Charted: Top-1 Holds 45% of AI Power Demand — Top-3 Holds 85%

Global averages hide the real story of AI power. Data centres used about 415 TWh in 2024 — only ~1.5% of world electricity — and the IEA Base Case still has them near ~3% by 2030. Those percentages are true, and they are also the wrong unit for planning. What matters for grids is concentration: how much of the load sits in the top of the distribution, which clusters absorb the next gigawatts, and whether wires can keep pace where the GPUs actually plug in.

The interactive dashboard above answers that distribution question directly. Top-1 country share is 45% (United States). Top-3 regional share is 85% (US + China + Europe). US + China capture roughly 80% of growth to 2030. Pair this concentration cut with our global IEA scenario frame for the 415→945 TWh path, and with the US-only data-center power vs grid capacity post for LBNL demand versus transmission miles.

The concentration scoreboard

Read the table as a ladder, not a single scare statistic. Global share stays modest. Country share is already oligopolistic. Cluster and state shares are where brownout politics live.

Top-1 and top-3: the regional ladder

In 2024 the United States accounted for 45% of global data-centre electricity consumption, China 25%, and Europe 15%. Add those three and you have 85% of the world’s data-centre load in a handful of jurisdictions. Japan is a distant single-digit slice; the rest of the world shares the remainder.

That is not how most electricity sectors look. Aluminium, steel, and cement are geographically wide even when capital is concentrated. AI-era data centres behave more like specialised industrial clusters: a few markets with deep fibre, latent interconnects, tax regimes, and hyperscaler campus footprints. The dashboard’s cumulative curve makes the inequality visual — the top bucket alone is nearly half the pie; the equal-split diagonal is nowhere near the actual path.

Absolute levels reinforce the share story. The US consumed about 187 TWh of data-centre electricity in 2024; China about 104 TWh; Europe about 62 TWh. By 2030 in the Base Case those become roughly 427 / 279 / 107 TWh. The United States adds about 240 TWh (+130%); China about 175 TWh (+170%); Europe more than 45 TWh (+70%). Roughly four-fifths of incremental demand lands in two countries.

Growth is more concentrated than the stock

Stock concentration (who holds today’s TWh) and growth concentration (who captures the next TWh) are related but not identical. The US already dominates the stock; China is catching up faster on a percentage basis; Europe grows from a smaller base. Together, US + China still dominate the delta.

That matters for equipment supply chains and for fuel. Transformers, high-voltage cable, gas turbines, and interconnection study bandwidth are finite. When ~80% of sector growth bids into the same two systems, lead times stretch everywhere those buyers touch — even markets that are not building AI campuses feel the queue. Our companion IEA global frame shows renewables meeting about half of incremental supply to 2030 while gas and coal still cover more than 40%; concentration decides which local grids absorb that fossil bridge.

Clusters: half of US capacity in five hubs

Country totals still average away the problem. The IEA is explicit that AI-focused data centres can draw as much power as aluminium smelters but are much more geographically concentrated. Nearly half of US data-centre capacity sits in five regional clusters. About 50% of US capacity under development remains inside those pre-existing large hubs.

Named clusters in the IEA’s global map — Northern Virginia, Beijing, Shanghai, Dallas, Pearl River Delta, Singapore, Chicago, Dublin, London, Omaha — are the practical map of AI power risk. Northern Virginia alone is the world’s largest market: roughly 4.9 GW of operating IT load and about 13% of reported global operational capacity in public market tallies cited by Virginia’s JLARC. It is more than double Beijing, the next-largest global market in those same tallies.

The dashboard’s cluster scatter puts operating load against estimated pipeline intensity. Markets that are already large and still filling pipelines are the ones where local congestion, capacity-market prices, and siting politics collide. Dallas-area large-load forecasts and Northern Virginia’s multi-gigawatt pipeline are not abstract “US demand” — they are specific substations and transmission corridors.

