AI Data Center Power Demand Outruns Grid as Racks Near 600 kW

Power, not silicon, now sets the pace

For most of the AI buildout, the scarce input was the accelerator. That has changed. In the 2026 forecasts, rack specifications and grid filings point to electricity as the main limit on how fast AI capacity can grow. Two trends are driving this. AI hardware now packs far more power into each rack, and the grids serving the biggest data center hubs cannot add generation, transformers and interconnections fast enough to keep up. The result is a widening gap between the compute the industry has paid for and the power that can actually be delivered to it.

The numbers show how fast this is moving. The International Energy Agency says global data center electricity use rose 17% in 2025, and consumption at AI-focused facilities jumped 50%.2 Gartner estimates that AI-optimized servers used about 95 TWh in 2025 and will use roughly 175 TWh in 2026, about 31% of the 565 TWh it expects data centers worldwide to consume this year.20 TrendForce measures demand in capacity rather than energy. It puts global data center power demand at 161 GW in 2026, up 31% from 2025, with AI servers making up about a third of that total.16

Every forecast keeps moving up

The forecasts disagree on exact figures, but they have all been revised in the same direction. BloombergNEF now expects U.S. data center capacity to reach 194 GW by 2035. That is 83% above its December 2025 forecast and about 2.5 times its April 2025 estimate.3 On that path, data centers would use about 12% of U.S. electricity in 2030 and 20% in 2035, up from 5.9% today.15 A separate BNEF model built from expected AI chip deliveries reaches 207 GW of U.S. demand by 2033.6 Analysts tracking these updates note that BloombergNEF, Gartner and Bank of America all raised their numbers within a few weeks of each other this summer.4

The IEA remains the most cautious major forecaster. It expects global data center consumption to roughly double from 485 TWh in 2025 to 950 TWh in 2030, or about 3% of world electricity demand, and it describes the period after 2030 as having possible upside rather than a firm trajectory.2 Gartner's warning that consumption could top 1,200 TWh by 2030 sits well above the IEA's base case.4

On near-term AI-specific load, estimates are far apart. Uptime Institute expects only about 10 GW of global AI-related data center load by the end of 2026, and it attributes that low figure to grid and generation limits rather than weak demand. Deloitte's outlook puts global AI data center draw near 96 GW this year.11 Much of that spread likely comes from different definitions of what counts as AI load. But Uptime's reasoning makes the central point: the measurable load is limited by what the grid can supply.

Chip roadmaps are now grid roadmaps

The hardware side explains why the electrical problem is getting harder, not just bigger. Uptime Institute's annual survey found that the most commonly reported rack density was 11 kW in 2026.24 An Nvidia GB300 NVL72 rack, by contrast, draws up to 142 kW under Nvidia's reference architecture.24 Even published figures for that rack differ. Nvidia's Mission Control documentation lists 120 kW for GB200 and 135 kW for GB300, while Vertiv's infrastructure reference designs assume 130 kW and 142 kW.23 Across a 1,000-rack campus, that gap adds up to about 7 MW before any cooling overhead.23

The next generation goes further. Nvidia has not published an official power rating for Vera Rubin NVL72. Supply-chain reports put it between 190 and 230 kW.24 Schneider Electric cites figures up to 246 kW.27 Morgan Stanley research estimates about 234 kW per Vera Rubin rack and about 600 kW for the later Rubin Ultra design.14 The Kyber rack that will house Rubin Ultra is expected around 600 kW in the second half of 2027.24 Nvidia has said its 800-volt DC architecture is designed for 1 MW racks and above, starting in 2027.29

This is how AI workloads are changing chip and system design. The IEA says the power density of AI servers rose elevenfold between 2020 and 2025 and is set to quadruple again by 2027. By then, a rack the size of a refrigerator could draw as much peak power as 65 households.2 At those levels, traditional 54-volt power delivery inside the rack stops working, which is why the industry is moving to high-voltage DC.2228 At 1 MW, an 800 V bus carries about 1,250 amps. Delivering the same power at 54 V would take more than 18,500 amps.28 Silicon carbide and gallium nitride power devices handle different steps of that conversion chain.28 Air cooling has reached its limit. Direct-to-chip liquid cooling, which handles 100 to 150 kW per rack, held 55% of the liquid-cooling market in 2026.24

