Real-time load, annual demand forecasts, and substation feeds — the dashboards behind the AI data center electricity story.
Ask five research houses how much electricity AI data centers will use by 2030, and you'll get five different numbers — but they all point the same direction: sharply up. Global data center electricity demand is on track to roughly double by 2030, and in the US, data centers could account for anywhere from 9% to 20%+ of national electricity within a decade. The forecasts vary by methodology, but the trajectory is no longer in serious dispute — what's still up for debate is how fast it arrives, who absorbs the cost, and which regions get squeezed first.
That last question is where this stops being an abstract energy-sector statistic and starts being a live variable in site selection, utility rate cases, and capital allocation decisions being made this year. Below, we break down the numbers behind the headlines — and what they actually mean if you're the one building, financing, or regulating this next wave of infrastructure.
The Four Numbers That Matter Most
Skip the methodology debates for a second — these are the figures that should actually change how a developer, utility, or investor plans the next 24 months.
Global data center electricity demand roughly doubles between 2024 and 2030, per the IEA's Base Case (415 TWh → ~945–950 TWh).
Share of global demand growth through 2030 concentrated in just two countries: the US and China, per the IEA.
The year Gartner expects AI-optimized servers to overtake conventional servers as the largest share of data center power draw.
Ceiling of EPRI's modeled range for how much of Virginia's electricity data centers could draw by 2030.
The Global Baseline: Where Demand Stands Today
The IEA's Energy and AI analysis puts global data center electricity use at roughly 415 TWh in 2024 — about 1.5% of global electricity — after growing around 12% annually over the prior five years. That growth accelerated sharply in 2025: the IEA estimates overall data center demand rose about 17% to roughly 485 TWh, with AI-focused facilities alone up around 50%.
Gartner's separate tracking lands close to the same 2025 figure, at 447 TWh globally, and projects a 26% jump in 2026 to 565 TWh — with global power demand climbing from roughly 104–105 GW to 132–133 GW in a single year.
Modeled from IEA, Energy and AI (April 2025) with updates in Key Questions on Energy and AI (2026).
2030 Forecasts, Side by Side
No two research houses model this the same way — some track global TWh, others US capacity in GW — but laid out together, the forecasts converge on the same doubling-or-more story.
| Source | 2030 Forecast | Scope |
|---|---|---|
| IEA | ~945–950 TWh (Base Case), ~3% of global electricity | Global |
| Gartner | ~980 TWh / ~290 GW power demand in some outlooks | Global |
| LBNL | 649 TWh Reference Case (range 521–843 TWh); ~11.8% of US electricity | United States |
| BloombergNEF | ~118 GW capacity (2030); 194 GW by 2035, ~12% of US electricity by 2030 | United States |
| EPRI | 380–790 TWh (9–17% of US electricity); peak 45–94 GW | United States |
| McKinsey | ~121 GW of IT demand; ~75% of US power demand growth this decade | United States |
| Goldman Sachs | Up to +165% vs. 2023 baseline (some updates higher) | Global |
| Bank of America | ~125 GW added US load; electricity demand CAGR ~4.1% (2026–2030) | United States |
Figures compiled from primary reports published through mid-2026; GW (capacity) and TWh (consumption) are related but not directly interchangeable, since both depend on utilization and power usage effectiveness (PUE).
Strip away the different units and one comparable metric emerges from three separate models: what share of total US electricity data centers will actually consume by 2030. Lined up, the spread is narrower than the headline TWh and GW figures suggest — a rare point of convergence worth paying attention to.
Bar width scaled to the 2035 figure (20%) as the reference maximum.
Why the US Carries an Outsized Share
The IEA finds that the US and China together account for nearly 80% of global data center demand growth through 2030 — with the US alone adding roughly 240 TWh, a jump of about 130% from 2024 levels. Put in industrial terms, the IEA projects US data centers will use more electricity by 2030 than aluminum, steel, cement, chemicals, and other energy-intensive manufacturing sectors combined.
That growth isn't spread evenly. EPRI's state-level modeling shows Virginia already drawing more than 20–25% of its electricity from data centers, with that share potentially climbing to 39–57% by 2030. BloombergNEF sees similar concentration in grid regions rather than states — projecting data centers could reach roughly 34% of demand in PJM and 22% in ERCOT by 2035.
US + China share of growth
Nearly 80% of global data center demand growth through 2030, per the IEA.
Virginia's exposure
Already 20–25%+ of state electricity; EPRI models up to 39–57% by 2030.
PJM by 2035
Data centers could reach ~34% of regional demand, per BloombergNEF.
ERCOT by 2035
Roughly 22% of regional demand in the same BloombergNEF modeling.
This regional concentration has knock-on effects well beyond utility bills — it's already reshaping site selection and land economics for developers. If you're evaluating markets, our data center site selection checklist walks through the power-availability criteria that now sit ahead of most other siting factors.
