Irish AI infrastructure boom: Why data centres are betting billions on liquid cooling
Heat is the enemy of every euro invested in Irish AI infrastructure. A modern AI server rack can draw 50 kilowatts of power. That is roughly the same as running 25 electric kettles continuously, every hour of every day, in a room packed with hundreds of identical racks. The cooling bill is not a footnote in the operating budget. It is the operating budget.
Why Ireland's Data Centre Boom Has a Physics Problem
Ireland hosts more data centre capacity per capita than any other country in Europe. The IDA's own figures put the Republic's share of European data traffic processing at somewhere between 25 and 30 percent. That position did not happen by accident. Tax policy, transatlantic fibre cables, a stable regulatory environment, and a decade of aggressive foreign direct investment built it. The problem is that the infrastructure model those centres were built on, forced-air cooling where giant fans push cold air across hot servers, was designed for a world before AI workloads existed.
A standard cloud server from 2015 drew around 5 to 10 kilowatts per rack. The Nvidia H100 GPU clusters that power large language models today draw 40 to 80 kilowatts per rack, with next-generation configurations pushing past 100 kilowatts. Air cooling, at any practical scale, stops working above roughly 30 kilowatts per rack. The physics are unforgiving. You cannot move enough cold air fast enough to keep those chips from throttling, and throttled chips mean wasted capital expenditure on hardware that is running at half its designed capacity.
This is not a marginal efficiency problem. It is the difference between a data centre that earns its construction cost and one that does not.
The Numbers That Force the Decision
Building a hyperscale data centre in Ireland costs between 800 million and 1.2 billion euro for a facility in the 100 megawatt range. That is a number worth making concrete. At the higher end, it is roughly the same as building 2,400 average Irish homes. The investors writing those cheques expect a return built on utilisation rates above 85 percent and power usage effectiveness ratios below 1.4, meaning for every unit of energy delivered to computing, no more than 0.4 units are wasted on cooling and other overhead.
Air-cooled facilities running modern AI workloads routinely breach a PUE of 1.6 or higher. At that level, a 100-megawatt campus wastes 60 megawatts on heat management alone. At Irish grid prices, that is millions of euro in annual operating cost that flows directly to the bottom line as loss. The hidden cost of AI agents compounds this further, as Irish companies renting compute time at inflated rates do not always realise they are subsidising the thermal inefficiency of ageing infrastructure.
Liquid cooling changes the equation. Directing chilled water or dielectric fluid directly to the chip surface removes heat 3,000 times more efficiently than air. PUE ratios for fully liquid-cooled facilities drop to 1.1 or below. That 60-megawatt waste figure shrinks to 10 megawatts. On a facility running for 20 years, the operating saving runs into hundreds of millions of euro.
How the Retrofit vs. Rebuild Decision Actually Works
Irish operators are facing a three-way choice, and the order in which they evaluate it matters.
Step 1: Audit the existing thermal envelope. Most Irish data centres built before 2020 were designed for air cooling. The raised floor plenum, the CRAC unit placement, the ceiling height, all of it was engineered around airflow. The first question is not whether to install liquid cooling but whether the building's structure can carry the weight of liquid distribution manifolds and whether the floor loading tolerates rear-door heat exchangers.
Step 2: Model the rack density trajectory. An operator taking a five-year view on AI workload demand needs to know what rack density they will be supporting in 2028, not 2024. If that trajectory goes above 40 kilowatts per rack, air cooling is not a bridge technology. It is a dead end.
Step 3: Choose the cooling architecture before choosing the hardware. This is where operators consistently get the sequencing wrong. They buy the GPU cluster first and then discover the cooling system cannot support it. The cooling architecture, whether direct liquid cooling to the chip, rear-door heat exchangers, or full immersion in dielectric fluid, determines which hardware configurations are even viable.
Step 4: Price the water supply and treatment infrastructure. Liquid cooling is not plug-and-play. It requires a reliable water supply, treatment systems to prevent corrosion and biological growth, and leak detection throughout the distribution loop. In an Irish context, proximity to municipal water infrastructure and planning permission for the associated civils work is a constraint that can add 12 to 18 months to a project timeline.
The Homegrown Opportunity
The story here is not just about global hyperscalers dropping billion-euro campuses in Meath and Kildare. It is about the tier-two and tier-three Irish operators, the colocation providers serving domestic enterprise and public sector clients, who face the same physics problem with a fraction of the capital resources.
A 10-megawatt Irish colo facility serving financial services clients, healthcare IT, and government workloads sits at exactly the inflection point where the upgrade decision becomes existential. Their enterprise customers are buying AI tooling. That tooling needs GPU compute. That GPU compute generates heat their facility was not built to manage. The operator who solves that problem first, who can credibly offer AI-ready rack space with guaranteed thermal performance, owns the next five years of that market. The operator who waits is competing with infrastructure that is already obsolete.
The fear is real: liquid cooling retrofits are expensive, disruptive, and require engineering expertise that is genuinely scarce in Ireland right now. The opportunity is equally real: the demand for AI-capable compute in Ireland is not speculative. It is contracted, it is growing, and the constraint on supply is thermal management, not planning permission or power availability.
What Comes Next
Ireland's position as Europe's data centre capital was built on being first to offer what global tech companies needed, stable regulation, competitive tax, and reliable connectivity. Maintaining that position through the AI infrastructure wave means being first to offer thermally capable compute at scale.
The operators who treat liquid cooling as a future consideration are making a producer's mistake: assuming the asset they built yesterday is still the product the market wants today. The ones moving now are building scarcity. AI-ready, liquid-cooled rack space in Ireland is a constrained commodity. Constrained commodities do not need to discount. They need to deliver.
The physics were never negotiable. The business model just caught up with them.
