Grid connections, on-site generators, and substations working in parallel — the new normal for data center power infrastructure in 2026.
Power, not chips, is now the thing standing between the AI industry and its next growth phase. Land is available, capital is abundant, and GPUs keep shipping — but the electrons to run them are stuck behind interconnection queues, transformer backlogs, and a grid that was never designed for this kind of load. Data center power infrastructure has become 2026's defining bottleneck, and the numbers behind it are staggering.
How Much Power AI Actually Needs, By the Numbers
Every major forecaster revised its numbers upward in 2026, and they all point the same direction: demand is growing faster than anyone expected twelve months ago.
| Source | Forecast | Timeframe |
|---|---|---|
| BloombergNEF | 118 GW of US data center capacity (52% above its December 2025 outlook) | By 2030 |
| BloombergNEF | 194 GW of US capacity; up to 207 GW in an unconstrained AI-chip scenario | By 2035 / 2033 |
| Goldman Sachs Research | US data center power demand more than doubles, from 31 GW to 66 GW | 2025 to 2027 |
| McKinsey | Data centers drive ~75% of US power demand growth; national capacity need of 30–55 GW | Through 2030 |
| IEA | Global data center electricity use grows from ~460 TWh to over 1,000 TWh | 2024 to 2030 |
| Rystad Energy | Global installed capacity rises from 141 GW to 262 GW, with a pipeline above 620 GW | End-2025 to 2030 |
| Bank of America | 125 GW of new US data center load, against only 93 GW of utility capacity additions | Through 2030 |
Modeled from Goldman Sachs Commodities Research (Hongcen Wei, Daan Struyven, Samantha Dart), May 2026.
Perhaps the most telling figure is the demand share of the grid itself: Goldman's research shows data centers moving from 4.1% of US peak summer demand in 2025 to 8.5% by 2027 — a doubling of load concentration on infrastructure that upgrades on a multi-year, not multi-month, timeline.
Why the Grid Can't Keep Up
Interconnection queues are the single biggest chokepoint developers describe in 2026. Regions like Northern Virginia, Texas (ERCOT), and the PJM footprint are seeing transmission delays and thin headroom stack up simultaneously. PJM's capacity auction prices are the clearest market signal of the strain, having spiked past $300/MW-day as reserve margins tighten.
Hardware lead times compound the problem. Large power transformers — the unglamorous devices that step voltage up and down between the grid and a facility — now carry lead times around 128 weeks, and generator step-up units run even longer. A data center can be built faster than the transformer it needs to energize it.
Interconnection queuesMulti-year waits in constrained regions; grid operators are rationing available headroom. |
Transformer lead times~128 weeks for large power transformers, longer for generator step-up units. |
Auction price spikesPJM capacity prices above $300/MW-day reflect tightening reserve margins. |
Historical ceiling~10 GW is the largest annual grid connection volume on record — far below near-term demand. |
For a closer look at how these constraints translate into real operational risk, see our breakdown of data center grid connection problems and how they're delaying project timelines across major US markets.
The Rise of On-Site and Behind-the-Meter Generation
Faced with multi-year interconnection timelines, developers are increasingly bypassing the grid queue altogether — at least partially. On-site generation, using gas engines and turbines, has moved from a backup strategy to a core part of the power plan. Bank of America tracks more than 7.5 GW of on-site projects already under construction, with tens of additional gigawatts sitting in pre-construction planning.
Behind-the-meter capacity is expected to scale meaningfully by 2028, effectively creating a parallel power track that runs alongside — rather than through — the traditional utility interconnection process. This shift changes the risk profile too: outages and reliability now depend as much on a facility's own generation assets as on grid stability. That's a topic worth its own deep dive if you're evaluating AI data center power outage and grid risk in 2026.
Rack Density Is Rewriting the Power Playbook
The power problem isn't only about how many gigawatts a region can deliver — it's also about how much power a single rack can draw. Traditional enterprise racks ran 5–15 kW. AI and HPC racks now routinely exceed 25 kW, with some configurations pushing past 100 kW.
That jump is driving a shift toward higher-voltage DC architectures — 800 VDC systems are gaining real traction, backed by initiatives from Vertiv and NVIDIA-aligned hardware partners, because they cut conversion losses at these extreme densities. Modular, factory-built power and cooling blocks are also becoming standard, letting operators assemble multi-gigawatt campuses in repeatable units rather than one-off custom builds.
We've charted this shift in detail in our piece on GPU rack power density evolution from 2020 to 2027, which is worth bookmarking if you're sizing electrical infrastructure for next-generation deployments.
Other Bottlenecks Slowing Deployment
Power dominates the conversation, but it isn't the only constraint. Construction costs are climbing alongside demand for skilled electrical and mechanical labor, and EPC (engineering, procurement, and construction) capacity is stretched thin across every major market. Community opposition is also a growing factor — driven by concerns over higher local electricity prices, water consumption for cooling, and strain on regional grids. Taken together, some forecasts estimate that 30–50% of planned 2026 capacity is at risk of delay or outright cancellation.
How Hyperscalers and Utilities Are Responding
The response from both sides of the meter is aggressive. Hyperscalers including Microsoft and Google are locking in long-term power purchase agreements, restarting retired nuclear assets, and funding dedicated generation projects rather than waiting in line. Utilities, for their part, are extending the operating life of existing plants, deploying batteries for peak shaving, and prioritizing transmission upgrades in the most constrained corridors.
Further out on the horizon, carbon capture retrofits and early-stage nuclear and microgrid concepts are entering utility and developer pipelines — early bets on solving the power problem structurally rather than patching it project by project.
What This Means Going Forward
Every major research house — BloombergNEF, Goldman Sachs, McKinsey, the IEA, Rystad, and Bank of America — is converging on the same conclusion from different angles: demand trajectories through 2030 are relatively certain, but the infrastructure to meet them is not. Power, not chip supply, is likely to remain the binding constraint on AI infrastructure deployment through at least the end of the decade.
For developers, operators, and investors, that makes power availability — not just land or capital — the real due-diligence item on any new data center project. If you're modeling out a project's power requirements, our data center power calculator is a practical starting point for sizing load against realistic grid and on-site generation scenarios.
Related Reads
Figures referenced are 2026 projections from BloombergNEF, Goldman Sachs Research, McKinsey, the IEA, Rystad Energy, Bank of America, Vertiv, EPRI, and supporting analysis from SemiAnalysis, ABI Research, and Utility Dive. Figures are subject to revision based on actual build rates, policy, and equipment availability.
