DeepSeek delays V4 release to optimize for Huawei hardware

DeepSeek has ported its core libraries to Huawei chips, achieving gross margins of approximately 80%, a figure comparable to those of Anthropic. To maximize efficiency on this domestic infrastructure, the company reportedly delayed the release of its V4 model. The decision prioritized optimizing inference performance for Huawei hardware rather than launching the model earlier on Nvidia systems. Huawei’s Ascend chips currently lag behind Nvidia by about five years in performance, a gap that analysts suggest is compensated for by volume. However, production remains constrained by a shortage of High Bandwidth Memory, a product primarily supplied by SK Hynix. The broader supply chain for these chips relies on ASML for lithography machines, Zeiss for mirrors, and TSMC as a key manufacturer. In a related development, the open-weights Kimi K31 Pro model has emerged as a competitive alternative, performing at a level comparable to US frontier models like Fable. Kimi ranks behind Opus, Sonnet, Muse, and Wonder, with pricing set at 2 dollars per million input tokens and 10 dollars per million output tokens. Reports indicate that influencers were instructed not to disclose that their positive coverage of Kimi was paid content. This ecosystem reflects a broader strategy where China leverages volume and inference optimization to mitigate hardware performance gaps.