Trust · how we work
Methodology
We count a data centre as a campus: one site and operator, joined across planning registers, EPA licences, the valuation list, Building Control and OpenStreetMap (79 campuses, Oct 2026). Status needs evidence of operation, not announcements. Grid draw comes from EPA annual reports; rates are estimates from valuations. Anything we could not check is marked unchecked.
Updated 10 October 2026 · data build 2026-10-10 · how we know · corrections
How to read our numbers
Every number on this site is one of these kinds. They answer different questions with different certainty, so we label them where they appear.
| Label | What it means |
|---|---|
| Measured | Metered or recorded by the source for the period stated, such as the CSO's metered electricity or an EPA annual report. |
| Calculated | Worked out from measured figures, such as an average MW from a year's GWh. The inputs are shown beside it. |
| Our classification | Our reading of public records, such as whether a campus is operating, permitted or proposed. Not a live feed. |
| Reported | Stated by an agency or company, such as EirGrid's contracted capacity. Not independently measured. |
| Forecast | An agency's projection of future values. Earlier editions have over- and under-shot. |
| Scenario | What follows from stated assumptions. Not a prediction or a recommendation. |
| Illustration | Round, made-up numbers that show a mechanism. Not Irish data. |
| Editorial judgement | Our assessment, such as a risk score. Not a measured probability. |
| Machine reading | Claude (an AI model) read public documents with a published prompt. Every quote is checked word for word against the source; stance scores and summaries are the model's reading, not ours or the speaker's. Follow the links. |
Labels appear beside headline numbers, figures and sections. Hover or long-press a label for its meaning.
Counting data centres
Each OSM building, EPA licence, valuation-list property and planning application is a point with an operator brand where its own source names one. Same-operator points within 600 m are chained into one campus. A point with no operator joins the nearest operator's campus only if that operator is within 150 m and the next operator is 1.6x further, or within 400 m with no other operator within 800 m; otherwise it forms an 'operator not identified' campus. Status takes the strongest evidence: operating (EPA generators in use, valuation-list entry, completion certificate, pre-2014 permission with the operator's building on site, or an application describing an existing data centre) > under construction (commencement notice without completion certificate, or OSM buildings with no other proof) > permitted (live permission or EPA licence not commenced) > proposed (appeal pending, EPA application, undecided) > lapsed (lapsed, refused or withdrawn only). gridMWAvg2025 = grid electricity reported in each EPA licensee's 2025 AER divided by hours in the year (sites without a combustion licence report nothing).
Caveats
- A campus here is a cluster of records, not an operator's own site definition. Dense parks (Grange Castle, Profile Park, Clonshaugh, Blanchardstown) hold several operators side by side; unattributed records there are only attached when one operator is clearly nearest, so some appear as separate 'operator not identified' campuses.
- Operators are never invented. operatorConfidence 'high' = an official register names the brand (EPA licensee, planning applicant, valuation address); 'medium' = OSM tag/name or an inference from commencement-notice names made in the lifecycle pipeline; operatorBasis says which. Co-location attributions are flagged.
- gridMWAvg2025 exists only where the site holds an EPA combustion licence (>= 50 MWth of generators) and filed a 2025 AER. Smaller sites and Google's ratings (confidential) are absent, so operator totals are a floor.
- 'Operating' from a valuation-list entry or completion certificate proves a building, not its IT load or energisation. 'Operating (low)' means OSM shows buildings and no register record was found: usually older colocation or telecom sites that pre-date the planning text filter.
- Planning coverage is applications whose description says data centre / data hall / data storage facility, plus the applications the hidden-data-centre model found ('ICT facility', 'electrical rooms for electronic operations', substations and generators on known campuses), each placed by hand from its planning history (planningRefsFoundByModel). Their status comes from the register decision and BCMS notices only.
- Operators named through an SPV (e.g. 'Tunis Properties LLC' -> Amazon, 'Crag Wicklow' -> Echelon) use only high-confidence mappings from pipelines/entities, and operatorBasis names the mapping.
Rates
Downloaded the full Tailte Éireann valuation list for all 31 rating authorities via the open-data API and kept every property whose 'Uses' or any valuation-report floor use is DATA CENTRE. NAV is the list's 'Valuation' (net annual value, euro). Estimated annual rates = NAV x the council's 2026 Annual Rate on Valuation from Tailte Éireann's ARV table (parsed via pdf). Operator is taken from the address text where the list names it; Clonee is attributed to Meta from public site knowledge; the rest are 'Unattributed' because the list publishes no occupier.
