Key insights
- AI workloads are pushing data centre design beyond traditional assumptions for power, cooling and resilience.
- Real data centre optimisation comes from modelling power, cooling, heat rejection, cost and carbon as one connected system.
- Liquid cooling is becoming essential as AI racks reach 100kW and beyond, but it needs to be designed around the whole facility.
- Future-ready AI facilities need climate-resilient heat rejection, flexible power strategies and scenario modelling to avoid overbuilding or underpreparing.
Why traditional data centre design is reaching its limits
AI data centre cooling and power demands have outpaced the design principles most facilities were built around.
Mohammad Royapoor explores why the assumptions that guided a generation of data centre design are no longer fit for purpose - and what it actually takes to build facilities that can handle what AI workloads demand.
Optimising data centre design in the age of AI
AI is changing what data centres need to do, and in turn what it takes to design them well. High-performance racks built for AI workloads now draw 100kW and beyond - several times the density of a traditional enterprise and mission-critical hall. At the same time, the chips inside them perform best at lower temperatures - so operators are being asked to push more power into a rack while keeping the silicon cooler than before.
Those two pressures, more power and tighter thermal limits, are why air cooling alone is reaching its practical limits. They also explain why AI infrastructure is forcing a wider rethink of how facilities are designed, powered and cooled.
The answer isn't any single new technology, it's a new method: treating the whole facility as one system to be optimised, rather than a collection of parts to be engineered separately.
What optimisation means in data centre design
Data centre optimisation comes down to trade-offs. Every decision is balancing performance against energy use, carbon, cost and cooling demand, while working within the constraints you can't change, and the variables you can.
In practice, that means choosing your rack densities, cooling topology and power architecture to hit a target - whether that's the lowest energy use, the lowest carbon, the lowest cost of cooling, or a balance of all three.
These constraints pull against each other as soon as you start designing around them:
- Thermal limits sit within ASHRAE classes, from tightly controlled conditions through to wider operating ranges.
- Efficiency is tracked using metrics like Power Usage Effectiveness and Water Usage Effectiveness, with embodied carbon also becoming part of how performance is judged.
- Climate plays a role too, with the same design behaving very differently in northern Europe compared to hotter regions like the Middle East.
The value of a genuine data centre optimisation approach is that it finds the overlaps between these things. You can optimise a condenser circuit on its own and gain something, but you're still working inside one energy vector. Model the whole facility together and you can balance improvements across cooling, power and heat rejection at once - and that’s where the biggest gains are found.
None of this is theoretical. RED has built the approach into a set of multi-vector master-templates spanning 5 to 50 MWe, each one a self-contained design that balances power, cooling and heat rejection together rather than in sequence.
To pressure-test them, the same 100 MW campus was modelled across four very different climates: London, Frankfurt, San Jose and Johor - so the design holds up whether the constraint is a cool maritime grid or tropical heat and humidity.
That’s the difference between claiming a facility is optimised and being able to show, climate by climate, exactly where the trade-offs land.
AI cooling and the move to liquid
For years, air did the job. The thermal resistance of a typical CPU or GPU sat around a tenth of a degree per watt, and conventional cooling kept everything in range. AI has broken that arrangement from both ends.
We're now asking chips to draw far more power, while also wanting to hold them cooler to keep them in their most efficient operating window. More heat to move, less room to move it.
The high-performance racks shipping today land at 100 kW and beyond, several times the density of a traditional hall. At those levels, air cooling simply runs out of road, which is why most of the industry has now accepted liquid cooling for high-density workloads.
The harder questions are the design and commercial ones:
- How does liquid cooling change the form factor of the data hall?
- Do you retrofit an existing facility or design for liquid from the outset?
- How do you scale without over-committing to infrastructure that hardware generations will outpace?
There's also a split developing in the market. Some hyperscale operators work directly with cooling manufacturers to integrate chip and cooling into a single standardised product. Colocation providers, particularly in emerging markets, often need independent advice on which approach actually fits their operating model.
