Executive Summary

• What makes a data centre ‘AI-ready’? I explore power, cooling strategies, redundancy, latency and the skills shortage to determine if the infrastructure can cope with the high-density AI workloads.
• Being “AI-ready” in terms of power means the electrical infrastructure and racks can deliver the density without downtime or compromising the resilience in faculty.
• A data centre can only be as “AI-ready” as the team responsible for keeping it running.

The term “AI-ready data centre” has been floating about for a while, but what does this mean? As AI settles in and is here to stay and change our evolving digital world forever, we explore what it means when a data centre is classed as “AI-ready”.

Simply, it is infrastructure that is built to handle the intense compute demands of AI workloads – from the racks themselves to power, cooling and resilience requirements.

The time for AI experimentation in organisations has passed; we’re moving to production-scale deployment and we need the facilities to handle the demand and be ready for more. The largest gap is between what standard data centres deliver and what AI workloads need, and it’s now impossible to ignore because of the high demand and reliance on technology.

Does it handle the power?

Traditional racks run at 5-15 kW, which isn’t so demanding compared to the AI workloads demanding 50-100kW and beyond that, in the case of some HPC deployments. Being “AI-ready” in terms of power means the electrical infrastructure and racks can deliver the density without downtime or compromising the resilience in faculty. This is a world away from the densities most existing facilities were designed for.

Alongside having the power on-site, UPS systems, switchgear and busbar systems need to be built for the new load requirements. However, if grid connections don’t exist or can’t support it, none of this investment on-site matters because the site won’t be able to handle those AI workloads.

Can it keep its cool?

Air cooling is redundant after 30-40kW per rack, so liquid cooling holds together the cooling strategy to handle intense workloads – or at the very least, a hybrid solution to optimise infrastructure for efficiency.

A credibly cooling strategy for mixed-density environments, as well as solutions for peak loads in isolation, means that an environment is AI-ready.

Can the physical environment handle it?

For retrofits, turning them into AI-ready infrastructure can be a challenge as they face floor loading, containment architecture and structural headroom – all of them potential constraints that can stop the project in its tracks. Usually, there are physical limits to infrastructure before there are electrical ones and the best way around it is a detailed, honest assessment of the building and what the existing environment can handle physically.

Can the network keep up and stay up when it matters most?

GPU clusters are data hungry and latency sensitive, so being AI-ready means fibre density, high speed and interconnect capacity matching the workload intensity. Not only that, but AI workloads are volatile and have power surges, so it means there must be real-time monitoring and DCIM systems to detect and respond to issues in near real-time. Especially with an operational expectancy of near 100%, infrastructure must have a redundancy strategy, maintenance windows reclibrated and tiers of resilience.

Can it prove sustainability credentials?

PUE, WUE and carbon reporting and renewable energy sourcing are becoming commercial requirements fast; with hyperscalers embedding sustainability thresholds and requirements into the contracts with co-location providers. As reporting becomes more transparent and mandatory, sustainability commitments must be part of the data centre strategy.

Does it have the people to run it?

One of the hardest questions in this article, and the question that is asked the least. With AI being pretty ‘new’ in the digital world, there aren’t enough people with the skills to operate the new high-density AI infrastrucutre and this could be the largest constraint of them all.

A data centre can only be as “AI-ready” as the team responsible for keeping it running.