Two people familiar with the matter told Reuters that DeepSeek did not supply its forthcoming flagship V4 model to U.S. chipmakers for the sort of performance tuning that developers commonly seek before a major release. Instead, the lab provided early access to domestic hardware suppliers, among them Huawei Technologies, the sources said.
DeepSeek had been expected to unveil the V4 update around the Lunar New Year holiday. The decision to limit or withhold pre-release access from U.S. semiconductor firms represents a notable departure from the prevailing industry pattern, under which AI developers typically circulate prerelease model builds to leading chip vendors so that software can be optimized to run efficiently on widely used hardware.
Industry practice often involves close technical collaboration between model teams and chipmaker engineers to tune performance, memory usage, and throughput for specific accelerators. DeepSeek has in the past worked closely with technical staff at Nvidia, according to the information provided, underscoring that the change in distribution for V4 is distinct from some of the lab's earlier interactions with chip vendors.
Separately, Reuters reported earlier that a senior Trump administration official told the news agency that DeepSeek's latest model was trained on Nvidia's most advanced Blackwell chip using a cluster located in mainland China. That account said the training setup appears to run afoul of U.S. export controls, though the reporting framed this as an appearance rather than a definitive finding.
The lab's choice to restrict distribution of the V4 prerelease to non-U.S. chipmakers and to favor domestic suppliers will likely draw scrutiny because sharing pre-release models with hardware partners is commonly used to ensure software performance across dominant platforms. The available information does not provide additional detail on the technical scope of the early access given to domestic suppliers or on any subsequent testing results.
With the expected release window near the Lunar New Year, stakeholders across AI development and semiconductor manufacturing are positioned to watch how DeepSeek's rollout proceeds and how hardware partners respond to the constrained prerelease access.