With China's AI inference data volume surpassing training for the first time, the industry's compute demand center of gravity is migrating from centralized training toward scaled inference services, requiring infrastructure to sit closer to the application side, Communication World reported on November 1.
In related moves, Nvidia demonstrated its next-generation CPO (co-packaged optics) electro-optical integration scheme, pushing optical engines closer to compute chips and extending optical-communications competition into chip packaging. Sanan Optoelectronics broke through the AIDC power-efficiency ceiling with a "low-resistance substrate plus soft-switching" approach, while UNIS (Unisplendour) deployed its next-generation domestic thousand-card supernode — based on a self-developed G-Link open high-speed interconnect — jointly validated with multiple domestic GPUs and piloted at a top internet firm, with large banks beginning trials.
Commentators said system-level integration and domestic GPU interconnect ecosystems are entering real-world stress tests, and that scaled deployment capability will determine subsequent competitiveness.
Source: Communication World