What's Driving the Growth
The core driver across every forecast is the same: AI training and inference workloads run on high-density accelerated servers that draw far more power per rack than traditional enterprise computing. Gartner's data captures this shift directly — AI-optimized servers accounted for 95 TWh of global demand in 2025, are projected to reach 175 TWh in 2026 (an 84% jump), and are expected to overtake conventional servers' share of total data center power by 2027.
Efficiency gains are real but are being outrun by volume. Per-task hardware and software efficiency continues to improve rapidly, but that's being offset by sheer growth in usage — more models, more inference calls, and increasingly intensive use cases like video generation and autonomous agents.
On the supply side, renewables paired with storage and natural gas lead most near-term buildouts, while nuclear, small modular reactors, and on-site generation are growing fast in project pipelines — largely because grid interconnection, transformers, turbines, and even chip supply remain persistent bottlenecks. For a closer look at how power shortfalls translate into real operational risk, see our piece on AI data center power outages and grid risk in 2026.
Why the Ranges Are So Wide
Look closely at any single forecast and the range inside it is often as telling as the headline number. EPRI's own scenarios for 2030 US demand span from 380 to 790 TWh depending on how fast development accelerates. LBNL's Reference Case of 649 TWh comes with a sensitivity band of 521–843 TWh, driven by variables like AI chip lifetimes, idle utilization rates, and the mix of specialized graphics chips deployed.
Three factors explain most of that spread: how fast AI adoption actually scales, how much hardware efficiency gains offset rising volume, and how quickly supply-chain bottlenecks in chips, power equipment, and grid connections get resolved. McKinsey's research adds a further wrinkle — interconnection queues already contain far more proposed capacity than will realistically be built, meaning headline pipeline numbers tend to overstate what actually comes online.
Advisory: What Each Stakeholder Should Do Now
Forecasts are only useful if they change decisions. Here's how the numbers above translate into concrete moves for the groups most exposed to this shift.
For Developers & Investors
Treat power availability, not land cost, as the primary underwriting variable. With interconnection queues holding far more proposed capacity than will realistically be energized, verify a project's actual position in the queue — not just its listed megawatt allocation — before modeling returns. Markets with early on-site generation approval (gas, or nuclear power purchase agreements) are increasingly de-risking timelines faster than markets waiting purely on utility upgrades.
For Utilities & Grid Planners
The concentration risk in regions like Virginia, PJM, and ERCOT means load forecasting needs to shift from aggregate regional demand to facility-level tracking of announced, permitted, and under-construction projects. Rate design and cost-allocation frameworks that separate data center load from residential and commercial classes are worth revisiting now, before concentration reaches the higher end of EPRI's modeled range.
For Corporate Energy Buyers
Long-term power purchase agreements are becoming a competitive necessity rather than a hedge — hyperscalers securing nuclear restarts and dedicated generation are effectively buying certainty that smaller buyers can no longer count on from spot markets. Locking in supply now, even at a premium, is increasingly cheaper than absorbing multi-year delays later.
For Policymakers
The IEA's finding that US data centers could out-consume aluminum, steel, cement, and chemicals combined by 2030 is a planning-level signal, not a talking point — it implies transmission and permitting reform now has a direct bearing on industrial competitiveness, not just tech-sector growth. Streamlining interconnection review without weakening reliability standards is the highest-leverage lever available in the near term.
What This Means Going Forward
Despite the differences in methodology, the IEA, Gartner, LBNL, BloombergNEF, EPRI, McKinsey, Goldman Sachs, and Bank of America are all converging on the same basic shape: global data center electricity use roughly doubling by 2030, with the US absorbing a disproportionate share of that growth and bearing the infrastructure strain that comes with it. Even the IEA's more conservative framing — that data centers drive less than 10% of global electricity demand growth overall — still concedes nearly half of US growth and more than 20% of growth across advanced economies broadly.
For developers and investors, that combination of scale and concentration is reshaping where and how projects get built — and increasingly, how much they cost to build in the first place. Our recent analysis of AI data center development costs in 2026 breaks down how power constraints are feeding directly into capital budgets, and our earlier look at AI data center electricity demand tracks how these forecasts have shifted over the past several quarters. The broader ripple effects are also worth watching beyond the data center fence line — see our coverage of AI infrastructure growth's effects on commercial real estate.
Related Reads
Figures referenced are 2025–2026 projections from the IEA, Gartner, Lawrence Berkeley National Laboratory, BloombergNEF, EPRI, McKinsey, Goldman Sachs Research, and Bank of America. Figures are subject to revision based on actual build rates, policy, and equipment availability.
Core Insights Review contributors publish research-based analysis and editorial insights on commercial real estate, PropTech, smart infrastructure, sustainable construction, industrial real estate, and emerging technologies shaping the future of the built environment.
Check for more information: Core Insights Review
