|
| Cover Image |
The Briefing, in Five Points
VERIFIED Global data center electricity demand grew 17% in 2025 to roughly 485 TWh — AI-focused facilities alone grew 50%, three times faster than the sector average.
REVISED THIS MONTH BloombergNEF lifted its 2035 U.S. forecast to 20% of national electricity, an 83% jump over its own December estimate — the sharpest single revision of the year.
MODELED Goldman Sachs expects U.S. data center power demand to double from 31 GW (2025) to 66 GW (2027), contingent on 95 GW of capacity actually being built on schedule.
STRUCTURAL SHIFT Gartner projects AI-optimized servers will draw more electricity than every conventional server combined by 2027 — a crossover that redefines what a "typical" data center consumes.
BOTTLENECK Bank of America reports large gas turbines are sold out through 2030; the constraint on AI growth has shifted from chip supply to grid and generation capacity.
What changed since our last update
Three forecasters revised their numbers upward in the past three weeks: BloombergNEF (July 21), Gartner (July 14–20), and Bank of America (June 26). Every revision moved in the same direction — higher and sooner. That consistency across independent methodologies is itself a signal worth weighing more heavily than any single number below.
In this article
01 · Snapshot Dashboard
565 TWh
Gartner's 2026 Global Forecast
~950 TWh
IEA's 2030 Base Case
20%
US Share of Electricity by 2035 (BloombergNEF)
66 GW
US Data Center Power Demand, 2027 (Goldman Sachs)
02 · Latest Developments
Five updates from the past four weeks, in order, each moving forecasts in the same direction:
July 21, 2026 — BloombergNEF
Raised its 2035 forecast for U.S. data center electricity share to 20%, 83% above its own December 2025 estimate, citing faster AI chip deployment than modeled six months ago.
July 14–20, 2026 — Gartner
Confirmed AI-optimized servers are on track to overtake conventional servers in total electricity draw by 2027, with power availability now the binding constraint on capacity growth.
July 7, 2026 — NTT Data
Reported that power, equipment, land, and permitting limits are colliding simultaneously, making electricity access a deciding factor in where new capacity gets sited.
July 3, 2026 — US Heatwave
A nationwide heatwave pushed several utilities to ask data center operators to switch to backup generators — an early stress test of grid strain from AI load growth.
June 26, 2026 — Bank of America
Warned that large gas turbines are largely sold out through 2030, forcing developers toward on-site gas engines while utilities extend coal plants and add batteries.
03 · The 2025 Baseline: What Actually Happened
Forecasts are only as credible as the actuals they're anchored to. The IEA's latest Energy and AI update confirms global electricity demand from data centers rose 17% in 2025 — in line with its own prior projection — while electricity use inside AI-focused facilities specifically jumped 50%, far outpacing the rest of the sector. That gap is the story: the "data center" category is no longer one thing, and treating it as one understates where the growth is actually concentrated.
Gartner's independent tally corroborates the direction: global data center electricity consumption reached about 447 TWh in 2025, with AI-optimized servers alone drawing roughly 95 TWh of that. Gartner now expects total consumption to climb to 565 TWh in 2026, a 26% single-year increase, with AI-optimized draw rising 84% to 175 TWh. Gartner analyst Linglan Wang frames the shift directly: power availability, not compute capacity, has become the new competitive constraint on scaling AI.
A countervailing force is real but insufficient on its own. The IEA reports that efficiency gains in chips, software, and cooling have cut the energy cost of a typical AI task by roughly an order of magnitude annually — a routine text query today draws less electricity than running a television for the same stretch of time. But usage volume is outrunning those gains: major model providers reported a roughly threefold rise in active users and a fivefold rise in revenue over the past year.
"Power availability [is] the new battleground for scaling and protecting margins in the global AI race." — Linglan Wang, Director Analyst, Gartner
04 · The IEA's 2030 Outlook: Toward 950 TWh
Under the IEA's Base Case, global data center electricity consumption nearly doubles again, from about 485 TWh in 2025 to roughly 950 TWh by 2030 — just under 3% of total global electricity consumption, up from 1.5% only a few years earlier. AI-focused data center electricity specifically is expected to nearly triple across that window, though bottlenecks in permitting, transformers, turbines, and skilled labor temper the odds of the most aggressive near-term scenarios materializing on schedule.
The broader power system is accelerating in parallel: the IEA's Electricity 2026 report puts global electricity demand growth at an average 3.6% a year through 2030 — about 50% faster than the previous decade — with data centers named alongside industry, electric vehicles, and air conditioning as core drivers. China alone is expected to account for close to half of that total increase across all sectors, not data centers specifically, which is a useful reminder that this is one growth driver among several, not the whole story.
