Two levels: the building and the computers
A data centre turns grid power into steady, cooled power for computers. Building measures say how well it does that and at what cost in water and heat. Computer measures say how much useful work the chips do with the power they get. EU reporting covers the first kind only; nobody is required to publish the second.
What one megawatt buys, by generation
The same megawatt at the meter runs about 3.3 times more AI arithmetic with B200 servers (2024) than with A100 servers (2020). Where grid power is the limit, newer chips are how a site gets more out of it.
Start with 1 MW drawn from the grid. At the Irish reporting sites' average PUE of 1.18, about 847 kW reaches the servers. Each NVIDIA server generation turns that into more peak arithmetic: 385 PFLOP/s per MW of A100 servers, 775 PFLOP/s per MW of H100 servers, 1,259 PFLOP/s per MW of B200 servers (dense BF16, from datasheets). Large training runs use only part of the peak: Meta reports 38–43% model FLOPs utilisation for Llama 3 on H100s, and the chart applies the midpoint to every generation.
EstimatePer-MW figures are dense peak ÷ maximum server power, our arithmetic; NVIDIA's headline figures are sparse, twice the dense ones. Applying one MFU to all three generations and ignoring networking and storage power make this an illustration, not a measurement of any site.
Sources: NVIDIA DGX A100 datasheet, 2020 · NVIDIA DGX H100 datasheet, 2023 · NVIDIA DGX B200 datasheet, 2024 · Llama Team, The Llama 3 Herd of Models, arXiv 2407.21783, Jul 2024 · DG ENER, first technical report on EU data-centre reporting, Jul 2025
The building: PUE, water and heat
Worldwide, the average PUE has been flat for six years at about 1.54 (Uptime Institute, 2025). Across the 770 EU sites that reported under the Energy Efficiency Directive, about 36% of those required to, it was 1.36. In Ireland only 18 of an estimated 123 sites reported in the first round, and the European Commission's technical report says Ireland "struggled to implement the necessary systems". Those 18 averaged 1.18: 1.14 for sites over 10 MW and 1.15 for enterprise sites, against 1.28 for colocation and 1.65 for small sites of 1–2 MW.
Irish data centres score well on PUE
The Irish sites that reported to the EU averaged a PUE of 1.18 (class B on the EU's new label), against 1.36 across the EU and 1.54 worldwide. PUE says how little power the building wastes, not whether it can host AI.
PUE divides a data centre's total energy by the energy reaching its computers. The EU's rating scheme, adopted in September 2026, puts it into classes from A (1.15 or less) to G (above 1.9). Only 18 of an estimated 123 Irish sites reported in the first round (2023 data), so the Irish average covers a minority of sites. Operators' own figures for their Irish sites are shown below the scale, each for the latest year published. At 1.18 against 1.54, the Irish sites waste about 23% less energy per unit of computing than the global average.
EstimateThe class of each figure is our reading of the EU bands; the "% less" comparison is our arithmetic.
Sources: European Commission, IP/26/1667 and C(2026) 3472 annexes, 21 Sep 2026 · Uptime Institute, Global Data Center Survey 2025, Jul 2025 · European Commission, COM(2026) 500, 21 Sep 2026 · DG ENER, first technical report on EU data-centre reporting, Jul 2025 · Amazon sustainability, AWS cloud, 6 Oct 2026 · Google data centres, efficiency, 6 Oct 2026 · Microsoft datacentre fact sheet, Ireland (uploaded Aug 2024), Aug 2024
On water, the Irish reporting sites used 0.64 litres per kWh of IT energy, against 0.58 across the EU (every licensed site's measured water). Their renewable energy factor was 0.99 (EU 0.86), but it counts certificates and power-purchase agreements over a year, so it says little about supply hour by hour. Ireland reported no heat-reuse figure; across the EU only 67 sites reported any. From 15 August 2027 each site's EU label, with its PUE and water classes, energy sources and heat-reuse readiness, becomes public; everything else stays confidential.
Utilisation is the measure Ireland can estimate itself: data centres drew about 38–46% of the grid capacity contracted to them in 2025, leaving 1.0–1.4 GW unused (our estimate).
The computers: chips, MFU and goodput
A chip's datasheet gives its peak arithmetic in FLOP/s, floating-point operations per second. Real training runs reach only part of it, because chips wait for data, for each other and for failures to be fixed. Model FLOPs utilisation (MFU) is the share of peak a run actually spends on the model. The published figures for large runs:
| Training run | Organisation | Hardware | MFU |
|---|---|---|---|
| PaLM 540B | 6,144 TPU v4 | 46.2% | |
| Llama 3 405B | Meta | up to 16,000 H100 | 38–43% |
| 175B model (MegaScale) | ByteDance | 12,288 GPUs | 55.2% |
| GPT-3 175B (as reported in the PaLM paper) | OpenAI | V100 | 21.3% |
Goodput goes wider. Google Cloud defines ML goodput as the product of scheduling goodput (share of time all resources are available), runtime goodput (useful training steps while they are) and program goodput (in effect, MFU). Failures matter at this scale: Llama 3 405B training saw 466 interruptions in 54 days (419 unexpected, about 78% of those hardware, 58.7% GPUs) and still achieved more than 90% effective training time.