One point often gets lost. Denser racks do not mean more compute per megawatt within a generation. A GB200 NVL72 and a half-sized NVL36 both draw about 1.83 kW per GPU, so a megawatt of IT load supports roughly 545 GPUs either way.29 What density changes is the load on the building and the grid: more power concentrated on less floor space, and more strain on busways, cooling loops and substations.

The efficiency paradox

Efficiency gains are real, but they have not stopped the growth. The IEA reports that the energy needed per AI query has fallen sharply, while far more energy-intensive uses of AI are spreading.2 Brookings says the same thing: chips keep getting more efficient, but model size and query volume are growing faster, and the pattern from 2015 to 2019, when data center power stayed flat despite rising workloads, has reversed.7 Our reading is that efficiency now mostly determines how much AI each megawatt produces. It does not reduce how many megawatts the industry wants.

AI loads also behave differently from older ones. The IEA notes that server load in AI data centers can swing by more than half of rated capacity within a single second.2 The Los Angeles Times reported facilities rated at 1 GW briefly drawing as much as 1.5 GW.11 Utilities now have to plan for this volatility as well as for the size of the load.

The grid's side of the ledger

On the supply side, the strain is visible. The EIA reported that PJM set an hourly peak record of 162.6 GW on July 2, 2026, and that ERCOT reached 91.1 GW on July 22, 6.5% above its previous high.17 Texas is weighing more than 474 GW of large-load connection requests, about 90% of them from data centers, which is more than five times its record peak.17 Dominion Energy reports about 53.8 GW of contracted data center capacity in Virginia.17 PJM's market monitor has linked roughly 12 GW of added data center demand in 2026/27 to a doubling of capacity costs in the region.11

The equipment shortages are just as serious. Sightline Climate data reported by Bloomberg shows about 16 GW of U.S. capacity announced for 2026, but only about 5 GW under active construction. It expects 30% to 50% of the year's pipeline to be delayed or cancelled, with transformers, switchgear and batteries named as the main bottlenecks.18 Grid connections in some markets can take 24 to 72 months, while the data center itself can be built in under 18.18 Morgan Stanley's model puts new U.S. power needs at 97 GW for 2026 through 2028, against about 19 GW of available grid capacity and 21 GW under construction.14 TrendForce expects demand and grid supply to start diverging this year, with the gap growing sharply after 2028.16 Its global projection shows about 491 GW of demand against about 223 GW of grid capacity available to data centers by 2030.12

Real demand, inflated queues

The strongest counterpoint concerns how much of the demand is real. Wood Mackenzie counts 1,066 GW of data center connection requests on file with U.S. grid operators and expects only about 28% to be built. It calls the rest speculative, often the same project filed with several utilities.19 Relae data shows 46 projects worth more than $170 billion blocked, withdrawn or stalled since early 2024. One analysis argues that utility load forecasts filed with FERC exceed credible build-out estimates by about 40%.13

That does not mean the shortage is exaggerated. It means queue totals and announcements are poor measures of it. The firmer evidence is shipping hardware, record grid peaks and rising capacity prices, and all three point the same way. Our assessment is that the AI power crunch is real but will be settled project by project, not by headline gigawatt totals. Advantage will go to projects that have a power connection, equipment orders and a rack design to match the next GPU generation. BNEF's own framing supports this: whether its 194 GW forecast happens depends on how many projects secure customers, equipment and a firm grid position.3

For chip designers, the conclusion is that a product's real specification now includes the facility around it: the voltage it needs, how it is cooled, how sharply its load swings, and whether a utility can serve it before the hardware is out of date.