compared to Chinese chips like those from Huawei.
At a time when AI workloads are becoming deeply intertwined with national competitiveness, DeepSeek’s decision may reflect strategic choices to bolster domestic technology ecosystems. However, it raises questions about global interoperability and whether such moves will fragment the AI landscape along geopolitical lines.
Observers will be watching how this development influences chipmakers’ relationships with AI developers and how export control policies evolve in response to such tactical shifts in industry practice.