For serving, the comparable measure is energy per answer. Median Gemini Apps text prompt: 0.24 Wh (Google Cloud blog). Includes idle capacity and data-centre overhead (PUE 1.09); 0.10 Wh counting active accelerators only. Typical GPT-4o query (estimate): about 0.3 Wh (Epoch AI). Rising to about 2.5 Wh with 10,000 input tokens. Tokens per second per MW, H100 vs B200: about 0.9 million vs 2.8 million (SemiAnalysis). Counting all-in provisioned power; an analyst benchmark.
None of these is published for any workload run in Ireland.
Can Irish halls host the newest chips?
Each NVIDIA generation has packed more power into a rack:
| Generation | Rack | Power | Cooling |
|---|---|---|---|
| A100 (2020) | 8-GPU server | 6.5 kW per server | Air |
| Hopper (2023) | 4 × DGX H100 servers (8 GPUs each) | 41 kW | Air |
| Blackwell (2025) | GB200 NVL72 (72 GPUs) | 132 kW | Liquid |
| Blackwell Ultra (2025) | GB300 NVL72 (72 GPUs) | 135 kW (peak 155) | Liquid |
| Vera Rubin (2026) | VR NVL72 (formerly NVL144), from H2 2026 | ~200 kW, estimate | Liquid |
| Rubin Ultra (2027) | NVL576 'Kyber' rack, announced for H2 2027 | ~600 kW, announced | Liquid |
Ireland's fleet was built before liquid-cooled AI racks
Every campus with measured 2025 grid draw and a completion record got its first certificate by 2022, before NVIDIA's first liquid-cooled rack-scale system shipped. Of 27 campuses whose files state a cooling method, 1 plans liquid-cooled racks.
The top panel places 23 of the 55 operating campuses at the year of their first completion certificate in the national Building Control register, which starts in 2014; the rest were finished earlier or have no certificate on record. Filled circles are sized by average grid draw in 2025 from EPA reports (726 of the 776 MW measured is at campuses first completed by 2022; the rest is at campuses with no completion certificate on record). Campuses add buildings over time, so a first completion date doesn't say how old every hall is. The bottom panel classifies the cooling method stated in each campus's planning or EPA licence files: 17 air: free air, evaporative or adiabatic, 9 chilled-water plant, 1 liquid-cooled racks; the liquid-cooled one is Amazon (AWS) Kildare Innovation Campus, Leixlip, building B1, in a 2026 licence application.
EstimateCooling classes are our grouping of the stated methods. Meta Clonee's "StatePoint Liquid Cooling" is an evaporative system that cools room air (Meta, 2018), so it counts as air-based.
Sources: Building Control Management System completion certificates, via StackPath directory, 10 Oct 2026 · EPA annual environmental reports (2025 grid draw), 10 Oct 2026 · Planning and EPA licence documents (cooling), 3 Oct 2026 · NVIDIA DGX A100 datasheet, 2020 · NVIDIA DGX H100 datasheet, 2023 · Uptime Institute Journal (D. Bizo), 25 Jun 2025 · Tom's Hardware, GTC 2025 report, 19 Mar 2025 · Meta Engineering, StatePoint Liquid Cooling, 5 Jun 2018
Bitpower's Q4 2024 market update counted 1,062 MW of IT capacity in "well over 100,000" racks in Ireland: at most about 10.6 kW per rack on average (our arithmetic). Bitpower's Q4 2024 update also reported 'zero announcements of AI-style campuses' and average draw around half of capacity.
Converting an air-cooled hall is more than adding pipes. A 63 A rack power strip at 230/400 V tops out at 43.5 kW; racks for dense IT should be rated above 1,800 kg static load; shallow raised floors and weak floor loading complicate retrofits. Above 40–50 kW per rack on average, power distribution becomes a bigger problem than cooling (Uptime's Daniel Bizo). Fewer than 4% of US data centres can support racks above 100 kW (Uptime's Daniel Bizo, estimate). Workarounds exist: meta and Microsoft run liquid-cooled GB200 racks in air-cooled halls using sidecar heat exchangers that turn room air into cooled liquid. Installing one HGX B200 system at CloudCIX/AlloComp in Cork, a rack close to a tonne, needed structural changes and a new access route into the building. (TechCentral, Dec 2025).
Why density matters more in Ireland
Our reading. Where grid connections are the scarcest input, compute per contracted megawatt is the lever. The same megawatt runs about 3.3 times more arithmetic with B200 servers than with A100s (Figure 1), and 1.0–1.4 GW contracted to data centres is unused. Using it needs halls that take dense, liquid-cooled racks and, since the CRU's ride-through decision, compliance before a site's demand can grow past its cap (the 900 MW limit).
What is not known
- 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.
- Irish sites' PUE, water and renewable figures one by one. EU reporting keeps them confidential; per-site rating labels are due from 15 Aug 2027.
- Model FLOPs utilisation, goodput or energy per query for any workload run in Ireland. No operator publishes them by country.