Data Centers Expected to Use 4x More Electricity by 2035

BloombergNEF published a report on July 21, 2026 projecting that US data centers will consume one-fifth of all electricity generated in the country by 2035 — four times their current share. That single figure would have been striking enough on its own, but the more telling detail buried in the report is that this estimate is 83% higher than what BloombergNEF itself predicted just seven months ago in December 2025.

The Claim

The BloombergNEF forecast centers on a surge in AI compute as the primary driver. The report projects US data center capacity will reach nearly 200 gigawatts over the next decade, with roughly half of that capacity devoted to AI training and inference workloads. The geographic concentration is also notable: by 2033, the United States is expected to host 64% of all AI chips measured by power demand, cementing its position as the global center of gravity for AI infrastructure.

The regional picture is stark. The PJM Interconnection — the grid spanning Virginia to Illinois, which already covers some of the most data-center-dense territory in the world — is expected to direct 34% of its electricity output to data centers. ERCOT, which covers most of Texas, faces a similar trajectory at 22%. These aren't marginal increases in load; they represent a structural transformation of what regional grids exist to do.

BloombergNEF isn't alone in revising upward. EPRI, the electrical industry research nonprofit, has more than doubled its 2024 estimate for data center electricity demand. S&P Global's forecast rose by more than a third between October 2025 and April 2026. Multiple independent organizations are all sprinting to catch up to a buildout that appears to be outrunning every model designed to track it.

What We See

The 83% upward revision in seven months is the most important number in this story, and it's worth sitting with. Forecasters don't revise by 83% because they refined a parameter — they revise that sharply when the underlying phenomenon is moving faster than the models assumed. Our read is that the real story here isn't the 2035 endpoint; it's the velocity. If BloombergNEF's own December 2025 estimate was already considered reasonable at the time, and it's now regarded as nearly half the likely reality, the honest position is that no one has a reliable model for where this lands.

The IEA's analysis, published separately, offers a useful global frame. It projects global electricity generation for data centers to grow from 460 TWh in 2024 to over 1,000 TWh by 2030 and roughly 1,300 TWh by 2035 — a figure broadly consistent with the BloombergNEF US-centric projection when you account for the US's outsized share of AI infrastructure. The IEA's work also adds texture on the supply side that the BloombergNEF report does not: renewables are expected to meet nearly half of the additional demand over the next five years, but coal currently holds a roughly 30% share of electricity consumed by data centers globally, with natural gas at 26% and renewables at 27%. The energy transition and the AI buildout are happening simultaneously, and they are in direct competition for the same grid capacity.

The PJM pricing signal is already legible in market data. Electricity prices in that interconnection rose 76% in the past year. That isn't a prediction or a model output — it's a recorded price movement in the most data-center-heavy grid region in the US. For developers and companies paying cloud inference bills, this is the mechanism by which the energy story eventually becomes a cost story. Cloud providers buy power under long-term contracts and can absorb short-term spikes, but a structural, decade-long demand shift of this magnitude will eventually flow through to API pricing. The direction of travel is set even if the timing is uncertain.

Where It Falls Short

The World Resources Institute raised a legitimate methodological concern in a September 2025 analysis: forecasting data center electricity demand is genuinely hard, and the extreme variance across estimates from credible organizations reflects real uncertainty in the underlying models, not just conservative versus aggressive assumptions. WRI noted that forecasts vary so widely in part because data center operators don't consistently report energy consumption, because the efficiency trajectory of AI hardware is contested, and because the buildout pace itself is a function of policy, permitting, and financing that can shift quickly. The BloombergNEF number is plausible and well-sourced, but treating it as a firm endpoint rather than a point on a range would be a mistake.

There's also a meaningful efficiency question left largely unaddressed in the primary forecast. AI chip generations are improving in performance-per-watt at a meaningful rate. If the next generation of training and inference hardware delivers a 2× or 3× efficiency gain — which is within historical range for this industry — total electricity demand could land substantially below the 200 GW headline even with the same volume of compute work being done. The forecast appears to weight capacity buildout more heavily than efficiency gains, and that framing may be more appropriate for real estate and grid planning purposes than for projecting actual consumption. The IEA's Base Case similarly treats efficiency improvements as a partial offset rather than a primary variable.

What neither source addresses directly is the feedback loop between energy cost and AI adoption rate. If grid congestion pushes US electricity prices high enough for long enough, some workloads will move offshore, some operators will delay buildouts, and some applications simply won't get built. The 1/5-of-US-electricity figure implicitly assumes demand grows to meet capacity, but demand is also elastic. The honest version of this forecast acknowledges that the number could be directionally right while still being wrong at the margins in ways that matter enormously for grid planning.

Sources