DeepSeek, the Chinese artificial intelligence laboratory whose low cost model unsettled global markets last year, has not shared its upcoming flagship system with United States chipmakers for performance optimisation, according to two sources familiar with the matter.
The move marks a departure from standard industry practice ahead of a major model release. Typically, developers provide early access to key hardware partners to fine tune performance before launch. However, DeepSeek has taken a different path as it prepares its next major update.
Break From Industry Practice
DeepSeek is expected to launch its new model, known as V4, in the near future. Yet the company has not shown pre release versions to leading United States chipmakers for testing and optimisation.
AI developers usually collaborate closely with firms such as Nvidia and Advanced Micro Devices before unveiling large scale systems. By sharing early builds, they allow hardware partners to ensure that software runs efficiently on widely deployed chips. This process often improves stability and performance at launch.
Nevertheless, DeepSeek has chosen not to follow that established approach this time. Instead, the company withheld access from American chipmakers as it prepared its next flagship release.
Early Access For Domestic Suppliers
Rather than working with United States suppliers, DeepSeek granted early access to domestic partners, including Huawei Technologies, the sources said. As a result, Chinese hardware providers gained the opportunity to optimise the model for their own systems ahead of release.
This strategy signals a potential shift in collaboration patterns within the global AI ecosystem. While international partnerships have long shaped model development, domestic cooperation now appears to be taking priority in this case.
Previous Collaboration With Nvidia
In earlier projects, DeepSeek worked closely with Nvidia’s technical staff. Therefore, the company had previously aligned with standard industry practice by engaging leading chipmakers during model development.
Developers often rely on such partnerships because they help align software advances with hardware capabilities. Consequently, early technical coordination can improve efficiency across widely used computing platforms.
However, DeepSeek’s decision not to share V4 with United States chipmakers represents a notable change from that pattern. As the anticipated update approaches, attention now turns to how the model will perform and which hardware ecosystems will benefit most from its optimisation.
Inputs from Reuters