The AI data center buildout: the power edge
The AI buildout is the biggest infrastructure project of the decade. But what does it actually do — to power grids, to materials, to jobs, to your electricity bill? We traced the connections, sourced every figure, and tried to explain each link in plain language. Where we don't know, we say so.
The players
Who's driving the buildout — the labs training the models, the chipmaker, the hyperscalers writing the checks, and the power side keeping up. Click a name for context.
AI labs
Chipmakers
Hyperscalers & cloud
Power & grid
By the numbers
Four charts, each connected. Read them left to right, top to bottom — the blue annotation under each one explains how it connects to the next.
The five largest tech firms spent $400B+ in 2025 and are set to raise it ~75% in 2026 — most of it AI infrastructure. [11] IEA
This spending builds the data centers that consume the electricity above — and drives the price spike below.
From $28.92 to the $333.44 cap in three years; the grid's market watchdog pinned ~63% of the jump on data centers. [12] RTO Insider [5] Utility Dive
This cost gets passed to electricity bills — that's where regular people start paying for the AI buildout.
Thousands to build (~800–1,200 at peak); ~25–50 to run it. Billions of capex, dozens of permanent staff. [8] Brookings
The construction boom is real but temporary — and it pulls workers from housing and other projects that also need them.
Numbers to know
Every number that matters for understanding the buildout. Click a row to inspect it — or read the "Why it matters" column first if the metric is unfamiliar.
| Metric | Latest | Trend / projection | Why it matters | Confidence | Source |
|---|---|---|---|---|---|
| Global data-center electricity | ~485 TWh in 2025 (+17% YoY) | → ~950 TWh by 2030 (~3% of world power) | This is the raw demand signal — everything else on this page flows from this number going up. | Official | [11] [1] |
| US share of US electricity | ~4.4% in 2023 | → 6.7–12% by 2028 | Data centers are going from a rounding error to a meaningful fraction of America's total power use in just five years. | Official | [2] |
| Big Tech capital spending | >$400B in 2025 (5 largest tech firms) | → up ~75% in 2026, mostly AI infrastructure | This is the money that funds the buildout. When this goes up, electricity demand, construction jobs, and materials demand all follow. | Official | [11] |
| PJM grid capacity price | $333.44/MW-day (2027/28, at the cap) | from $28.92 in 2024/25; ~$530 uncapped | This is the cost of reserving future electricity in the eastern US. It 12x'd in three years — the clearest signal of data centers competing with everyone else for power. | Press-reported | [12] [5] |
| Generation waiting to connect | ~2,300 GW in US interconnection queues | ~4.5-year average wait | New power plants need years just to plug into the grid. This bottleneck means demand is growing faster than the system can add supply. | Official | [4] |
| Nuclear restarts for AI | Three Mile Island: 835 MW, 20-yr Microsoft PPA | back online ~2027 | Companies are restarting retired nuclear plants to get clean, reliable power — a move nobody expected five years ago. | Official | [3] |
| Copper demand | ~330k–500k tonnes/yr by 2030 | ~2% of global demand | Every data center and power line needs copper. AI is one of several reasons copper is getting more expensive — alongside EVs and housing. | Proxy / estimate | [6] |
| Power transformers | ~30% US supply deficit (2025) | ~2–4 year lead times | You can't connect a data center (or anything else) to the grid without transformers. There aren't enough, and the wait is years. | Proxy / estimate | [7] |
| Jobs per ~100 MW campus | ~800–1,200 building it (peak) | vs ~25–50 permanent staff | Massive construction boom, tiny permanent workforce. Great for electricians now; the long-term local job impact is much smaller than the construction phase suggests. | Press-reported | [8] |
| Compute contracted | Stargate $500B / 10 GW; Anthropic–Google up to 1M TPUs | tens of $B; >1 GW online in 2026 | Computing power is being bought and sold in advance, like oil futures. The scale of these contracts is what's pulling all the other numbers on this page. | Press-reported | [9] [10] |
| Nvidia → OpenAI | Up to $100B for ≥10 GW (~4–5M GPUs) | first GW H2 2026; Nvidia funds ~$10B equity per GW — 'circular financing' | The chip supplier is funding its biggest customer's purchases. This circular arrangement is either an ecosystem bet or a bubble warning, depending on who you ask. | Press-reported | [13] [14] |
| Largest single AI site | SpaceXAI (ex-xAI) Colossus: ~555k GPUs, ~2 GW, ~$35B (Memphis) | world's largest AI training site; xAI merged into SpaceX → SpaceXAI (2026) | One facility drawing 2 GW — roughly what a mid-sized city uses. This is the physical scale of AI training made concrete. | Proxy / estimate | [15] [18] |
| AI datacenter power (SemiAnalysis) | ~20 GW of new capacity energized in 2026 | +~30 GW expected 2027; US AI power ~3→28 GW (2023→26) | The buildout isn't slowing down — it's accelerating. US AI-specific power capacity grew nearly 10x in three years. | Proxy / estimate | [16] |
| AI-lab revenue vs. capex | Anthropic ~$65B, OpenAI ~$40B run-rate (mid-2026) | against $400B+ (2025) / ~$700B (2026) capex — the 'bubble' gap | The entire AI industry earns less in revenue than tech companies are spending to build it. Whether this gap closes or crashes is the single biggest financial question behind the buildout. | Press-reported | [17] |
How the numbers connect
Each card below is a connection between two things — with a plain-language explanation of the mechanism, what type of link it is (direct, indirect, correlated, or circular), who benefits, and who pays. The cascade reads top-down by order of consequence; the graph shows the same connections visually.
