MiTAC shows at WAIC what really holds back agentic AI: it's not the chips, it's the cooling

By Macau Business (PR Newswire) · July 17, 2026.

MiTAC Computing Technology, a subsidiary of Taiwan's MiTAC Holdings (TWSE:3706), took advantage of WAIC 2026 in Shanghai to showcase its infrastructure catalog for the 'agentic era': four complete racks and several individual servers under the banner 'Liquid Cooling. Infinite Compute. Sustainable Future.' The flagship is a 52U liquid-cooled rack that packs 12 G4826Z5 servers with up to 96 AMD Instinct MI355X GPUs and dual EPYC processors, with 50% more GPU density than a standard 48U rack. Alongside it are an air-cooled version (G8825Z5, up to 32 MI350X GPUs), a rack compliant with the open OCP ORv3 standard designed for hyperscale clouds, and a high-density storage rack that integrates the DDN Infinia data platform to accelerate RAG and agentic inference. The lineup is rounded out by servers for compute (Intel Xeon 6, up to 8 GPUs per node) and for enterprise storage.

This is, in essence, a product press release: there are no sales figures, named customers or independent performance comparisons, so it is best treated as what it is, a commercial catalog at a trade show, not a verified technological milestone. But the interesting angle lies not in the specific specs, but in where MiTAC has decided to place its focus: not on 'more GPUs,' but on how to cool them, package them and feed them data without bottlenecks.

That choice confirms something we have been observing in the industry: as agentic AI moves from generating one-off text to executing chains of tasks continuously, the competitive limit shifts from the GPU itself toward the infrastructure that sustains it. Compute density per rack is no longer limited by the chips available, but by how much heat can be extracted from a finite physical space and how much energy it costs to do so (PUE, or power usage effectiveness, appears explicitly as a selling point). That MiTAC builds on AMD Instinct GPUs instead of Nvidia is also relevant as context: the AI compute supply chain is diversifying, and niche server manufacturers like MiTAC —with decades of experience in motherboards and systems, but little known outside the sector— are the ones that in practice decide which architectures reach real data centers.

Our reading is that announcements of this kind, however modest they may appear, are a more reliable thermometer of the industry's real state than many model launches: while labs compete for the benchmark headline, the physical capacity to deploy those models at scale —liquid cooling, high-performance storage, open standards like OCP ORv3— is being built in parallel, and in a less flashy way, at trade shows like WAIC. It is the boring but indispensable part of the computational abundance that underpins long-term optimism about AI: without these advances in energy efficiency and density, scaling artificial intelligence to the point of generating the surplus of resources the abundance thesis promises would be simply unfeasible on energy cost. The transition does not depend solely on better models; it depends, in equal measure, on someone solving how to keep 96 GPUs cool without sending the electricity bill soaring.