Get it right and high-density liquid cooling can lower your PUE, cut operating cost and footprint, quieten the place down, and open up far better opportunities to reuse the heat you reject.
Heat rejection and climate resilience
Rejecting that heat is its own design problem - one that AI densities make harder.
Size your heat rejection plant on historical weather data alone, usually drawing on ASHRAE extreme-condition figures, the way much of the industry still does, and you're designing for a climate that no longer exists. Historical data doesn’t account for a warming climate, and dry coolers in particular are less resilient to extreme heat than hybrid coolers or cooling towers - so plants can end up undersized for the conditions they’ll actually face.
A more robust method adds a climate-change uplift to your design temperatures based on recognised warming scenarios, so the plant is sized for the conditions the facility will actually face over its lifecycle.
Layout matters just as much as sizing: pack heat-rejection units too close and hot exhaust gets pulled back into neighbouring kit, creating localised microclimates that push temperatures well above design assumptions.
This is exactly the kind of thing computational fluid dynamics (CFD) lets you catch before it's built and mitigate through spacing, gantries, blanking or higher-velocity exhausts - rather than discovering on a hot afternoon two years in.
The numbers back this up. In one modelled dry cooler array, units on the leeward side were seeing inlet air up to 18°C hotter than those facing into the wind, driven purely by recirculated exhaust. That was enough to push them beyond their design conditions. Spotting that in CFD meant it could be designed out with spacing and baffling before a single unit was installed.
Demand, resilience and the grid
The power side is moving just as fast. AI workloads make demand more variable and less predictable, which puts pressure on both supply and on-site generation - and it's pulling facilities into a much closer relationship with the electricity grid than most were designed around.
In some regions, data centres already account for a large slice of total demand, and grid operators are starting to look at whether those facilities can help balance the network rather than just draw from it.
This is a real opportunity, because when supply is tight, flexible power becomes extraordinarily valuable. In January 2025, with the grid short, NESO paid gas plants up to £5,000/MWh to run for a few hours - around fifty times the going rate. That's an extreme case, but it shows what flexibility is worth at the moment the grid needs it.
A facility designed with the right on-site generation and the ability to flex its load can step into markets like these, turning power infrastructure from a pure cost into a potential revenue stream.
Resilience is being optimised at the same time. Blanket facility-wide redundancy rules are giving way to distributed approaches that optimise resilience much closer to the rack - so you hold very high availability without paying for components you don't need. The aim is the same as everywhere else in the design: meet the requirement without over-provisioning.
If you want to learn more about the methods behind this approach, take a look at Designing Resilient, Future-proof Energy Systems for Modern Data Centres.
Designing for an uncertain future
The hardest part of designing for AI is the uncertainty. Nobody knows exactly how far rack power densities will increase, or how fast. Some clients want to be ready for the most extreme power densities AI might bring; others have conventional mission-critical workloads at the core of their business and want to hedge.
Data centre optimisation is what makes that uncertainty manageable. Modelling scenarios rather than committing to a single forecast lets a developer decide how much future-proofing to build in without overspending today.
The trade-off between stranded capacity (cooling and power paid for but never used) and underutilised white space only becomes visible when the whole system is modelled together. It's not a decision that can be made well any other way.
Designing this way also keeps the route to net-zero carbon open. Choosing energy and cooling approaches that can adapt alongside cleaner fuels and water-efficient heat rejection technologies means the facility doesn't design itself into a dead end.
The common thread across all of it is integration. The most efficient, resilient and sustainable facilities are the ones where cooling, power and heat rejection are optimised as one problem, against a clear objective and an honest set of constraints, not the ones where each system was perfected in isolation.
At RED, that integrated thinking runs through everything we do. We work with developers, operators and investors across multiple continents, on hyperscale campuses, colocation facilities and the next generation of AI infrastructure, helping them make better design decisions earlier - before they become difficult and expensive to revisit.
If you want to understand where the real gains are in your design - across cooling, power, heat rejection or all three - our team is ready to work through it with you. Get in touch with us today.
