Seven readings of the build-out, each built from public data and updated as it moves. Every one is read first in the US, the sharpest case, then extended to every major market.
Who pays for the build.
In America's largest grid, the cost of the build is now landing on the neighbours' bill. Read straight, that exposure is also the industry's strongest defence.
The fastest way to lose the public is to be seen raising the neighbours' electricity bill. In the largest US grid, that is now measurable, and it is a real exposure for the industry. But the honest reading of the data is also the industry's strongest defence: the pass-through is a feature of one market's design, not an inevitable consequence of AI. Read it straight, and it becomes a case you can win.
An eleven-fold rise in three years. The last two auctions cleared at the federal price cap; PJM estimates that without the cap agreed with Pennsylvania, the 2027/28 price would have landed near $530 / MW-day.
Keep the claim honest, or it breaks. The wholesale price only moves the supply component of a bill, which is 30 to 50% of the total, so a capacity spike does not multiply the whole bill, and a national fact-check rated the simplistic version mostly false. Data centres are not the sole driver either: an ageing grid, equipment costs and generator retirements all push the same way. The defensible line is narrow and strong: in a capacity market, data-centre demand is a majority of a real, measured increase.
Do not deny the cost. Fix who pays it. The market monitor's own remedy is to require large loads to bring or contract their own firm generation, so a campus's demand does not raise its neighbours' bills, an approach now echoed by a White House-backed ratepayer-protection pledge. Operators who move first, with transparent cost allocation and bring-your-own-power commitments, defuse the backlash. Those who wait for it to become a ballot question will inherit the rules others write.
PJM is one market's design. The same pressure surfaces wherever always-on demand concentrates, even where the mechanism is nothing like a capacity auction.
| Market | How the cost is set | Household exposure | The number that binds |
|---|---|---|---|
| PJM | Capacity auction clearing price | Direct, shared across 67m people | $9.3bn extra in a single year |
| Ireland | Gas sets the wholesale price more often | Cumulative, paid through the bill | 22% of national electricity |
| United Kingdom | Levies and network charges | Disputed, under active policy debate | 10+ yr connection waits compound it |
| Germany | Energy-only with capacity payments | Indirect, via price volatility | 4.26 GW load, highest in the EU |
| Middle East | State-directed pricing | Very low, by policy | $0.05 to 0.06 per kWh |
| China | State-directed, spot trading emerging | Managed through industrial policy | 2026 first data centre VPPs in spot markets |
Ireland is the sharpest non-US analogue. A study for Friends of the Earth Ireland estimated the average household paid an extra 360 euro across 2015 to 2023 because data-centre demand pushed gas to set the price more often, a cumulative 715 million euro, with a further 1.4 billion euro projected over the coming decade if expansion goes unregulated. The opposite case is the Gulf, where large-consumer tariffs of five to six cents are already pulling power-intensive workloads toward the region. Same physics, opposite bill.
Sources: Friends of the Earth Ireland and Beyond Fossil Fuels, The Cost of Data Centres, 2026; Ireland CSO; IEA Energy and AI 2025; TechUK position paper on electricity rates; PwC Middle East. Household figures are graded projected and vary by market design.
Sources: PJM Interconnection Base Residual Auction results, delivery years 2024/25 to 2027/28; Monitoring Analytics, PJM independent market monitor; NRDC and the Citizens Utility Board of Illinois; PolitiFact; SemiAnalysis. Capacity-price figures verified against PJM releases; household-cost figures are forward projections, graded projected.
Small in water, loud in power.
Against the sectors that move the meter, data centres barely register in water; their power draw is small but the fastest-growing thing on the grid.
The industry's instinct is to fight the water story. That is the wrong fight, fought the wrong way. Set against the sectors that actually move the meter, data centres barely register in water, while their energy footprint is small but the fastest-growing thing on the grid. The winning move is not to deny either number. It is to get the proportion right, in public, before someone else frames it.
Lead with the proportion. Never deny the local truth. Honesty about where you do bite is what makes the context believable.
Water is the wrong battlefield to fight on dishonestly. Nationally, data centres are a rounding error, dwarfed by power generation and farming. But around 40% sit in water-stressed basins, and most do not publish their consumption, so a blanket "we barely use water" claim collapses the moment a reporter stands in front of a stressed reservoir in Arizona. The credible claim is specific: small nationally, managed locally, measured everywhere.