Local intensity: Virginia 25%, Ireland 20%

Zoom one more step and the global 1.5% statistic collapses. In Ireland, data centres consume around 20% of metered electricity supply. In the United States, six states already see data centres above 10% of electricity supply, with Virginia leading at about 25%. That is an order of magnitude above the world average — and it is the everyday reality for the grids hosting the densest AI and cloud capacity.

Local intensity is why interconnection reform and transformer manufacturing matter more than another global percentage-point debate. A country can have ample national generation and still fail a hyperscale campus if the cluster’s local network is saturated. Conversely, siting outside saturated hubs can unlock power that national models already count as available.

Emerging and developing economies outside China illustrate the mirror image: they hold roughly half of the world’s internet users but under 10% of global data-centre capacity. Concentration is not only too much load in a few places — it is also computing power locked away from places that might want to host it if power quality and interconnection were investable.

Can electricity and grid build-out keep pace?

The core question of this theme is whether power and wires can keep up with AI load. Concentration changes the answer from “maybe globally” to “not yet where it counts.”

On the demand side, a hyperscale campus can stand up in roughly two to three years once power is secured, and accelerator refreshes cycle even faster. On the grid side, building new transmission in advanced economies often takes four to eight years. US median time from interconnection request to commercial operation exceeded five years for projects that reached operation in 2025 (LBNL Queued Up 2026). Wait times for transformers and cables have roughly doubled in three years. The IEA estimates that unless grid risks are addressed, around 20% of planned data-centre projects could face delays.

Pace mismatch is lethal precisely because load is clustered. If demand were sprinkled evenly across every balancing authority, five-year queues would still be painful but diffuse. When half of US capacity — and half the pipeline — packs into five clusters, every delayed transformer and every restudy hits the same corridors. Firm generation announcements do not clear that bottleneck by themselves; deliverability does.

What would change the concentration story

Several shifts would rewrite this scoreboard:

1. Siting outside saturated US clusters at scaleso the “50% of pipeline in existing hubs” figure falls for several consecutive years.
2. Faster interconnection and transmission that compress median IR→COD below the campus construction clock in PJM, ERCOT, and other stressed regions.
3. Meaningful non-US/China growth that drops the US+China share of incremental TWh well below ~80% without simply shifting the same congestion to Dublin, Singapore, or the Pearl River Delta.
4. Virginia and Ireland intensity peakinglocal DC shares of electricity stabilising or falling as efficiency, flexibility, and geographic diversification bite.
5. EMDE capacity share rising toward internet-user share, which would both diversify global concentration and test power-quality readiness in new host markets.

None of those are guaranteed by another round of campus announcements. They are grid, siting, and industrial-policy outcomes.

Caveats and reading notes

• Regional shares are IEA disclosed; Japan and rest-of-world splits in our five-bucket ladder include estimated residuals that reconcile to 415 TWh.
• Cluster IT-load gigawatts beyond Northern Virginia are ranked estimates consistent with public “NoVa more than double Beijing” market tallies and the IEA named-cluster listtreat them as order-of-magnitude, not meter reads.
• Pipeline intensity by cluster is estimated from announced-development narratives; the IEA chart of installed vs under-development capacity is the qualitative source, not a full disclosed table in our module.
• HHI is an approximate index from five regional buckets, useful for ranking concentration across posts, not a full reconstructed global plant-level distribution.
• AI vs non-AI workloads inside data centres remain hard to separate; accelerated-server growth is the best available proxy in the IEA frame.
• Figures can revise as OMDIA, utility IRP filings, and IEA updates refreshdirectionally the concentration ladder has been stable even when point estimates move.

Bottom line

The AI power problem is a concentration problem. Top-1 holds 45%, top-3 holds 85%, and US + China take about 80% of the growth to 2030. Nearly half of US capacity sits in five clusters; Virginia and Ireland already run data-centre shares of electricity an order of magnitude above the global average; and roughly one in five planned projects sits at grid-delay risk while interconnection medians stretch past five years. Electricity can be built for AI. The open question is whether it can be built where the distribution is thickest, on a clock that matches campus construction rather than transmission planning.

Related reading: IEA global data-centre scenarios and US data-center power vs grid capacity.