IEA Global Data Center Electricity Demand
| Year | Electricity Demand (TWh) | Share of Global Electricity |
|---|---|---|
| 2024 | 415 | 1.5% |
| 2025 | ~485 | ~1.7% |
| 2030 (Base Case) | ~950 | ~3% |
05 · Goldman Sachs: U.S. Power Demand Set to Double by 2027
Goldman Sachs Research's commodities team — Hongcen Wei, Daan Struyven, and Samantha Dart — published an updated U.S.-focused forecast in May 2026 built on facility-level tracking data from infrastructure intelligence provider Aterio, covering permitting status, construction progress, and satellite imagery. The central case: U.S. data center power demand rises from 31 GW in 2025 to 41 GW in 2026 and 66 GW in 2027, implying roughly 95 GW of installed capacity by the end of 2027, on a 70% capacity utilization assumption.
The scheduled pace of additions is the operative detail for planners: 13.6 GW in 2026 and 36.3 GW in 2027, against just 6.4 GW realized in 2024 and 8.5 GW in 2025. Goldman Sachs is explicit about execution risk — historically, only about 60% of capacity scheduled for the coming year, and roughly 50% scheduled two years out, actually comes online on time, constrained by supply chains, labor, and the typical 18-to-24-month build cycle once permits clear. Even discounted for that risk, the bank still expects data centers to reach close to 8.5% of total U.S. power demand by 2027, up from about 4.4% in 2023.
Goldman Sachs U.S. Data Center Power Demand Forecast
06 · Gartner's July Update: AI Servers Take the Lead
Gartner's newest figures show worldwide data center power demand rising 27% in 2026 to 132 GW, up from 104 GW in 2025, on a path toward roughly 290 GW by 2030. On electricity consumption, Gartner puts global data center use at 565 TWh in 2026, up from 447 TWh in 2025, and warns that total consumption could exceed 1,200 TWh by 2030 if current growth holds — a figure that sits above the IEA's own Base Case and worth flagging to anyone benchmarking against a single source.
The AI-specific breakdown is the sharper signal for planners: AI-optimized servers consumed roughly 95 TWh in 2025, are projected to draw 175 TWh in 2026 (an 84% jump), and are forecast to reach 258 TWh in 2027 — the point at which AI-optimized hardware consumes more electricity globally than every conventional server combined. AI-optimized servers already represent 31% of total data center power consumption in 2026, up from about 20% a year earlier.
AI-Optimized vs. Conventional Server Electricity Use (Global, TWh)
07 · Regional Breakdown
The United States, China, and Europe remain the markets every major forecaster expects to absorb most of the coming decade's growth, but the character of that growth differs sharply by region.
United States
McKinsey's data center demand model projects U.S. data center electricity consumption climbing to roughly 606 TWh by 2030, up from 147 TWh in 2023 — about 11.7% of total U.S. power demand. EPRI's pipeline-based analysis puts 2024 U.S. consumption at 177–192 TWh, rising to between 380 and 790 TWh by 2030 (9–17% of national generation), and BloombergNEF now projects the U.S. share reaching 12% by 2030 and 20% by 2035.
China
The IEA continues to identify China as the fastest-growing major electricity market overall, expecting it to account for close to half of the world's total electricity demand growth across all sectors through 2030, with data centers a meaningful part of that trajectory.
Europe
McKinsey projects European data center electricity consumption nearly tripling, from about 62 TWh today to more than 150 TWh by 2030 — roughly 5% of total European power consumption, up from about 2% currently — with capacity expanding from 10 GW to around 35 GW over the same period.
Ireland and Emerging Hubs
Concentration effects remain most visible in smaller grids. Data centers already account for more than a fifth of Ireland's total electricity demand, among the highest shares anywhere in the world, and similar dynamics are emerging across parts of Southeast Asia as cloud and AI infrastructure scales up.
➡️ Also Read: AI Data Centers and the Global Electricity Surge: Why Power Is Becoming the New Bottleneck of Digital Infrastructure
08 · Forecast Comparison Matrix
No single forecast should be read in isolation. The table below lines up every major projection side by side, with a confidence read on each based on methodology and how recently it was revised.