Publish site-level water data, report consumption and not just withdrawal, and commit to closed-loop or air cooling in water-stressed sites. Then lead every public conversation with the proportion: energy first, water second, local before national. Operators who measure and disclose own the argument. Those who deflect hand it to their critics.
The proportion holds worldwide. What changes is the local stress point, and it is sharper in Asia and the Middle East than anywhere in the US.
| Where the power goes | 2024 demand | Share of the global total | Trajectory to 2030 |
|---|---|---|---|
| United States | 180 TWh | 44% | Largest single market |
| China | 102 TWh | 25% | 19% CAGR to a projected 289 TWh |
| EU big four | 41 TWh | 10% | Germany, France, UK, Netherlands combined |
| India | ~12 TWh | under 3% | Almost 5x to 57 TWh, 0.8% to 2.6% of the nation |
The fastest growth is outside the US. In the tropics the northern-hemisphere fix, air-side cooling, barely works, so Southeast Asian sites stay on evaporative or chilled water and a single 1 MW hall can draw more than 25 million litres a year. The credible claim travels intact: small as a share, sharp where it lands, and most acute exactly where it is least disclosed.
Sources: IEA Energy and AI 2025; Uptime Institute 2025; Planet Tracker via the ASEAN Guide for Sustainable Data Centre Development; CEEW India 2025; hyperscaler sustainability disclosures. Global shares verified; 2030 figures graded projected.
Sources: USGS Circular 1441, 2015 national water-use compilation; Lawrence Berkeley National Laboratory 2024 US Data Center Energy Use Report (DOE); WRI Aqueduct for water stress. Sector shares verified; the data-centre water share is an order-of-magnitude estimate and the 2028 energy figures are projected ranges, graded projected.
The gigawatts that will never connect.
Read the interconnection queue and a smaller number appears: the share actually energised. The gap between announced and real is what matters most.
Every week brings another headline gigawatt figure. Most of it is noise. Read the interconnection queue, the list of everyone asking to plug in, and a different number appears: the share actually energised and drawing power. In the largest data-centre market in the country, that share is tiny, and the gap between announced and real is the single most useful thing an investor or a regulator can know.
The phantom cuts both ways. The boom numbers in the press, the gigawatts "announced", are mostly queue entries that will never connect, so treat any announced figure as an option, not a build. But ERCOT still has to plan for a credible slice of it, and even a tenth of 226 GW would swamp the grid. The signal is not the headline total. It is the conversion rate, and who is actually posting financial commitments. And this is only the power side of the gap: whether the chips to fill these gigawatts are even landing is a separate question, mapped in Landfall.
For capital and operators: track energised and financially-committed gigawatts, not announcements, and discount any market's headline queue by its historical conversion rate. For policymakers: copy SB 6's duplicate-disclosure rule and make a queue position cost something, so speculative filings clear out and the real projects connect faster.
ERCOT's 226 GW is a Texas artefact. The speculative queue, where the headline number is an option market and not a build forecast, is now everywhere.
The 2.3% energised in Texas and the 13% lifetime conversion in the Berkeley Lab data point at the same diagnosis: the headline queue is not the pipeline. Europe has reached the same conclusion independently. With backlogs in at least 16 member states, the EU is shifting from first-come to first-ready, first-served, requiring permits, land control and financing before a slot is held, while Amsterdam, Dublin and congested parts of Denmark have paused new connections outright. It is SB 6's logic, making a queue position cost something, arrived at from the other side of the Atlantic.
Sources: CERRE, From Gridlock to Grid Asset, 2025; Lawrence Berkeley National Laboratory interconnection data, 2024; European Commission grid-access reform. Queue totals verified; conversion rates move and are re-checked at each staged release.
Sources: ERCOT System Planning and Weatherization Update and Large Load Interconnection Process Q&A, November to December 2025; ERCOT board materials; Latitude Media; Mercom; analysis by Dave Friedman. Queue and energised figures verified against ERCOT filings; the conversion rate moves monthly and is re-checked at each staged release.
How long until it draws power.
