Three kinds of data centre
Most of Ireland's data centres were built for cloud computing: storing data and running websites, apps and business software for customers across Europe. AI adds two jobs with different needs. Training builds a model by running thousands of chips together for weeks. Serving, also called inference, uses a finished model to answer people's requests.
| Cloud and colocation | AI serving | AI training | |
|---|---|---|---|
| Power per rack | About 8–9 kW on average | 40–135 kW on current GPU racks | 120–155 kW today; 600 kW announced for 2027 |
| Cooling | Air (free air, evaporative or chillers) | Air for older GPU racks, liquid for current ones | Liquid to the chip; air works only below ~50 kW a rack |
| Size of one site | Tens of MW | Tens to hundreds of MW | Hundreds of MW to over 1 GW |
| Network inside the site | Ordinary data-centre networking | Fast links between GPUs in a rack | Every GPU linked: 1.8 TB/s inside a rack, 800 Gb/s between racks |
| Needs to be near users | Yes: services, storage, interconnection | Somewhat: answers stream to users, but budgets are hundreds of milliseconds | No: data stays inside the cluster |
| Where data must stay | Customer's chosen region | Often 'in the EU', not one country | Anywhere the operator chooses |
| How power use behaves | Steady | Steady | Swings of tens to hundreds of MW within seconds |
Sources: Uptime Institute, Global Data Center Survey 2025; NVIDIA DGX SuperPOD data centre design guide; Uptime Institute Journal; NVIDIA, GB200 NVL72 product page; IEA, Energy and AI; MLCommons, MLPerf Inference v5.0 LLM benchmarks; Microsoft Learn, Azure model region availability; Choukse et al.; NERC, Characteristics and risks of emerging large loads.
Power per rack is up more than tenfold
A rack is a cabinet of servers about two metres tall. In a typical data centre it draws almost 9 kW, and in 2025 more than 80% of operators had no rack above 30 kW (Uptime Institute). A rack of NVIDIA's current Blackwell chips draws about 132 kW and weighs about 1.4 tonnes. Air can't carry that much heat away, so coolant is piped to cold plates on the chips. Uptime puts the practical limit of air cooling at about 50 kW a rack; Schneider Electric's guidance is lower, about 20 kW.
Power per rack: ordinary racks against AI racks
A current AI rack draws about 132 kW, 15 times a typical rack and at least 12 times the average IT capacity per rack in Ireland's fleet. Above about 50 kW, a rack needs liquid cooling.
The bars compare how much power one rack of servers draws. A typical rack worldwide averaged almost 9 kW in 2025 (Uptime Institute survey). Bitpower counts 1,062 MW of IT capacity in "well over 100,000" racks in Ireland, so at most about 10.6 kW per rack on average. NVIDIA's AI racks went from about 40.8 kW for four air-cooled H100 servers to 132 kW for a liquid-cooled 72-GPU Blackwell rack, and NVIDIA has announced racks of about 600 kW for 2027. The key shows which racks are air-cooled, which are liquid-cooled, and which figures are estimates or announcements.
EstimateIreland's per-rack figure is our arithmetic (IT capacity ÷ racks), a design capacity, not measured draw. Vera Rubin: an analyst estimate; NVIDIA has published no rack power. Rubin Ultra: as announced on stage in 2025, not on an NVIDIA page.
Sources: Uptime Institute, Global Data Center Survey 2025, Jul 2025 · Bitpower (Skreen Engineering), Ireland Q4 2024 market update, 2025 · NVIDIA DGX SuperPOD data centre design guide (DGX H100), DG-11301-001 v4, May 2023 · Uptime Institute Journal (D. Bizo), 25 Jun 2025 · Lenovo Press LP2357, NVIDIA GB300 NVL72, 30 Aug 2026 · NVIDIA newsroom, Vera Rubin platform (no rack power given), 16 Mar 2026 · Tom's Hardware, GTC 2025 report, 19 Mar 2025 · Uptime Institute Journal (D. Bizo), 'Density choices for AI training are increasingly complex', 25 Jun 2025
The next steps are bigger again. GB300 racks draw 135 kW, peaking at 155 kW, and NVIDIA's CEO has described racks of about 600 kW for 2027. NVIDIA plans to move to 800-volt direct current for racks of up to 1 MW, because at today's voltages a 1 MW rack would need up to 200 kg of copper busbar (NVIDIA developer blog, 800 V HVDC architecture).