Caveats
- The public valuation list has no occupier name, so most properties cannot be tied to an operator; brand attribution covers only those whose address names one.
- Matching relies on Tailte Éireann's 'DATA CENTRE' use code. Data centres valued under another use (e.g. office, warehouse, 'industrial' with server rooms) or entered as a generic industrial unit are missed; mixed properties where one floor is a data centre are counted at full NAV.
- Properties still under construction, not yet entered on the list, or under appeal are absent or at their pre-appeal value; newly completed halls can take 1-2 years to appear.
- Valuation dates differ by council, so NAVs are not on a common basis: Clare County Council 01/02/2022; Dublin City Council 07/04/2011; Fingal County Council 15/09/2017; Limerick City And County Council 01/03/2012; Louth County Council 15/09/2017; Meath County Council 15/09/2017; South Dublin County Council 30/10/2015 (property-level valuationDate, most common per council). Dublin City data-centre NAVs date from the 2011 valuation; newer councils' halls are at later tone-of-list dates. A national revaluation (Reval 2026+) is in progress.
- Cork City and County are pre-revaluation (old Rateable Valuation, no use field) and are excluded; Cork has few hyperscale sites.
- Estimated rates = NAV x 2026 ARV; actual bills can differ (part-year occupancy, vacancy, rates waivers/supports, base-year adjustments). Councils do not publish rates paid per ratepayer.
Grid capacity and draw
Read by hand from cached primary documents: An Coimisiún Pleanála inspector reports (data/raw/acp_reports), EPA LEAP licence documents (data/raw/epa_generation/docs), EirGrid / CRU / Oireachtas / CSO papers (data/raw/policy). Quotes are verbatim (<25 words); page = PDF page of the cached file (pdftotext form-feed index). build() re-checks every quote against the cache. contractedMVA is filled only where a document states a contracted or offered Maximum Import Capacity (MIC); any other figure is a stated demand (statedDemandMW), not a contract. Usable MW = MVA x 0.95 (assumed power factor). Utilisation = 2025 average draw / usable MW. Headroom = usable MW - 2025 average draw. statedDemandMW comparisons are labelled separately.
Caveats
- Per-site contracted MIC is almost never public. Only four sites have a stated MIC (one of them an offer, one a 0.5 MVA utility supply); everything else is either a stated demand from planning/EPA papers or unknown. Do not read statedDemandMW as a grid contract.
- Utilisation assumes power factor 0.95 (MW = MVA x 0.95). Data-centre power factors are typically 0.95-0.99, so this slightly understates usable MW and overstates utilisation by up to ~4%.
- 2025 draw is the annual average from EPA Annual Environmental Reports (electricity GJ / 31,536,000 s). Peak draw is higher than the average, so headroom against MIC at peak is smaller than the average-based figure.
- EPA AERs cover only sites with >= 50 MWth of licensed generators (12 campuses, 776 MW in 2025, ~89% of CSO's national data-centre total). Smaller and colocation sites are missing from the per-site view.
- EirGrid's national contracted totals (2,100 MVA transmission + 300 MVA distribution) include 'other new technology loads', not only data centres, and are not broken down by customer. The May 2026 EirGrid overview rounds the data-centre figure to 2 GW.
- The earlier summary of '~1,300 MVA connected/contracted and ~1,600 MVA applied' is from EirGrid's July 2020 DCCOPP consultation and is out of date: CRU/21/124 (Nov 2021) gives >1,800 MVA; GCS 2023-2032 gives ~2,000 + ~300 MVA; AIRAA 2026-2035 gives ~2,100 + ~300 MVA.
Water
Per-site water read by hand from Table 5 'Water Used' in each EPA Annual Environmental Report (AER) for the data-centre Industrial Emissions licences already tracked in data/raw/epa_generation (same documents, same pdftotext output; page = PDF page). National, hosepipe-ban and modelled figures from Oireachtas PQ answers (cached XML in data/raw/water/pq/), the Maynooth MUSSI working paper 31 (cached PDF), the Water Forum (2024, quoting the Government's 2022 data-centre statement) and The Journal Investigates (secondary; used only where no primary source is public).
Caveats
- AER water is what the operator reports for the licensed installation; Amazon sites' 2025 falls of 84-87% are not explained in the AERs.
- 2023 rows are mostly part-year (licences granted mid-2023). Equinix 2023 and Microsoft P1187 2023 include construction water.
