Why AI labs and cloud vendors pick one country for data centres

AI labs and cloud vendors choose a country mainly by asking whether it can deliver large amounts of reliable, affordable electricity quickly, whether the law lets them build and serve customers there, and whether they can legally and safely import the chips. Tax, climate, fibre, talent and politics then break ties.

This post walks through each factor with named examples. Every figure is labelled by source and date, and gaps in the public record are flagged. Nothing here is investment advice, and no one factor decides a site on its own.

Key numbers at a glance

• US data centre share of electricity, 2023: about 4.4 percent, projected at roughly 6.7 to 12 percent by 2028. Source: US Department of Energy summary of a Lawrence Berkeley National Laboratory report, 20 December 2024.
• US data centre electricity use: 58 terawatt hours (TWh) in 2014 and 176 TWh in 2023. Same source and date.
• US computing share of commercial electricity: about 8 percent in 2024, projected to reach 20 percent by 2050 in the EIA reference case. EIA, 25 June 2025.
• Ireland, data centre share of metered electricity: 22 percent in 2024, 21 percent in 2023, 5 percent in 2015. Central Statistics Office, Ireland.
• Ireland, contracted data centre demand: 2 GW, mostly in greater Dublin, against an all island peak of about 7.5 GW; a further 6 GW of interest identified. EirGrid, 20 May 2026.
• PJM (a large US grid operator) summer peak growth: 3.6 percent a year on average over ten years, reaching 222,106 megawatts (MW) in 2036, up 65,733 MW. PJM load forecast, posted 14 January 2026.
• US industrial electricity price, July 2026: 9.77 cents per kilowatt hour on average. Examples from the same table: Texas 7.07, Washington 7.51, Oregon 8.92, Georgia 10.03, Virginia 10.07, Ohio 10.95. EIA preliminary monthly data.
• EU non-household electricity price, second half of 2025, for users of 500 to 2,000 megawatt hours a year: EU average 0.1837 euros per kilowatt hour; Ireland 0.2552 (highest in the EU); Germany 0.2264; Sweden 0.0970. Eurostat.
• US federal permitting order: Executive Order 14318, signed 23 July 2025, defines a covered data centre project as one needing more than 100 MW of new load.
• Anthropic US build: sites in Texas and New York, about 800 permanent and 2,400 construction jobs, online through 2026, per the company.
• European Union: 19 AI Factories and 13 antennas in 16 countries, and up to seven planned AI gigafactories of more than 100,000 advanced processors each. European Commission page, as of 19 October 2026.
• AWS European Sovereign Cloud: general availability announced 15 January 2026, first region in Brandenburg, Germany, with more than 7.8 billion euros planned in Germany. Company statement.

Factor 1: power price and power availability

A modern AI data centre is mostly a building full of electricity-hungry chips. Power is usually the largest running cost and the hardest thing to get in volume. A price per kilowatt hour matters, but availability matters more: a site with cheap power that cannot supply 300 MW for five years is no use to a buyer who needs the capacity next year.

The scale of the demand is why every country is now asked about it. The US Department of Energy, summarising a national laboratory report, put data centres at about 4.4 percent of US electricity in 2023 and projected 6.7 to 12 percent by 2028. The EIA, in a separate analysis, expects computing to grow from about 8 percent of commercial electricity use in 2024 to 20 percent by 2050, and says its reference case does not assume a rapid expansion of AI.

Prices differ a lot by place. In the EIA's July 2026 preliminary table, the US industrial average was 9.77 cents per kilowatt hour, with Texas at 7.07 and Ohio at 10.95. These are preliminary monthly figures, which the EIA revises. In Europe, Eurostat's second half of 2025 figures for mid-sized non-household users show a spread from 0.0970 euros per kilowatt hour in Sweden to 0.2552 in Ireland. Those bands are for users of 500 to 2,000 megawatt hours a year, far smaller than a large data centre, so they show direction, not what a hyperscale buyer pays. Large buyers often negotiate directly with generators through a power purchase agreement (PPA), a long contract to buy electricity at an agreed price.

Factor 2: how fast you can connect to the grid

An interconnection queue is the waiting list of projects asking a grid operator for a connection. Getting through it can take longer than building the data centre. Operators also have to decide which requests are real.

PJM's January 2026 forecast shows how seriously this is taken. It says near-term forecast years need "firm" commitments, such as signed service obligations or construction commitments, while longer-term projects are treated as "non-firm" and reduced because they are uncertain. It also says the 2026 forecast is lower than the 2025 one in the near term through 2032, partly because of those large load adjustments. In plain terms: a grid operator that sees far more requests than it can plausibly serve starts sorting them, and a buyer with a signed commitment moves ahead of one with a hope.

