From CDUs to diamonds: AI infrastructure at Computex 2026

AI systems are everywhere at Computex 2026. Here's what struck me about the latest offerings, and what they mean for data centres. I dropped by MiTAC this morning. Here are my thoughts.

Liquid cooling in the mainstream

Just a few years ago, liquid cooling was hardly mentioned. But soon after ChatGPT kicked off the AI era, CDUs started showing up at conventions. Last year, I wrote about the bewildering array of in-rack and in-row CDU offerings. Today, they are squarely in the mainstream, offering redundant configurations, built-in N+1 redundancy, high-capacity in-rack units, and massive multi-MW capacity.

The range I saw today bears this out. It ran from up to 250kW (OCP) for an in-rack unit by Nidec, to 2MW for an in-row CDU by CoolIT. The latter even offers N+1 redundancy for its pump.

Full-stack platform

GPUs are no longer deployed alone, but as part of an AI infrastructure stack. One can always buy fully into Nvidia's ecosystem, of course.

But most will prefer alternatives, and will need to think about several layers at once: the hardware, such as servers and storage; the orchestration for GPUs and AI workloads; and the platform for bare-metal control and monitoring, among others.

It's worth noting that many of these decisions are ideally made up front.

Hybrid deployments

We all know that liquid cooling is the future of data centres. But it is a future that will take time to materialise. In the meantime, air-cooled servers will likely stay with us for a long while yet.

This means data centre operators must cater to hybrid deployments, designing for both air-cooled and liquid-cooled servers. In fact, I can't think of any recent announcements of a liquid-only data centre in Southeast Asia, or even across the Asia Pacific.

Better performance

Liquid-cooled solutions make it possible to deploy more GPU servers than before. But solution providers continue to find ways to push the envelope.

By using either in-rack or in-row CDUs, MiTAC is able to fit 96 AMD MI355X GPUs in a single 52U rack, or 50% more than AMD's reference design. And by using lab-grown diamond from partner Akash Systems for its superior thermal conductivity, MiTAC claims up to a staggering 50% token throughput improvement for an air-cooled server.

To close on storage: while most of the talk around RAM tends to revolve around its current scarcity, it's worth remembering that AI has tremendously increased demand for storage too.