Why AI training can go almost anywhere
Inside a training cluster every chip exchanges data with every other, constantly: NVIDIA's 72-GPU rack links each GPU at 1.8 TB/s, and racks connect at 800 Gb/s per GPU. That traffic stays inside the building. Users never touch it, so a training site doesn't need to be near them.
"AI training is less latency-sensitive than traditional data centre workloads" (IEA, Energy and AI, Apr 2025)
Serving is more tied to users, but less than it sounds. The MLPerf benchmark's interactive test allows 450 milliseconds for the first word of an answer and 40 ms for each one after. Light in fibre adds about 5 ms per 1,000 km (our arithmetic from Epoch AI's figures). And EU data-residency offers promise processing "within any EU member nation" (Microsoft), not in a particular country.
So the largest AI sites go where big blocks of power are cheap and quick to get:
| Site | Where | MW | Reason given |
|---|---|---|---|
| Stargate Abilene (Crusoe, Oracle, OpenAI) | Texas, US | 1,200 | Local renewable power; 'low-cost green energy' |
| xAI Colossus 2 | Memphis, US | 946 | Not stated |
| Meta Prometheus | Ohio, US | 1,000 | Not stated |
| Start Campus, Sines | Portugal | 1,200 | 100% renewable power, seawater cooling |
| Stargate Norway (Nscale, Aker, OpenAI) | Narvik, Norway | 520 | Hydropower, cool climate, 'industrial readiness' |
| Fluidstack, France | France | 1,000 | Nuclear power. Withdrawal reported as due to a US focus and difficulty securing power (withdrawn) |
| Stargate UK | United Kingdom | — | Paused until 'regulation and the cost of energy' allow (paused) |
MW as announced; "IT power" where stated. Abilene, Colossus 2 and Prometheus are in the US, shown for scale.
Not every plan holds. Fluidstack pulled out of its 1 GW French project in March 2026, reportedly partly over securing power, and OpenAI paused Stargate UK in April 2026 until "regulation and the cost of energy" allow.
Where Ireland stands
What an AI data centre needs, and what Ireland offers
Ireland is strong where cloud and AI serving need it (connected capacity, networks, users) and weak where large-scale training needs it: 2 binding limits and 3 weaknesses, mostly about power.
Binding Large blocks of power, quickly
Matters most for: Training
AI needsSingle AI sites now run to hundreds of MW and beyond 1 GW (Sines, Portugal: 1,200 MW planned).
Ireland todayDublin is fully constrained; EirGrid reported one new data-centre contract under the 2021 criteria by mid-2025. The largest campus, Microsoft Grange Castle, averaged 154.5 MW in 2025.
Weakness Cheap power
Matters most for: Training
AI needsTraining can run anywhere, so it follows cheap power (Narvik: hydropower, cool climate, 'industrial readiness').
Ireland todayDay-ahead power averaged €114/MWh in 2025, against Norway €48, Sweden €43, Finland €41 and France €61.
Binding A grid that can absorb big swings
Matters most for: Training
AI needsTraining jobs swing by tens to hundreds of MW within seconds (Microsoft, OpenAI and NVIDIA, 2025).
Ireland todayThe grid can lose at most 900 MW of demand in one fault, and data centres already fill that limit (EirGrid, June 2026).
Weakness Dense, liquid-cooled halls
Matters most for: Both
AI needsCurrent AI racks draw 132–155 kW and need liquid cooling to the chip.
Ireland today1 of 27 campuses whose files state a cooling method plan liquid-cooled racks (Amazon (AWS) Kildare Innovation Campus, Leixlip, building B1). Average IT capacity per rack is at most about 10.6 kW (our arithmetic on Bitpower's figures).