Ireland shows what a crowded grid looks like. The Central Statistics Office reports that data centres used 22 percent of metered electricity in 2024, up from 5 percent in 2015. That statistic alone explains why a country with an existing data centre cluster can become a hard place to add capacity.

The grid operator's own numbers show the scale. EirGrid's Data Centre Constrained Area Overview, published 20 May 2026, says the all island power system has a peak demand of about 7.5 GW and that Ireland has 2 GW of contracted data centres, mostly in the greater Dublin area. Once those contracts are fully used, data centres will represent nearly a third of Irish demand. A market exercise found interest in a further 6 GW, which EirGrid says would double the current system peak if it all materialised. The document says the available headroom "has largely been exhausted".

The regulator responded with rules rather than a simple ban. The Commission for Regulation of Utilities published its Large Energy Users Connection Policy (CRU/2025236) on 12 December 2025. EirGrid's summary says it requires new data centres to provide dispatchable generation or storage on site or nearby that matches their maximum import capacity, and to show whether they sit in a constrained or unconstrained part of the network. The law firm Philip Lee, in its analysis of the decision, reports a further requirement that sites of 1 MVA (megavolt ampere, close to a megawatt) and above match at least 80 percent of annual demand with additional renewable generation in Ireland, phased in over six years. EirGrid also lists Dublin specific problems: high short circuit currents that limit new generator connections in the greater Dublin region, and stability issues when data centres drop demand sharply during voltage dips. For a site buyer, the result is that Ireland is open, but mainly to projects that bring their own power and choose a location outside the constrained areas.

Because grid connection is slow, some developers look for sites that already have a large connection, such as old industrial land. The UK's January 2025 AI Opportunities Action Plan proposes AI Growth Zones for that reason, pointing to areas with existing energy capacity, and the UK government announced the first at Culham in Oxfordshire, citing its grid connection.

Factor 3: permitting and planning

Permits cover land use, environmental review, water, and connection of new power lines. A country that can approve a large site in months has an edge over one that needs years.

The US federal government addressed this in Executive Order 14318, signed 23 July 2025. It sets out to speed federal permitting for data centres and their power equipment. Its main tests for a "qualifying project" are at least 500 million dollars of committed capital, more than 100 MW of added load, a national security role, or designation by a cabinet secretary. It directs agencies to seek categorical exclusions under the National Environmental Policy Act, uses the FAST-41 dashboard process for tracking reviews, and sets a 180 day review target for certain Army Clean Water Act permits. It also lists federal land that may be offered.

Note what this order does not do: it covers federal permits only. State and local approvals, which often decide whether a project goes ahead, remain separate.

The UK plan takes a similar route. It suggests a streamlined planning process for AI Growth Zones and considers a special planning use class for AI data centres.

Factor 4: tax, incentives and public money

Governments compete with tax relief, land, and direct funding. The EU is an example of public money shaping location. Its Commission says 19 AI Factories and 13 antennas are being set up across 16 countries, and plans up to seven AI gigafactories with more than 100,000 advanced processors each. For the gigafactories it describes up to 10 billion euros of EU and national funding, expected to unlock at least 20 billion euros of private investment. Those are planned figures from the Commission, not completed spending.

Sometimes the incentive is about jobs and prestige as much as tax. In its own announcement of a 50 billion dollar US investment, Anthropic cited about 800 permanent and 2,400 construction jobs and said the plan supports the US AI Action Plan. The announcement does not set out any tax terms for those sites.

Factor 5: climate, cooling and water

Chips make heat, and removing it costs energy and sometimes water. Cool climates reduce the energy used for cooling. Dry climates may allow evaporative cooling, which uses water, while wet or humid places make that less efficient. Water figures and the difference between direct and indirect use are covered in our post on data centre water use. The practical point for site choice is general: a site's cooling design is chosen to fit its climate, and a local water shortage can become a permit problem even where power is plentiful.

Factor 6: fibre, latency and the type of workload

Training a model is a long batch job. It can run far from users, so power and space dominate the site choice. Serving answers to customers (inference) is different: speed and reliability matter, and that favours capacity near users and near good network links.