| Source | Metric | Target Year | Methodology | Confidence |
|---|---|---|---|---|
| IEA | ~950 TWh (global) | 2030 | Bottom-up energy system modeling | High |
| Gartner | 1,200+ TWh (global) | 2030 | Vendor shipment & server-density modeling | Medium |
| Goldman Sachs | 66 GW (US power demand) | 2027 | Facility-level pipeline tracking (Aterio) | High |
| EPRI | 9–17% of US electricity | 2030 | State-level project pipeline scenarios | High |
| McKinsey | 606 TWh (US); 219 GW (global capacity) | 2030 | Proprietary demand model | Medium |
| BloombergNEF | 20% of US electricity | 2035 | Chip deployment & capacity build-out modeling | Revised recently |
The spread is itself the finding. As the World Resources Institute has noted, published 2030 projections for U.S. data center electricity use alone range from roughly 200 TWh to over 1,050 TWh — a gap wide enough to represent the electricity use of tens of millions of homes. That range reflects genuine uncertainty about how much of today's announced project pipeline actually gets built on schedule, not disagreement about direction. For planning purposes, treat the IEA and EPRI figures as the more conservative, methodologically transparent anchors, and the Gartner and BloombergNEF figures as upside-case signals worth monitoring rather than baseline assumptions.
09 · BloombergNEF: The 2035 View Just Got Bigger
BloombergNEF's July 21, 2026 report is the sharpest upward revision of the year. The firm now expects U.S. data centers to account for about 20% of national electricity consumption by 2035, up from an estimated 5.9% today and roughly 12% by 2030 — a 2035 estimate 83% higher than what BloombergNEF itself forecast only seven months earlier. The firm attributes the jump to faster-than-expected AI compute deployment, with data center capacity approaching 200 GW nationally over the next decade and nearly half of that devoted to AI training and inference. By 2033, BloombergNEF expects the U.S. to host 64% of the world's AI chips by power demand, with Virginia and Texas seeing even higher local concentration than the national average.
BloombergNEF: US Data Centers' Share of National Electricity
10 · AI as the Dominant Driver
Across every major forecast, one theme repeats: accelerated, GPU-driven computing — not traditional enterprise servers — is now the primary source of new electricity demand. The IEA's 2025 data shows AI-focused facility electricity use growing three times faster than the data center sector as a whole, and Gartner's numbers point to AI-optimized hardware overtaking conventional servers in total electricity draw as soon as 2027. EPRI notes that AI workloads currently make up 15–25% of total U.S. data center electricity use, a share expected to keep rising as generative AI, image and video processing, and increasingly autonomous AI agents proliferate across consumer and enterprise software.
Goldman Sachs frames this as a structural shift in power density: facilities built for AI training and inference require far higher electricity draw per square foot than legacy data centers, and that density is expected to keep climbing as denser racks and more advanced chips come online. For capacity planners, this means historical per-square-foot power benchmarks are no longer a reliable guide for new AI-focused builds.
➡️ Read Also: AI Data Center Power Crisis 2026: How It is Impacting U.S. Real Estate
11 · Grid Risk Register
The binding constraint on AI infrastructure growth is no longer chip supply — it is the grid. Bank of America's analysts note that large gas turbines are essentially sold out through 2030, pushing developers toward on-site gas engines while utilities extend coal plant lifespans, deploy batteries, and pursue transmission upgrades; BofA estimates data centers alone could add roughly 125 GW of U.S. electric load between 2026 and 2030, lifting overall U.S. electricity demand growth to a 4.1% compound annual rate. Early July's U.S. heatwave already forced utilities to ask some data center operators to switch to backup generators — a live preview of the reliability tension both Goldman Sachs and EPRI have flagged as a defining feature of the next few years.
Risk Register: Probability & Impact by Region
| Region | Constraint | Probability | Impact |
|---|---|---|---|
| Virginia | Transmission & interconnection queue | High | Very High |
| Mid-Atlantic US | Generation capacity lag | High | Very High |
| Mid-Continent US | Turbine & equipment shortage | Medium | High |
| Northwest US | Grid reliability during peak load | Medium | High |
| Texas | Peak demand coincidence with heat events | Medium | Moderate |
| Georgia | Rate impact on residential customers | Low-Medium | Moderate |
12 · Scenario Range to 2035
Because AI adoption itself remains hard to predict, the IEA models several scenarios rather than a single number:
Base Case
Global data center electricity demand reaches approximately 950 TWh by 2030, close to 3% of global electricity demand.
Lift-Off Case
Assumes faster AI adoption and fewer supply-chain bottlenecks; demand could exceed 1,700 TWh by 2035, around 4.4% of global electricity consumption.
High-Efficiency Case
Assumes strong gains in chip performance, software optimization, and cooling technology outpace demand growth; consumption still reaches roughly 970 TWh by 2035 despite continued strong AI adoption.
Headwinds Case
Assumes slower AI deployment, tighter supply chains, and permitting delays; demand plateaus near 700 TWh, keeping data centers below 2% of global electricity demand.
