Executive Summary

• This article explores the key takeaways from the DCi Horizon’s panel session, Building the AI-ready data centre. The panel consisted of three distinct specialists [Associate Director of Gleeds, Jonathan Clark, Sales Director, Regional Strategic Accounts – Northern Europe, of Vertiv Giuseppe Forgione, and Samuel Premkumar (Manager, Strategic Growth & Operations Europe Service, Rosenberger OSI.
• Being AI-ready is a moving target; operators must explicitly define their compute profile or risk deploying wrong infrastructure.
• Grid connection delays of 6–8 years are forcing operators to rely on interim gas while the massive density of direct-to-chip liquid cooling is breaking legacy footprint layouts.
• With operators demanding rapid speed-to-market, bespoke builds aren’t viable for the operators who need to get operational yesterday. Manufactured infrastructure is becoming the new industry baseline.

 

During our virtual summit, DCi Horizons, we brought together three specialists from different disciplines to discuss strategies for building the AI-ready infrastructure in a panel session. We heard from Associate Director of Gleeds, Jonathan Clark, Sales Director, Regional Strategic Accounts – Northern Europe, of Vertiv Giuseppe Forgione, and Samuel Premkumar (Manager, Strategic Growth & Operations Europe Service, Rosenberger OSI, as they discussed the definition problem of the term and strategies needed to help infrastructure is indeed able to cope with the growth of AI and its power demand and how to future-ready solutions for tomorrow’s demand. Here are the key takeaways from the session.

The ambiguous definition of “AI-ready”

Being “AI-ready” is an ambiguous marketing buzzword you’ll see in the DC industry, but the panelists emphasised that a functional design will depend on the specific compute workload. Jonathan Clark states, “A lot of sites are saying they are AI-ready. The big thing I’m trying to explain… is understanding what that AI-ready model actually looks like. Is it GPUs-as-a-service? Is it training? Is it inference? Is it sovereign? Is it enterprise? All of those are slightly different.”

Having the right technology for investors requires knowing the technical requirements at the earliest stage to prevent Day 1 obsolescence.

Bridging the grid connection gap

Grid connection timelines are reaching up to six and eight years in key regions in Europe; this is a huge delay, especially for operators needing to get to market like yesterday. The delay in grid connections is forcing project managers to alter deployment schedules drastically, whilst also calculating the lowest common denominator of initial available power that can assist in getting their site on the ground and operational. The harsh reality is that a site that’s been designed for ultimate capacities of 500 MW may need to go live with as little as eight megawatts.

As a temporary solution, some operators are using on-site interim fixes just to go live, with gas generation highly favoured for the short term because it’s available, cheaper and faster to deploy – though that comes with the environmental debate of meeting our climate goals. It’s also unlikely that European regulations will accept gas as a long-term, permanent power solution, despite it being heavily used over the transatlantic in the US.

Retrofitting legacy white space

Many severe structural and geometric challenges are created when operators are retrofitting an existing facility to accommodate AI clusters. While traditional DC designs tend to scale infrastructure out across wide footprints, AI needs compaction of white space.

Giuseppe Forgione explains:

“When you had a 100-megawatt data centre, you needed a certain number of chillers or a certain amount of UPS systems. What we’ve got now is a huge compaction of the white space… but an enormous change in the requirement for the infrastructure that surrounds it.”

Legacy layouts have distributed compute with a generous amount of white space while AI high-density layouts need ultra-compacted white space with large secondary fluid loops and grey space, the latter being for integrating direct-to-chip liquid cooling systems. This increases surrounding grey-space requirements for all of the equipment required – from pumps and valves to specialised UPS technology. This needs to happen in the physical limits of existing infrastructure, which was built with air, perimeter or simple water cooling technology. Every legacy design is also different, so the project ends up becoming an expensive, customised engineering issue.

The death of bespoke builds?

There is a shift away from bespoke builds, particularly for operators who want to achieve rapid speed-to-market. Giuseppe Forgione states: “Speed of deployment being the key… if you go for something bespoke, it’s going to take longer to deliver.”

The panel emphasised that the industry is quickly adopting pre-engineered reference models such as NVIDIA’s standardised pod architectures and Vertiv’s converged infrastructure solutions to eliminate design delays and compress deployment timelines.

Connectivity and cabling are the unsexy bottleneck

The industry is obsessing over power allocations and thermal technology, leaving the physical reality of the infrastructure overlooked – namely the data throughput and fibre routing. — Samuel Premkumar explains: “It is heavily underestimated and heavily understaffed… Just very practically opening up the roads, laying the fibre, making sure you have enough and account for that future capability is something which is not factored in enough yet.”

The panel noted that while fibre cables containing over 13,000 individual fibres aren’t the bottleneck from the perspective of connections and manufacturing, it presents as an operational hurdle, since AI training relies on point-to-point spine and leaf architectures linking thousands of GPUs; from day one, there must be meticulous cable plans to address this. Some operators are using the grid connection delays to perform rigorous due diligence on their networking infrastructure.

Monetising heat reuse: compliance vs revenue

Liquid cooling is far more efficient at removing waste thermal energy than traditional air cooling systems and from the engineering perspective, this is rather a large efficiency breakthrough. Giuseppe Forgione explores this more: “We know that we get a higher grade of heat back from the direct-to-chip cooling solutions. Therefore, to reuse that heat, we don’t need to use as many heat pumps. We don’t need to spend as much money or energy upgrading that heat so that it can be used.”

Whilst heat pump tech and heat exchanges are readily available, the viability of heat recovery financially remains dependent on the infrastructure location. If the infrastructure lacks neighbours who are viable consumers of thermal energy, then there is no reason to implement the requirement to capture waste heat. Currently, heat reuse is an ESG compliance that is tricky to monetise; there must be a shift in the industry where collaboration is essential. The panel referenced successfully localised examples such as Digital Realty’s communicable district heating integration in Vienna and boutique wood-constructed 5MW facilities.

Future-ready for tomorrow

The overriding theme of the panel session from all three lenses is the necessity of early communication and collaboration among investors, suppliers, design teams and project managers. The hardware sector is moving faster than the infrastructure, with operators already planning for rapid transitions from current-gen to next-gen architecture like NVIDIA’s Rubin, Rubin Vera and Feynman platforms.

To avoid planning to build a facility that requires an immediate retrofit the moment it’s finally operational, flexible, modularised frameworks that treat power, scaling, connectivity, and cooling as flexible, scalable blocks should be considered.