The AI data center buildout: the power edge
What the buildout is actually doing to power — traced three orders deep, every figure sourced. Not a pitch; the data.
By the numbers
The headline figures, from reputable sources. Full context and the trade-offs are traced below.
Set to roughly double by 2030 — from ~1.5% to ~3% of world electricity. [1] IEA
From 4.4% in 2023 toward as much as ~1/8 of all US power by 2028. [2] Lawrence Berkeley National Laboratory
+833% in a single year; the market monitor pinned ~63% on data centers. [5] Utility Dive
Thousands to build (~800–1,200 at peak); ~25–50 to run it. Billions of capex, dozens of permanent staff. [8] Brookings
The full trace
Every edge — who wins, who pays — three orders deep, sourced.
First order — the raw demand
Data centres used ~415 TWh in 2024 (~1.5% of world electricity) and are projected to roughly double to ~945 TWh by 2030 (IEA). In the US they were ~4.4% of electricity in 2023 (176 TWh) and could reach 6.7–12% by 2028 (Berkeley Lab). [1] [2]
Where the data runs out: Projections diverge widely — Berkeley Lab's 2028 US range alone is 6.7–12% — and AI's exact share within data-center load is uncertain.
Second order — who wins, who pays
Who wins
Utilities, IPPs, gas & nuclear operators
Who pays
The grid — interconnection queues & strain
Constellation is restarting the 835 MW Three Mile Island Unit 1 ($1.6B) on a 20-year PPA sending 100% of its output to Microsoft. Meanwhile ~2,300 GW of generation & storage sits in US interconnection queues, and the typical project now waits ~4.5 years to connect (Berkeley Lab). [3] [4]
Where the data runs out: Which specific plants are contracted to which campuses is largely undisclosed; queue volume actually fell ~12% year-over-year as withdrawals rose.
Who wins
Copper, transformer & turbine suppliers
Who pays
Every other buyer — price & lead-time pressure
AI data centres could consume ~330,000–500,000 tonnes of copper a year by 2030 (~2% of global demand). US power transformers face a ~30% supply deficit in 2025 (Wood Mackenzie), with large-transformer lead times around 2+ years. [6] [7]
Where the data runs out: Copper figures are analyst forecasts that vary; transformer lead times are reported by suppliers, not a clean public series.
Who wins
Electricians & construction trades
Who pays
Sectors that lose those workers; wage pressure
A 100 MW hyperscale campus can employ ~800–1,200 workers at peak over an 18–36 month build — but once open often runs on only ~25–50 permanent staff (Brookings; industry data). [8]
Where the data runs out: Permanent headcount varies by site (≈50–200 for larger facilities); peak construction numbers are self-reported by developers.
Third order — where it lands
Who wins
—
Who pays
Rate-payers; delayed / priced-out projects
PJM's 2025/26 capacity auction cleared at $269.92/MW-day — an 833% jump from $28.92. The market monitor attributed ~63% of the increase (~$9.3B) to data centers; the next auction hit the $329.17/MW-day cap. [5]
Where the data runs out: Attributing regional price moves to data centres specifically is contested; utilities and data-center firms dispute how costs should be allocated.
Who wins
Compute sellers & chip vendors
Who pays
Buyers — capex, dependency, concentration
Compute is being contracted at unprecedented scale: OpenAI's Stargate targets $500B and 10 GW over four years; Anthropic's deal for up to 1 million Google TPUs is worth tens of billions and brings >1 GW online in 2026. [9] [10]
Where the data runs out: Deal values are announced totals, not audited spend, and Stargate has reportedly faced delays. A reported 'SpaceX selling compute to AI labs' deal could not be verified, so it is omitted.
Where the data runs out
Surfaced, not hidden — every place this map is thinner than we'd like.
- Electricity demand: Projections diverge widely — Berkeley Lab's 2028 US range alone is 6.7–12% — and AI's exact share within data-center load is uncertain.
- Generation & the grid: Which specific plants are contracted to which campuses is largely undisclosed; queue volume actually fell ~12% year-over-year as withdrawals rose.
- Materials: Copper figures are analyst forecasts that vary; transformer lead times are reported by suppliers, not a clean public series.
- Labour: Permanent headcount varies by site (≈50–200 for larger facilities); peak construction numbers are self-reported by developers.
- Rate-payers: Attributing regional price moves to data centres specifically is contested; utilities and data-center firms dispute how costs should be allocated.
- Compute deals: Deal values are announced totals, not audited spend, and Stargate has reportedly faced delays. A reported 'SpaceX selling compute to AI labs' deal could not be verified, so it is omitted.
Sources
- [1]Energy and AI — Energy demand from AI — IEA
- [2]2024 United States Data Center Energy Usage Report — Lawrence Berkeley National Laboratory
- [3]Constellation to Launch Crane Clean Energy Center (Three Mile Island Unit 1 restart; 20-year Microsoft PPA) — Constellation Energy
- [4]Queued Up: 2025 Edition — Characteristics of Power Plants Seeking Transmission Interconnection — Lawrence Berkeley National Laboratory
- [5]Data centers 'primary reason' for high PJM capacity prices: market monitor — Utility Dive
- [6]AI data center copper demand & 2030 forecasts — Fastmarkets
- [7]Power transformers and distribution transformers will face supply deficits of 30% and 10% in 2025 — Wood Mackenzie
- [8]New evidence on data center employment effects — Brookings
- [9]OpenAI's first data center in $500 billion Stargate project is open in Texas — CNBC
- [10]Google and Anthropic announce cloud deal worth tens of billions of dollars (up to 1M TPUs) — CNBC
Every figure is sourced and confidence-tagged; consequences are read second- and third-order (who wins, who pays). Understanding, not investment advice. Found something wrong or missing? Tell us.