Strength Grid capacity already connected
Matters most for: Serving
AI needsThe fastest way to add AI is to densify sites that already have power.
Ireland todayAbout 1.0–1.4 GW of grid capacity contracted to data centres was unused in 2025 (our estimate). New load needs fault ride-through compliance first.
Strength Close to users and networks
Matters most for: Serving
AI needsServing answers and cloud services benefits from short routes to users and many networks.
Ireland todayNine subsea cables to Britain and direct links to the US; INEX, the internet exchange, with 128 members; the EU headquarters of most US tech firms.
Neutral Data kept in the EU
Matters most for: Serving
AI needsEU data-residency offers process AI requests "within any EU member nation" (Microsoft), not in one country.
Ireland todayIreland qualifies, and so does every other member state, so it competes with all of them for EU serving.
Neutral A cool climate
Matters most for: Both
AI needsMatters less than it did: liquid-cooled racks run on warm water (GB300 accepts inlet water up to 50°C).
Ireland todayMild all year; Microsoft says its Dublin sites need water for cooling less than 5% of the time.
Weakness Public AI compute
Matters most for: Both
AI needsEU AI Factories give researchers and firms shared AI supercomputers: 19 in 16 countries.
Ireland todayNone hosted; Ireland has an AI Factory Antenna linked to France and Luxembourg.
Each row sets a requirement of AI data centres, from public sources, against Ireland's position from this site's data. "Training" means building AI models, which can happen almost anywhere; "serving" (inference) means answering users' requests, which benefits from being near users and data. A requirement can matter for one and not the other, which is why Ireland can be a weak place for training and a reasonable one for serving at the same time.
EstimateThe verdict column is our judgement. Unused contracted capacity and per-rack IT capacity are our estimates.
Sources: Ember wholesale price data, 3 Oct 2026 · EirGrid via CRU202704 (load-rejection limit), 9 Oct 2026 · CRU2025236 (EirGrid contracts under the 2021 criteria), 12 Dec 2025 · Planning and EPA licence documents (cooling), 3 Oct 2026 · EuroHPC AI Factory selections, 10 Oct 2026 · PwC for the Department of the Environment, Climate and Communications, May 2024 · INEX, Oct 2026 · Choukse et al. (Microsoft, OpenAI, NVIDIA), Power stabilization for AI training datacenters, arXiv 2508.14318, 20 Aug 2025 · Microsoft Learn, Azure model region availability, 3 Sep 2026
AI hardware is arriving anyway. Ireland imported €19.7bn of servers, accelerator boards and network gear in the 12 months to Jul 2026, €4.8bn net of re-exports (more on server imports). CloudCIX and AlloComp's 'Boole' cluster in Cork runs liquid-cooled B200 GPUs (launched Jan 2026). GPU count and power were not disclosed. Equinix began building DB7x in Blanchardstown ($92m), aimed at AI workloads and run on power already allocated to the site, with 'no additional grid power'. It is due to open in early 2028. Pure DC and AVK are building a 110 MW gas microgrid in west Dublin aimed at AI, with no grid connection; the first two energy centres are due by end-2026. No public announcement of a GB200- or GB300-class cluster in Ireland by AWS, Microsoft or Meta was found.
What this means
Our reading. Ireland's data-centre economy was built for the cloud era: networks, users, EU headquarters and a mild climate. Those still suit serving AI to European users, especially inside the 1.0–1.4 GW of grid capacity already contracted to data centres but unused. Large-scale training values different things: big, cheap, quickly available power and a grid that can absorb sudden swings. On those Ireland is weak or blocked today, so training at scale is likely to keep going elsewhere unless power availability and price change.
That is not automatically a loss. Training sites are few, very large and often far from where the money is made; serving grows with use. Which kind of AI capacity Ireland should want, and at what cost in power and gas, is a separate question from where it can compete.
What is not known
- How much of the computing in Irish data centres is AI. No Irish or EU body publishes it; EU reporting data is confidential per site.
- Rack densities in Irish halls. No planning or licence file we read states a design kW per rack.
- How many AI accelerators are installed in Ireland. Import statistics show hardware arriving, not where it is installed or used.