One-line takeaway
Fewer than 100 projects, 50 GW, the peak load of an entire UK. Money moves from USD 565 billion toward USD 4 trillion; electricity moves from 415 TWh toward 945 TWh.
Scope and notation
This dataset collects public figures from the International Energy Agency's World Energy Outlook 2026 and its Energy and AI report, plus press coverage of both. Every number carries its source notation; figures from different notations are not combined.
Three concepts must be kept apart: investment (money), capacity and power demand (watts), and electricity consumption (energy). The units differ and cannot be converted into "dollars per kWh" without a load factor the IEA has not published.
Investment
| Metric | Value | Notation |
|---|---|---|
| Global data center investment, 2025 | USD 565 billion | IEA, annual |
| Cumulative investment by 2030 | About USD 4 trillion | IEA, cumulative |
| Share going to AI training and inference | About three quarters | IEA, projected increment |
| Non-AI load | Cloud storage, video streaming and other traditional workloads | IEA category |
Power and capacity
| Metric | Value | Notation |
|---|---|---|
| AI-dedicated mega projects | Fewer than 100 | IEA |
| Combined power demand of those projects | 50 GW | IEA |
| Reference: UK peak electricity demand | Roughly 50 GW | press conversion |
| AI-dedicated large data centers tracked by satellite imagery | 93 | IEA |
| Capacity growth over two years | About 10x | IEA, satellite view |
| Global data center electricity use, 2024 | About 415 TWh | IEA, Energy and AI |
| Share of global electricity | About 1.5% | IEA, Energy and AI |
| Projected use, 2030 | About 945 TWh | IEA, Energy and AI |
| Reference: Japan's annual electricity use | on the order of 945 TWh | press conversion |
| Data center emissions by 2030 | Up about 70% | IEA |
| 403 US data centers | Nearly 2% of US electricity | aggregated press notation |
Hard constraints on supply
| Constraint | Manifestation |
|---|---|
| Build-time mismatch | Large data centers can be built and connected in 2-5 years; transmission expansion often takes a decade or more |
| Local connection limits | Dublin has capped new data center connections |
| Tightening approvals | The Netherlands has tightened permitting over infrastructure and land constraints |
| Grid headroom | In some growth markets, planned projects exceed local current peak load |
How to read the tables
First, USD 565 billion is one year of spending; USD 4 trillion is cumulative through 2030. Calling that "a sevenfold increase" is wrong. The correct reading is that at the current pace, the next five years accumulate roughly seven times a single year's outlay.
Second, 50 GW is power, not energy. It describes how much generation and transmission capacity would be needed if everything ran at once. That is why grid interconnection, rather than annual energy, is the binding constraint.
Third, 415 TWh to 945 TWh is a doubling of consumption. Within a global total where data centers account for 1.5%, the absolute number is modest, but its growth rate far exceeds total demand growth, which is why the IEA treats it as a structural shift rather than a marginal variable.
Fourth, 93 projects and "fewer than 100" do not conflict: the first is the satellite-observable sample, the second the statistical set. The value of the satellite view is that it provides third-party observation independent of corporate disclosure. The 10x capacity growth figure comes from imagery, not from filings.
Fifth, the roughly 70% emissions increase refers to power-related emissions from data centers, not global emissions. The numerator grows fast while the denominator changes slowly, so the ratio means little; the absolute increment is what matters.
Notation caveats
Three discrepancies appear across coverage of the same facts.
One: investment figures may or may not include land, shells and supporting power infrastructure. USD 565 billion and USD 4 trillion can be compared directly only if they come from the same series.
Two: "10x capacity growth in two years" describes installed capacity visible from orbit, not compute (FLOPS). GPU generations raise compute per watt substantially, so capacity growth and compute growth are different numbers.
Three: 50 GW is demand; 945 TWh is consumption. Converting between them requires a load factor the IEA has not disclosed, so this dataset performs no such conversion.
Boundaries
- Investment figures, project counts, 50 GW, 415 and 945 TWh, 93 projects, 10x capacity and the 70% emissions figure come from IEA reports and press coverage, without page-level verification of every chart;
- "UK peak of about 50 GW" and "Japan's annual use of about 945 TWh" are press conversions offered for scale, not IEA figures;
- "403 US data centers at nearly 2%" is an aggregated press notation whose statistical boundary (self-generation, cooling) is undisclosed;
- The Dublin and Netherlands connection restrictions are policy facts; exact terms follow local regulatory documents;
- This dataset records and scales figures only; it contains no investment advice or forecast.
Sources
- Sina Finance: IEA energy demand warning, AI data centers push the grid to its limit
- IEA: Energy and AI, executive summary
- IEA: Energy and AI, energy supply for AI
- Data Centre Magazine: IEA expects USD 4tn in data centre investment by 2030
- China Economic Net: global AI spending and data center power figures
