DeepSeek is reported to have open-sourced a complete development platform for Huawei’s Ascend AI chips. The covered tools include TilesLang Ascend, DeepGEMM, DeepEP and FlashMLA, and the article outlines preparations Japanese companies should make from the perspectives of US-China regulations and inference costs.
On 30 September, DeepSeek released an open-source development platform for Huawei’s Ascend AI semiconductors. In addition to the Ascend version of TileLang, libraries for computation, communications and long-context processing, including DeepGEMM, DeepEP and FlashMLA, are being provided as a package, with joint optimisation also under way for supernode configurations using Ascend 950.
US export controls and economic security constraints have been cited as background, and the view has been expressed that Japanese companies cannot easily make the transition based solely on the availability of a low-cost option. At the same time, the expansion of abstraction technologies that do not depend on CUDA is broadening the potential for software to close the hardware gap.
Challenges for Japanese companies using Ascend include EAR/BIS regulatory risks, the location where data is stored and the applicable laws, and the requirement for prior notification before deployment in critical infrastructure. In light of these factors, the question is how to manage ease of migration in system design and procurement risks.
The article proposes 3 actions to ensure migration flexibility: building an organisation that can respond to regulatory changes across IT, legal and economic security functions; taking inventory of workloads and designing flexible infrastructure; and securing portability.