Capital has stopped asking what a site costs and started asking how soon it can be energised. In a constrained grid, the months between securing a site and drawing power decide the return, and they vary by years across markets. The fast routes now bypass the utility queue entirely.
These are indicative ranges, not quotes. Time-to-power is project-specific and shifts with every substation and policy change. But the pattern holds: the interconnection-queue average for a 2025 first-power project now exceeds 2,100 days, roughly six years, and behind-the-meter generation is the only route that reliably beats the queue. Treat any single market figure as a starting point to verify, not a promise.
Lock power before land. Secure an interconnection slot or an on-site generation plan first, then build to it. For capital, price each market by its speed to power, not its rent: a campus that energises two years sooner is worth more than one that is two dollars cheaper per square foot.
Switch the scope to redraw the same clock for the world's most-watched markets. The Gulf builds grid and digital infrastructure together, which is why it leads on greenfield speed; London, Dublin and Tokyo are fighting legacy grids never designed for gigawatt loads.
Sources: CBRE North America Data Center Trends, H1 and H2 2025; Enverus Intelligence Research, Time to Power 2026; CERRE 2025; Gulf Data Centre Association; PwC Middle East; operator and utility disclosures. Ranges are indicative and graded projected; verify per site at build.
The latency no one can beat.
A data centre can be the best in the world and still be in the wrong place. The speed of light through fibre sets a hard floor on how fast a user reaches a server, and that floor, not the brochure, decides which workloads a site can serve. Drag the distance, or pick a route, and watch what each kind of compute can and cannot do.
The physics is fixed at roughly five milliseconds per thousand kilometres. What varies is where the users are, where the cables land, and which workloads a route can carry.
| Route | Distance | Indicative round trip | What it can serve |
|---|---|---|---|
| NYC to Ashburn | ~330 km | ~8 ms | Every workload |
| London to Frankfurt | ~640 km | 15 to 20 ms | Every workload |
| NYC to London | ~5,600 km | ~70 ms | Interactive inference, not real-time edge |
| Mumbai to Singapore | ~3,900 km | 55 to 70 ms | Interactive inference |
| Johannesburg to London | ~9,000 km | 140 to 170 ms | Training and batch only |
| London to Singapore | ~10,800 km | 170 to 230 ms | Training and batch only |
Singapore is the reason its moratorium hurts: it carries sub-20 ms reach across 680 million people in ASEAN, so pushing capacity to Johor, Batam and Bangkok is a real latency trade-off, not just a land play. Africa sits the other way. Johannesburg to London clears 140 ms and Lagos to New York clears 160 ms, which places both outside real-time edge and at the margin of interactive inference. The build case there is sovereign cloud and batch processing until new subsea routes such as 2Africa and Equiano close the gap.
Sources: Atlas latency analysis; Equinix and Hibernia route data; Console Connect Africa Interconnection Report 2025. Round trips modelled from fibre-route latency and checked against known city pairs, graded projected.
Round-trip times modelled from real fibre-route latency, roughly 0.0125 ms per kilometre plus switching overhead, and checked against known city pairs. Indicative, graded projected. Training carries no user-latency requirement; its constraint is the internal fabric, not the distance to people.
Where the next campus should go.
The cheapest power on the grid is the power being thrown away. On grids from Texas to Australia to Inner Mongolia, wind and solar are curtailed or sell at negative prices because the transmission to move them does not exist. A flexible training load can sit on top of that stranded energy. This maps the candidates: how much surplus power is going begging, against how easy it is to build. The top right is the frontier.
The frontier is not where the announcements already are. West Texas is proven, but filling up and transmission-bound. The quieter opportunity is the wind belt, Oklahoma, Kansas and Iowa, where curtailment is climbing and the grid still has room. Find it before the queue does.
Stranded power is real but conditional. The same transmission gap that strands the energy can strand your campus, so co-location works only with on-site flexibility and a credible plan for firm backup. The axes here are an analytical scoring, not a survey: the curtailment figures are verified, the build-readiness ranking is judgement. Treat it as a shortlist to test, not a map to follow blindly.
