WHAT DOES AI REALLY MEAN FOR DATA CENTRE DESIGN?

Article contents

Key Insights:

• GPU-led workloads concentrate power and heat into smaller footprints, creating tighter dependencies between power, cooling, hall design and grid strategy that must be addressed as one connected system.
• Higher rack densities, growing energy demand and the move towards liquid cooling are forcing organisations to rethink how facilities are powered, cooled and scaled from the outset.
• Grid capacity, renewable energy access, water availability and environmental impact all play a larger role in determining what can realistically be built and operated over the long term.
• Facilities designed around today's GPU requirements risk becoming constrained within a few years, making adaptable power, cooling and structural infrastructure essential for long-term viability.

AI has changed the data centre design brief

There’s a version of the AI data centre story that goes like this: AI is growing fast, so data centres need to get bigger. More servers, more power, more space. That version is incomplete - and it can lead to early design decisions built on the wrong assumptions.

AI increases demand, but it also changes what that demand actually looks like across every layer of the engineering brief.

The assumptions behind decades of data centre design - how power moves through a building, how heat is managed, how a hall is laid out, how a site connects to the grid - are all being tested by the arrival of GPU-led AI workloads at scale.

At RED, we’re working through these questions with clients across multiple continents.

What we see consistently is that organisations that treat AI data centres as scaled-up versions of what came before tend to run into problems that are expensive to fix later. The ones that do better take a step back and rethink the design fundamentals from the start.

This article sets out what that rethink involves and why AI-ready design depends on seeing the data centre as one connected engineering system from the start.

What do AI workloads change inside the facility?

The move from traditional to AI workloads is well documented - we’ve covered it ourselves in detail in our guide to AI data centres vs traditional data centres. But what does that change actually mean for the design of the facility?

The core issue is how power and heat are now being packed into much smaller areas. Traditional data centre loads are distributed relatively evenly across a facility - predictable and easier to manage. AI workloads do the opposite. 

GPU-led training clusters pull extreme amounts of power into dense, compact footprints, generating concentrated heat loads that change how the building has to be designed and operated.

As demand becomes more concentrated, the interaction between power, cooling, layout and connectivity becomes more critical to overall design, with decisions in one area quickly affecting another:

• Cooling requirements affect how the building is structured and how space is used
• Power architecture limits the cooling systems that can be used
• Grid capacity and connection strategy determine what can feasibly be built on site

In a traditional data centre, these systems can be designed separately. In an AI data centre, they have to be designed together - because getting one wrong has consequences that run through everything else.

The design fundamentals being changed by AI

AI data centre demand is increasing rapidly, but the more pressing issue is how that demand changes the engineering requirements behind the facility. From power and cooling to environmental impact and future flexibility, AI is transforming the design fundamentals that determine whether a data centre can perform effectively over the long term.

Grid and power strategy

Power has always shaped where and how data centres are built; AI raises the stakes. High-density GPU workloads require significantly more energy than traditional compute environments - according to the IEA's April 2026 report, electricity consumption from AI-driven data centres alone is forecast to quadruple by 2030.

That scale of demand raises questions that go beyond on-site infrastructure: Can a site secure the grid capacity it needs, and within a viable timeframe? What is the renewable energy mix available locally, and how does that affect both carbon performance and long-term operating costs? 

Power strategy has to be part of the site selection and design process from the outset, not something resolved after the facility is committed.

We look at how sites are actually navigating those constraints, and what happens when grid capacity becomes the limiting factor, in our article Can the grid keep up with AI data centres?

Rack density and hall design

AI hardware places new demands on the physical design of the data hall. GPU racks are heavier than traditional server equipment, increasing floor loading requirements and placing greater demands on structural design. Higher rack densities also create more complex cabling challenges, both in terms of routing and ongoing management.

There's also a longer-term consideration. Data halls designed today will be tested in the future, as the next GPU generation is likely to push power and weight requirements even higher. That makes flexibility an important design consideration from the outset.

Cooling strategy

Air cooling works for traditional data centre environments, but at the power densities that GPU clusters generate, it begins to reach its limits. Once you move towards racks in the 100 kW range, removing heat effectively requires a different approach - typically some form of liquid cooling such as direct-to-chip, rear-door heat exchangers or full immersion.