Gartner's newer 1,200+ TWh estimate for 2030 sits above even the IEA's Lift-Off trajectory pace, and BloombergNEF's sharp July upgrade for the U.S. specifically suggests forecasters are still revising toward the higher end of this range rather than the lower one — a pattern worth tracking over the coming year rather than treating any single scenario as settled.
13 · Where Capital Is Flowing
The AI buildout is reshaping investment priorities across the power sector. Utilities and developers are directing capital toward grid expansion, transmission upgrades, renewable energy projects, battery storage, natural gas generation, and small modular nuclear reactors (SMRs) to meet the load coming down the pipeline. Technology companies, for their part, continue signing large long-term power purchase agreements to secure supply directly.
Where Investment Is Flowing
📘 Quick Definitions
TWh (terawatt-hour): one trillion watt-hours of electricity — roughly what a mid-sized country consumes in a year.
GW (gigawatt): a measure of power capacity; 1 GW can broadly supply several hundred thousand homes.
Interconnection queue: the waitlist a power project joins to physically connect to the grid, often the slowest step in bringing new capacity online.
14 · Stakeholder Action Matrix
The direction of this trend is no longer in question; the open questions are speed, sequencing, and who bears the transition cost. The table below sorts practical implications by urgency, so each stakeholder group can prioritize.
| Stakeholder | Immediate (0–12 months) | Structural (2027+) |
|---|---|---|
| Utilities & grid operators | Treat announced pipeline capacity as a signal, not a firm commitment — historical realization rates run 50–60%. | Build flexible interconnection and load-shaping agreements rather than planning to a single peak-load number. |
| Policymakers & regulators | Monitor rate impacts in high-concentration states (Virginia and emerging hubs) as new load comes online. | Accelerate permitting for transmission and generation while building in residential rate safeguards. |
| Investors & infrastructure financiers | Stress-test underwriting against the 200–1,050+ TWh forecast spread, not just the headline growth case. | Price in turbine and interconnection queue delays as a base-case timeline risk, not a tail risk. |
| Hyperscalers & AI developers | Secure power through long-term contracts and on-site generation alongside GPU procurement. | Sustain chip, cooling, and software efficiency gains — they're already offsetting a meaningful share of raw demand growth. |
| Local communities & grid customers | Track electricity rate filings directly through state utility commissions in high-concentration markets. | Engage in local planning processes before 2030, when concentration effects are projected to peak in several states. |
Frequently Asked Questions
How much electricity do data centers use today?
Globally, data centers consumed roughly 485 TWh in 2025 according to the IEA, and Gartner projects that figure will rise to about 565 TWh in 2026. In the U.S. specifically, data centers account for about 5.9% of national electricity consumption today, according to BloombergNEF.
Why do forecasts vary so widely between organizations?
Estimates diverge because they rely on different methodologies — some track announced project pipelines, others model chip shipments or historical growth trends — and because a large share of announced capacity never gets built on schedule. Published 2030 U.S. estimates alone range from about 200 TWh to over 1,050 TWh.
When will AI servers use more electricity than regular servers?
Gartner projects that AI-optimized servers will overtake conventional servers in total global electricity consumption by 2027, reaching about 258 TWh that year.
Which U.S. states face the biggest grid strain from data centers?
Virginia faces the most acute exposure, with data centers projected to consume 39–57% of the state's electricity by 2030 according to EPRI. Arizona, Indiana, Iowa, Nebraska, Nevada, Oregon, and Wyoming are also flagged as emerging high-concentration markets.
Is renewable energy keeping pace with AI's power demand?
Not entirely. While utilities and hyperscalers are investing heavily in renewables, battery storage, and small modular nuclear reactors, natural gas is expected to remain the dominant source of new generation in the near term, and Bank of America notes that large gas turbines are largely sold out through 2030 — a bottleneck pushing some developers toward on-site generation instead.
Bottom Line for Decision-Makers
Every organization studying the intersection of artificial intelligence and energy — the IEA, Gartner, Goldman Sachs, EPRI, McKinsey, and BloombergNEF — has revised its numbers upward within the past few months, and every revision has moved in the same direction: higher, and sooner than previously modeled. Data center electricity consumption is growing significantly faster than overall electricity demand, with AI-accelerated computing now the dominant source of new load growth.
The question for decision-makers is no longer whether AI will increase electricity demand — it already has, measurably, in a single year. It is whether utilities, grid operators, regulators, and energy developers can build the infrastructure needed to support this expansion without compromising reliability or affordability elsewhere on the grid. For investors, policymakers, and energy planners, this has moved from a niche technology issue to a central input in capital allocation and energy strategy — and the forecast range above should be treated as a working range to plan around, not a single number to plan to.
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