For developers and capital chasing latency-insensitive training: shortlist the upper right before the announcements do, and underwrite the transmission risk explicitly. For policymakers in those regions: the curtailment that frustrates your renewable investors is also your strongest pitch to the one load that can actually use it.
Switch the scope to re-plot the same frontier worldwide. China runs the most systematic version: its East Data, West Computing programme relocates training to Inner Mongolia and the western provinces where curtailment is highest, while the Gulf's sub-two-cent solar makes bring-your-own-power viable at sovereign scale.
Sources: Amperon, Modo Energy, S&P Global, and CAISO and ERCOT market data on curtailment and negative pricing, 2024 to 2025; AEMO on the Australian NEM; China MIIT computing-hub mandates; PwC Middle East on MENA solar; CBRE on emerging markets. Curtailment and negative-price figures verified; stranded-power and build-readiness scores are Entelligencia's indicative composite, graded projected.
The moves, by who you are.
The same evidence asks something different of each reader. Choose your seat, and read the moves it leaves you.
The same evidence asks something different of each reader. Choose your seat: tap a panel to open its moves. Every set below is drawn straight from the report's findings on the screens above: the workload lens, the legitimacy ledger, the proportion and the phantom load.
The same evidence sends each workload somewhere different on the map. Globally a fourth force, sovereignty, now overrides cost and latency outright.
| Workload | The constraint | Where it goes worldwide |
|---|---|---|
| Training | No user latency, power-hungry | West Texas, the SPP belt, Inner Mongolia, Australia's NEM, Gulf solar |
| Inference | Within 100 ms of users | N. Virginia, London, Frankfurt, Tokyo, Singapore, Sydney, Sao Paulo, Mumbai |
| Edge, real-time | Under 20 ms, metro-bound | No market has enough; the gap is widest in Africa and SE Asia |
| Sovereign | Must run in jurisdiction | India, Brazil, Saudi Arabia, the EU, regardless of cost or latency |
Two forces reshape the map the US version does not see. The first is sovereignty: India's data-protection law, Brazil's LGPD, Saudi localisation and the EU's GDPR now pin certain workloads in-country whatever the power, water or latency penalty, which is the real reason Africa and India build despite the disadvantages. The second is the cooling supply chain. Rack density has climbed from 8 kW in 2021 toward 50 kW by 2027, and the manifolds and immersion gear behind direct-to-chip cooling carry equipment lead times of 8 to 24 months, a constraint operators meet alongside the grid queue on every continent.
Sources: IEA Energy and AI 2025; CBRE Global Data Center Trends 2025; Bain on build timelines; national data-protection statutes. Geographies are analytical mappings, graded projected.
Seven readings, one market.
Read alone, each instrument is a single force. Put together, they describe one global build-out that is roughly doubling its power draw, tripling its installed base and pulling record capital toward whichever market clears power, permission and sovereignty first. The US is the sharpest case in the set. It is no longer the whole story.
| Region | Installed, 2024/25 | 2030 projection | CAGR | Lead markets |
|---|---|---|---|---|
| North America | ~40 GW | 90 to 100 GW | ~14% | N. Virginia, Chicago, Phoenix |
| Europe | ~18.7 GW | 35 to 40 GW | ~12% | Frankfurt, London, Paris, Dublin |
| Asia-Pacific | ~13 to 15 GW | 30 to 40 GW | 15 to 20% | Tokyo, Singapore, Sydney, Seoul |
| China | 32 GW | 60+ GW | ~19% | Beijing, Shanghai, Ulanqab |
| Middle East | ~1 to 1.2 GW | 3.3 to 5 GW | 25 to 30% | Riyadh, Abu Dhabi, Dubai |
| Latin America | ~1.36 GW | 3+ GW | ~16% | Sao Paulo, Mexico City, Bogota |
| Africa | ~0.4 GW | 1.4 to 2.2 GW | 15 to 20% | Johannesburg, Lagos, Nairobi |
Sources: IEA Energy and AI 2025; Uptime Institute, CBRE and JLL 2025 to 2026 outlooks; Deloitte Powering Asia Pacific 2026; MUFG and Dell'Oro on hyperscaler capex; regional portfolio databases for Latin America and Africa. Installed figures verified where disclosed; 2030 projections and CAGRs are graded projected and move with each capex cycle.
