Can the US grid absorb the data-center decade?
The question of the decade has an arithmetic, and this report runs it from the primary records. On one side, the load: data centers took 176 TWh — 4.4% of US electricity — in 2023, and the national laboratory that counts them projects 325 to 580 TWh by 2028. On the other side, the machine that builds supply: a fleet of 1,405 GW, a build rate that just set its modern record at 57 GW in 2025 — and an interconnection queue holding 2,363 GW, 1.7 times the entire existing fleet, of which history says only about 15 cents on the dollar ever gets built. The verdict view runs the collision honestly, floor first. Every figure is computed from the records beside this page.
published 2026-08-28 · data cutoff 2026-08-28 (LBNL queues 2025 edition · EIA-860M June 2026) · edition 1 — updated only by a new dated edition · sources: LBNL · EIA · SEC EDGAR
The wedge, measured then projected.
What the demand side actually weighs. The grid's observed all-time peak — 761 GW, set in the July 2025 heat — against the data-center load: measured at 176 TWh in 2023, projected by LBNL to 325–580 TWh by 2028. In average-power terms the wedge is 17 to 46 GW of new around-the-clock load in five years — a band whose width is itself a finding: the laboratory that counts this industry will not commit to better than 2×.
2,363 gigawatts are waiting. History says 15% arrive.
Every power plant must stand in line for a grid connection, and the line is now 1.7 times the size of the entire operating fleet — 2,363 GW active across the country's interconnection queues, dominated by solar (1,089) and batteries (780). The waterfall below is the report's central exhibit: what entered the queue in the mature cohorts, what withdrew, and the 14.8% that actually reached operation — after a median of 3.1 years spent in line.
2025 set the modern record. The decade needs it every year.
What the machine has actually delivered, year by year, from the generator inventory itself: the all-time record is 70 GW in 2002 — the gas-turbine boom — and 2025 just posted the renewables-era record at 57 GW. Ahead stands the EIA's planned list: 286 GW with in-service dates, front-loaded into 2026–2028. The solid bars are history; the hollow bars are the promise; the dark line is what retired.
The floor clears. The ceiling does not — not from the planned list alone.
At the floor, yes — with room. If the data-center decade lands at LBNL's low case, the wedge is ~17 GW of average load by 2028, and even a pessimistic read of the planned list delivers roughly twice that in average-equivalent supply. At the ceiling, no — not without the queue moving faster than it ever has. The high case demands ~46 GW around the clock — more than the planned list yields under any slippage assumption short of heroic. And the national arithmetic hides the real bind: the load concentrates where the queue is thickest — PJM alone holds 256 GW in line, after a median wait of over three years.
Where every number comes from.
Three primary records, one computed collision. The demand wedge is LBNL's United States Data Center Energy Usage Report (2024) — measured history to 2023, published range to 2028, drawn as published and never averaged. The supply machine is EIA-860M (June 2026): every operating, retired, planned and canceled generator in the country, with nameplate capacity and vintage. The queue is LBNL's Queued Up (2025 edition): 33,256 interconnection requests with status, fuel, capacity and dates. All three sit in data.json beside this page.
The completion rate is measured, not assumed. 14.8% is the capacity-weighted share of 2000–2018 queue entries that reached commercial operation — cohorts old enough for their story to be over. Recent cohorts are still undecided and are never counted as failures. The 76.9% withdrawal rate is the same cohorts, same weighting. The median 3.1 years is time from queue entry to operation, for the projects that made it.
Average-equivalent gigawatts. Nameplate capacity is not around-the-clock power. The verdict converts the planned list with standard capacity factors (solar .25, wind .35, offshore .40, gas .55, nuclear .90) and counts batteries at zero new energy — they move energy in time. The wedge needs no conversion: TWh ÷ 8.766 = average GW.
What this page does not do. It does not project electrification, manufacturing or export load — the wedge here is data centers only, which makes the verdict conservative on demand. It does not assume queue reform. It does not net the retirement schedule beyond drawing it. And it does not resolve the locational question — a national yes can still be a Virginia no; that is stated, not solved.