MiTAC unveils liquid-cooled AI servers for agentic AI era

MiTAC’s WAIC showcase is a snapshot of where the AI infrastructure race is heading: from generic GPU racks to tightly engineered compute “appliances” explicitly marketed for agentic AI. A 52U cabinet packing up to 96 AMD Instinct MI355X accelerators, combined with custom cold‑plate liquid cooling, is not just about raw FLOPs—it’s about making frontier‑scale experimentation feasible in power‑ and space‑constrained data centers. By wrapping high‑density GPU clusters, OCP‑standard liquid‑cooled systems and DDN‑integrated storage into a coherent product suite, MiTAC is positioning itself as a one‑stop shop for enterprises that want OpenAI‑style capabilities without building hyperscale infrastructure from scratch.([news24.tw](https://www.news24.tw/tech/6175/))

Strategically, this reinforces two trends that matter for the race to AGI. First, Nvidia’s dominance is being seriously challenged at the rack level: AMD‑centric designs like MiTAC’s give buyers credible leverage and could ease some of the GPU supply bottlenecks that currently gate large training runs. Second, the marketing emphasis on “agentic AI” hints at how vendors see demand evolving: from batch training and inference toward always‑on, multi‑agent systems that coordinate complex workflows. Those systems are far more sensitive to memory bandwidth, interconnect topology and data‑pipeline design—exactly the domains these integrated cabinets are meant to optimize. As more Tier‑2 cloud and enterprise players adopt this kind of hardware, frontier‑class experimentation will no longer be confined to a handful of US hyperscalers.