Chinese AI infrastructure firms push supernodes, inference software and cooling
China’s AI compute sector saw new domestic accelerator systems, inference software, liquid-cooling plans and foundation-model product expansion.
• H3C launched its UniPoD S81000 X1 supernode, supporting up to 40 domestic accelerator cards in one Scale-Up domain with 448GB/s peer-to-peer bandwidth.
• Shishi Technology launched its Meta-Infer inference engine, which aims to improve the performance of existing GPUs through software optimization without changing model architectures or task semantics.
• Sichuan Run Investment removed energy storage from a private-placement fundraising plan and shifted its focus toward liquid cooling.
• DeepSeek released the Harness v0.2 desktop preview for macOS and Windows, adding plugin installation and management alongside office and development features.
• Doubao is testing a personal-assistant product internally under the codename Spell and is expanding its AI assistant ambitions into travel, hotels, ride-hailing and other local services.
• Mentech began small-batch shipments of 800G LPO optical modules, adding capacity to the high-speed networking supply chain.