Click any connection card to open it in the right rail — then dig deeper with Ask.
First order — the raw demand
Every AI model runs on servers, and servers need power 24/7. More data centers means more electricity consumed — this is physics, not a forecast. The scale is what's new: a single large AI campus can draw as much power as a small city.
Data-centre electricity use jumped ~17% in 2025 to about 485 TWh (~1.5% of world power) and is projected to roughly double to ~950 TWh by 2030 (IEA). In the US, data centers were ~4.4% of electricity in 2023, heading to 6.7–12% by 2028 (Berkeley Lab, 2025 update). [11] [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
All that new electricity has to come from somewhere. Power companies are signing long-term deals to supply data centers — in some cases, spending billions to restart retired nuclear plants. The demand is so large that new power plants are waiting years just to connect to the grid, creating a bottleneck for everyone.
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.
Data centers and the power plants feeding them are built from real physical stuff — copper wiring, transformers, turbines. AI isn't the only reason these are expensive (electric vehicles and housing need copper too), but it's adding major new demand to markets that were already strained. The connection is real but shared with other industries.
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.
You can't build a data center without electricians, welders, and construction crews — hundreds of them, for 18 to 36 months. But once it's built, a massive campus might employ only a few dozen people to run. The construction boom is real but temporary, and it pulls workers away from housing, roads, and other projects that also need them.
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
When data centers book up the available power supply, the cost of reserving future electricity goes up for everyone. In the eastern US grid (called PJM), the price to reserve capacity 12x'd in three years — and the grid's own market watchdog attributed about 63% of that spike to data center demand. But this is contested: utilities and tech companies disagree on how costs should be split. Either way, the cost eventually shows up in regular people's electricity bills.
Who wins
—
Who pays
Rate-payers; delayed / priced-out projects
PJM's grid capacity price ran from $28.92/MW-day (2024/25) to $269.92 (2025/26) to $333.44 (2027/28) — clearing at the regulatory cap; uncapped, the 2027/28 auction would have hit ~$530. The market monitor tied ~63% of the 2025/26 jump to data centers. [12] [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.
Computing power itself is becoming something bought and sold in massive advance contracts, like oil futures. And some of the financing is circular — Nvidia is investing billions in OpenAI, which turns around and buys Nvidia's chips. A supplier funding its own customer's purchases. Critics call it bubble mechanics; supporters call it an ecosystem bet. Either way, the numbers are staggering and the interdependence is real.
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; Nvidia is investing up to $100B in OpenAI as it deploys ≥10 GW of Nvidia systems — the chipmaker partly financing its own sales ('circular financing'); Anthropic's deal for up to 1 million Google TPUs is worth tens of billions, >1 GW online in 2026. [9] [13] [14] [10]
Where the data runs out: Deal values are announced totals, not audited spend, and Stargate has reportedly faced delays. The Nvidia–OpenAI structure is partly 'circular' — a supplier funding its customer's purchases — which critics flag as a bubble risk. 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. The Nvidia–OpenAI structure is partly 'circular' — a supplier funding its customer's purchases — which critics flag as a bubble risk. 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]United States Data Center Energy Usage Report: 2025 Update — 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
- [11]Data centre electricity use surged in 2025, even with tightening bottlenecks — IEA
- [12]PJM Capacity Auction (2027/28) Clears at Max Price, Falls Short of Reliability Requirement — RTO Insider
- [13]OpenAI and NVIDIA announce strategic partnership to deploy 10 GW of NVIDIA systems — OpenAI
- [14]Nvidia plans to invest up to $100 billion in OpenAI as part of data center buildout — CNBC
- [15]Colossus (data center) — xAI's Memphis supercluster (~555k GPUs, ~2 GW, ~$35B) — Wikipedia
- [16]AI Datacenter Energy Dilemma — Race for AI Datacenter Space — SemiAnalysis
- [17]Anthropic revenue run rate surpasses $65 billion pre-IPO (OpenAI ~$40B) — Axios
- [18]Musk's xAI, SpaceX combo is the biggest merger of all time, valued at $1.25 trillion (xAI later rebranded SpaceXAI) — CNBC
Keep exploring
Power & grid is one edge of the buildout. Follow connected perspectives into chips, materials, capital, policy, jobs, and more — or see what we're adding next.
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.