Each option brings its own infrastructure implications: pipework, leak detection, integration with wider building systems, compatibility with future hardware. Cooling strategy is a series of linked design choices that must reflect the workload, site conditions and operating model.

Water use and local impact

Advanced cooling strategies typically involve water, bringing additional considerations for design and delivery. AI workloads increase water demand both directly, through cooling systems, and indirectly through the water intensity of electricity generation.

In regions where water resources are already constrained, this can limit the cooling approaches that are practical and influence how proposals are assessed by planning authorities and local stakeholders. 

Water impact, therefore, needs to be considered early in the design process. That includes selecting sites and cooling technologies with local availability in mind, using closed-loop systems where appropriate, and defining clear targets for water efficiency across operations.

Power architecture

Getting power onto a site is one challenge - getting it to the rack reliably and efficiently is another entirely. AI data centres place extreme demands on power distribution infrastructure, from the substation through to the busbar, and the decisions made at this level have direct implications for performance and future scalability.

The debate around AC versus DC distribution, and the use of higher-voltage systems to reduce conversion losses, is becoming increasingly central to AI data centre design. They also carry long-term implications, as power architecture is not something that can be easily changed once a facility is built.

Environmental impact

Energy use and water consumption sit at the centre of discussions around AI data centres environmental impact, but they only capture part of the picture.

Embodied carbon in construction and equipment, carbon reporting obligations, community expectations and the scrutiny that large-scale infrastructure attracts all influence how facilities are designed and delivered.

Sustainability in the construction industry used to be a case of setting a few PUE targets and calling it done - green credentials sorted. For AI data centres, the conversation now starts much earlier in the design process.

The most credible facilities are now designed with lifecycle carbon in mind from the outset, accounting for operational energy, supply chains, grid carbon intensity and the long-term environmental profile of the facility.

Energy consumption

Power demand and energy consumption are related, but they are not the same thing. AI data centres, the question of how much energy is being used, and how that can be managed without affecting performance, sits across workload design, cooling, power architecture and infrastructure planning.

As AI workloads become more diverse and more demanding, energy consumption is becoming an increasingly important design input - rather than something tracked in operation.

Understanding how compute density drives cooling requirements, and how both feed into overall energy use, is key to designing facilities that remain viable over time.

Future flexibility

Perhaps the most important design consideration of all is the one that is easiest to underestimate: the rate of change in AI hardware.

GPU architectures are evolving at pace, and while rack power requirements can feel high today, in three to five years they could be baseline. That creates a clear risk: if a facility is locked too tightly around current hardware assumptions, it can lose capacity limits long before its intended lifespan is reached.

Designing flexibility into power capacity, cooling infrastructure and structural loading is what keeps a facility relevant long-term.

Six questions every AI-ready design needs to answer

Across all of the themes above, the design challenge ultimately comes back to a consistent set of questions.

These questions are tightly linked. The answer to the cooling question affects the water question, the answer to the power question affects the environmental question - and so on. That interdependency is exactly why AI-ready design has to be treated as a connected system, rather than separate workstreams.

For organisations working through these questions now, our piece on optimising data centre design in the age of AI and our broader overview of AI in data centre design help frame the wider context behind these questions.

From design principles to AI factories

The nine themes in this series cover the engineering fundamentals, but there’s one step we’re yet to mention: translating those fundamentals into the design of purpose-built AI factories - the next generation of hyperscale infrastructure built specifically to run AI workloads at the highest densities and the greatest scale.

AI factories are a different design brief again compared with hybrid facilities or retrofitted campuses. They require every decision across grid strategy, hall layout, cooling architecture, power distribution and sustainability to be made as a single integrated system from the outset, rather than as separate elements of design work.

Our existing work on AI factories and HPC facilities explores what that looks like in practice.

AI-ready design has to be integrated from day one

In AI data centre design, everything is connected - decisions around power affect cooling, cooling affects water use, site selection affects power strategy. Every design choice has knock-on effects elsewhere.

The facilities that will serve AI workloads well are built by making those decisions together, with a clear view of how each one affects the others. That requires a different approach to the design process, not just different technical specifications.

At RED, that integrated approach is how we work - across hyperscale campuses, hybrid facilities and everything in between. The complexity of AI data centre design is real, but the right expertise, applied early enough, makes it manageable.

If you’re planning an AI data centre, we can help translate the engineering requirements into a clearer, more workable design brief. Get in touch with our team to move it from concept to design reality.