As an enterprise buying compute, you’re likely looking for a facility that can host AI workloads. Sure enough, every facility is tripping over itself to tell you that they’re an AI-ready data centre. When it comes to density, power, and cooling, what does this term actually mean?

Just as ‘sustainable’ doesn’t tell you much about a data centre’s environmental impact, ‘AI-ready’ doesn’t get into the nitty-gritty of how a facility is prepared for AI workloads. The term is being thrown around by legacy facilities that, in reality, would require costly, timely retrofitting before they’re truly AI-ready.

Here’s exactly what AI-ready means, along with an enterprise buyers’ checklist for purchasing AI-ready compute.

Why does AI need a different data centre set-up?

Before we get into what an AI-ready data centre is, why do we need them in the first place? AI workloads have different requirements than traditional IT workloads. Firstly, they are much more energy-intensive, requiring higher-density compute (with more kWs per rack) to run. Using more power also means that servers running AI workloads give off more heat. Because of this, traditional air conditioning isn’t sufficient; AI-ready data centres must be equipped with direct-to-chip or immersion cooling systems.

What’s more, AI-ready equipment is simply heavier than traditional compute, so facilities need to account for this weight by reinforcing infrastructure. Meanwhile, inference workloads in fields such as automotive driving or robotic manufacturing require low latency and must be hosted in a data centre near your enterprise’s headquarters.

What is an AI-ready data centre?

An AI-ready data centre is a facility where enterprise compute customers can host their AI workloads. To be AI-ready, the data centre must provide a minimum density of 50 kW per rack, along with sufficient cooling equipment to cool down the servers.

What is the reality of retrofitting for AI

Some data centres claim they’re AI-ready. What they really mean: they’ll retrofit their facility so it’s suitable for AI workloads, but only when you ask. As an enterprise compute customer, watch out for these promises. The reality is, retrofitting for AI can take months, or even years. Power supplies may need to be upgraded to support higher-density racks. New cooling systems must be installed to handle the heat from high-density servers. The whole building may even need to be reinforced to cope with heavier equipment. A longer deployment timeline can cost your enterprise dearly, as you wait to scale your AI workloads while your competitors get ahead.

Where does the heat from AI workloads go?

We’ve already outlined that AI workloads produce extra heat because of higher-density servers and larger power requirements. Most facilities simply throw this heat away. However, when it’s captured and redistributed via district heating networks, the facility receives cooled water in return. In this way, donating heat reduces a data centre’s cooling requirements, making it more efficient and cost-effective for enterprise compute customers. AI-ready data centres that donate heat also earn their social license to operate by providing benefits for local people, ensuring your compute purchase is fully defensible.

What makes a data centre AI-ready? The ultimate checklist for enterprise compute customers

Power

AI workloads require higher power than traditional IT infrastructure.

Ask: What’s the density per rack?

Answer: Minimum of 20 kW per rack

Cooling

AI-ready servers can’t be cooled with air conditioning units alone.

Ask: What cooling system do you use?

Answer: An AI-ready data centre should be able to support a combination of  air cooling, direct liquid cooling (DLC), and immersion cooling technologies, with the solution tailored to the hardware, rack density, and site requirements.

Ask: Is the cooling system already fitted?

Answer: Yes.

Waiting for retrofits can cost your enterprise, so ensure the facility is already AI-ready and live, then check the deployment timeline.

Ask: What’s your speed to deployment?

Answer: 4 weeks or less.

Latency

If you’re running inference workloads, make sure the facility is situated nearby.

Ask: How far is the facility from my HQ?

Answer: 50 miles or less.

Heat donation

AI-ready data centres that reuse their excess heat are more efficient and cost-effective for enterprise compute customers.

Ask: How much of your facility’s heat is donated? (This is demonstrated by the metric: Energy Reuse Factor, or ERF.)

Answer: 75% or more.

Ensure your compute provider is truly AI-ready

‘AI-ready’ is more than just a slogan: it’s a checklist of specs that your enterprise must validate before signing a compute contract. Make sure your chosen facility can fulfil these requirements so you can deploy your AI workloads in as little as four weeks. Want to see an example of an AI-ready data centre? Check out our Urmston facility.