Network links also pull toward certain countries. Singapore is an example from Anthropic's 14 October 2026 announcement of an office there. The company said Singapore is one of its most active markets, ranking second of 121 countries for claude.ai usage relative to population. The page did not mention a Singapore data centre. It noted that regulated organisations there can run Claude with in-region inference on Amazon Bedrock. So the office, the customers and the processing location are three separate decisions.

Factor 7: data residency and sovereignty rules

Many customers must keep certain data in a country or region, or under that region's laws. Providers respond in two ways: they add regional processing, and they build separate sovereign offerings. Microsoft said in April 2025 that European datacentre capacity would grow 40 percent over two years across 16 countries, and that its EU Data Boundary was complete. AWS announced general availability of its European Sovereign Cloud on 15 January 2026, with a first region in Brandenburg, operated by EU residents under European law, according to its press release.

This factor has its own post: see data residency and sovereignty for AI, explained.

Factor 8: chip export controls

A country with cheap power is useless to a lab if the most advanced chips cannot legally be shipped there. The United States regulates exports of advanced computing chips, and in July 2026 it eased those rules for the United Arab Emirates. The detail is in our export controls explainer.

Factor 9: talent, geopolitics and trust

A data centre needs electricians, network engineers and operators, but an AI lab also needs researchers, and those cluster in a few cities. Offices and data centres are therefore often in different countries. Trust runs through everything: a government that signs an agreement on AI safety and security, as Singapore's ministry did with Anthropic on 14 October 2026, signals openness. A government that can compel access to data or is under sanctions signals risk. This factor is the hardest to measure and the easiest to overstate.

How the factors combine

No place wins on all factors. Texas has a low industrial price in the EIA table, but price is only one input and the queue for a connection is another. Ireland has an established data centre cluster, but data centres already use a large share of national metered electricity, and new sites now have to bring their own dispatchable power. The UAE has capital and, since July 2026, easier chip access under US rules. Read the pattern, not a ranking: buyers trade one scarce thing for another, and the trade is different for every project.

For more context on who is paying for this capacity, see Anthropic's compute spending explained and OpenAI's compute commitments and 26 gigawatts.

What is firm and what is uncertain

Firm: the DOE and EIA statistics, the PJM forecast text, the Irish statistics and EirGrid's constrained area document, the executive order text, the Commission's AI Factory page, and company statements about their own announcements. Each is a primary or near primary document.

Uncertain: electricity price tables change monthly and depend on customer class. Plans such as gigafactories and AI Growth Zones are intentions. Company job and investment numbers are company statements. Projections to 2028 or 2050 are projections, not outcomes. The 80 percent renewable figure for Ireland comes from a law firm's reading of the CRU decision rather than the decision paper itself.

What to watch next

• Whether more US states and grid operators adopt firm commitment tests like PJM's.
• EirGrid's heat maps of available demand capacity, which the CRU decision requires.
• Which AI Growth Zones and EU gigafactory sites are named.
• Whether sovereign cloud offerings add AI model hosting inside the region.
• Further BIS rules on chip exports.
• Updated EIA monthly prices; the figures above will age quickly.

Safeguard's SBOM and dependency scanning, with a free scan at https://gold.safeguard.sh/scan, shows which AI services your code depends on.

Quick check

Sources

• DOE releases new report evaluating increase in electricity demand from data centers, US Department of Energy, 20 December 2024.
• Computing is a growing share of commercial electricity use, EIA, 25 June 2025.
• EIA Electric Power Monthly, average price table, July 2026 preliminary data.
• Electricity price statistics, Eurostat, second half of 2025.
• Data centres metered electricity consumption 2024, Central Statistics Office, Ireland.
• PJM 2026 Load Forecast Report, posted 14 January 2026.
• Accelerating Federal Permitting of Data Center Infrastructure, Executive Order 14318, 23 July 2025.
• AI Opportunities Action Plan, UK government, January 2025.
• Prime Minister sets out blueprint to turbocharge AI, UK government, 13 January 2025.
• AI Factories, European Commission, as of 19 October 2026.
• Data Centre Constrained Area Overview, EirGrid, 20 May 2026.
• Re-balancing the digital bargain: Ireland's new CRU large energy user connection policy, Philip Lee, 22 December 2025.
• Anthropic invests 50 billion dollars in American AI infrastructure, Anthropic.
• Anthropic opens Singapore office and signs MOU, Anthropic, 14 October 2026.
• European digital commitments, Microsoft, 30 April 2025.
• AWS launches AWS European Sovereign Cloud, Amazon, 15 January 2026.
• Department of Commerce eases export controls for UAE, BIS, 10 July 2026.

Last reviewed: 19 October 2026.