- Only data centres with >=50 MW of on-site combustion plant hold EPA licences and file AERs; smaller sites are not covered (Uisce Éireann counts 57 on the network).
- litresPerMWh = public-supply water / grid electricity imported, both from the same AER. Electricity covers the whole licensed installation.
- The Uisce Éireann 57-site figures (4.78 ML/day; 2026 monthly) were released to a TD and are reported only by The Journal. No primary document is public, and the averaging period for 4.78 ML/day is not stated.
- Share of public supply uses 608,000 ML/yr from the Government's 2022 statement. About 38% of supply is lost to leakage (Water Forum 2024), so share of water actually used is higher.
Emissions
Indirect = data centres' share of metered electricity (CSO MEC02) x Electricity-sector emissions (EPA, SEC mapping) - an average, attributional allocation that handles imports and losses consistently. Cross-checks: CSO TWh x EPA generation intensity (direct CO2, per kWh generated; 2023-25 only) and x Ember intensity (lifecycle gCO2e, higher). Direct = stated annual CO2 from EPA inspector reports where given; otherwise rated gas thermal input x hours x 56.6 t/TJ. Measured direct = EU ETS verified emissions for the 27 Irish data-centre installations. Shares are against the July 2022 sectoral ceilings: Electricity for indirect; Buildings (Commercial and Public) for on-site combustion, which is where EPA inspectors say data-centre on-site combustion is counted.
Caveats
- P1202 (Crag Arklow Limited, 580.65 MWth) excluded: Withdrawn; Withdrawn Jan 2025 after the energy centre was relocated in planning; a new application is expected.
- P1203 (Vantage Data Centers DUB11 Limited, 508 MWth) excluded: Compliance Failure/Rejected; 2023 application the EPA could not consider (incomplete, Apr 2025); re-filed as P1227. MWth approximate sum (235 + 210 + 63).
- Sector placement of on-site combustion (Buildings - Commercial and Public) is the EPA inspectors' statement in licence reports (P1210 addendum p.41; P1186-02; P1181-02), not checked in the National Inventory Document.
- The EPA 2025-2030 projections describe the Commercial and Public sector as space and water heating; whether data-centre gas power is included is not stated (unchecked). If not, stated site emissions are on top of a sector already projected at 8.1 Mt vs a 5 Mt 2026-30 ceiling.
- Stated CO2 figures are licence-application estimates at expected operation, not measurements; actual output depends on when grid connections arrive. Measured on-site emissions so far are small (ETS verified ~9.5 kt in 2025).
- Ember intensity is lifecycle gCO2e and excludes imports; EPA and EirGrid figures are direct CO2 per kWh generated. The attributional method is the headline.
Who says what (machine reading)
Claude (Anthropic's model) read every document in the corpus with a fixed, published prompt and recorded, for each speaker: a stance on more data-centre development in Ireland (from −2, stop or pause, to +2, champion; or none when they took no position), up to four concerns from a fixed list, what they asked for, a one-sentence summary and a quote. Every quote was then checked word for word against the source text; quotes that could not be found were dropped. Counts, averages and concern shares are computed from those records, not by the model. A second prompt wrote each group's summary from its records only. Parliamentary questions are mostly factual, so most askers get no stance; their concerns still count. Party averages weight each speaker once. Ministers' answers count towards the government of the day, not their party.
Caveats
- Machine reading: summaries and stance scores are Claude's reading of the text, checked by quote, not by a person line by line. Treat them as a guide and follow the links.
- Coverage: parliamentary questions and debates are those whose heading mentions data centres (since 2015). Speeches on data centres under other headings (energy, climate, Leaders' Questions) are not yet included.
- Companies speak through the CRU consultation (2025) and committee hearings. Amazon, Microsoft, Google and Meta did not respond individually; their trade bodies (Cloud Infrastructure Ireland, Ibec, AmCham) did.
- Councils speak through their development plans and, for South Dublin, a CRU response. Councillors' motions are not yet read.
- Party is the member's party on the day they spoke (Oireachtas members data). Government terms use the dates the Taoiseach was elected (unchecked).
How we check claims
Each fact check names the claim, gives a verdict (true, mostly true, partly true, misleading, false or unknown), and links every figure it relies on to a source with a date. Where the evidence is a model rather than a measurement, we say so. When we find a mistake, we fix it and log it on the corrections page.
Rule from the project brief: only state checked facts, and mark anything from general knowledge as unchecked.