Subscribe to Our Digest
Goldman Sachs put a figure on what grid operators have been feeling since late 2024: US data-center power demand is expected to reach 66 gigawatts by 2027, up from 31 GW in 2025. No grid was designed for that rate of change.
The May 2026 Goldman Sachs Research forecast puts specific annual waypoints on what had previously been expressed as wide ranges. The 31 GW to 41 GW to 66 GW trajectory implies data centers will account for 8.5 percent of US peak summer electricity demand by 2027, up from 4.1 percent in 2025. At the 70 percent utilization assumption the firm uses, installed capacity reaches 95 GW by end-2027.
That 95 GW installed figure matters for grid planners. Installed capacity must be built, energized, and connected before demand arrives. Given transformer lead times above 128 weeks and interconnection queues stretching to 2031 in some zones, the gap between forecast and deliverable is the central tension in data-center infrastructure planning right now.
The Electric Power Research Institute took a longer view in its February 2026 Powering Intelligence report. EPRI projects data centers could consume 9 to 17 percent of US electricity by 2030, up from roughly 4 to 5 percent today, a jump from 177 to 192 TWh consumed in 2024 to 380 to 790 TWh by 2030. The width of that range reflects genuine uncertainty about inference efficiency improvements and whether capex commitments translate into commissioned capacity on schedule.
The revision size is the key signal: the 2026 EPRI estimate is 60 percent higher than its own 2024 projection. Institutional forecasts from two years ago have been overtaken by actual orders. The right response to any current forecast is to treat the high end as the planning case.
Virginia illustrates state-level concentration risk. EPRI estimates data-center demand could account for 39 to 57 percent of Virginia’s total electricity consumption by 2030. A grid majority-dependent on a single industry’s bursty load profile creates planning challenges utilities have not previously encountered.
The DOE and Lawrence Berkeley National Laboratory’s December 2024 report established the pre-acceleration baseline: data centers consumed approximately 4.4 percent of US electricity in 2023, projected to reach 6.7 to 12 percent by 2028 at 325 to 580 TWh. Given that the report predates the 2025-2026 acceleration visible in interconnection filings, those figures are likely the floor.
The DOE framing also disaggregates by facility type. Hyperscale cloud campuses account for the largest share of new capacity, but the enterprise colocation segment is growing faster on a percentage basis. Colo operators face interconnection timelines their hyperscaler counterparts can sometimes bypass through direct utility negotiations.
The Grid Strategies National Load Growth Report projects five-year peak demand growth of 166 GW at 3.7 percent annually, with data centers responsible for roughly 90 GW of that total. Sixteen GW-scale data-center projects totalling around 30 GW are slated for commercial operation in 2026-2027.
The same report flags that utility load forecasts submitted to RTOs may overstate near-term data-center additions by approximately 25 GW. The mechanism: a developer files interconnection requests with multiple utilities simultaneously. Each records the full load. The result is systematic double-counting without reflecting actual committed buildout. Demand is real; timing and geographic distribution may be less smooth than aggregate numbers suggest.
Bloom Energy’s 2026 Data Center Power Report estimates total IT load capacity at approximately 80 GW in 2025 growing to around 150 GW by 2028. The gap between IT load capacity and grid draw reflects PUE ratios and the growing share of behind-the-meter generation that does not appear in utility consumption data. As BTM gas and fuel cells account for more of the power mix, utility-side measurements will increasingly undercount true data-center energy consumption.
Why does the Goldman Sachs forecast jump so sharply from 2026 to 2027?
The step reflects the anticipated commissioning of a large wave of hyperscale facilities that broke ground in 2024-2025. Construction timelines run 18 to 30 months, so 2023-2024 capex announcements translate into energized capacity in 2026-2027. The slope mirrors the construction pipeline.
How reliable are these forecasts given they keep being revised upward?
Every major institution has revised data-center demand projections upward since 2022, consistently in the same direction. Treating the upper end of current ranges as the central planning case is more defensible than anchoring on midpoints that have repeatedly proven too low.
What does the Grid Strategies 25 GW overcount warning mean in practice?
It means some portion of the 166 GW five-year peak growth forecast represents the same physical facility counted in multiple utility service territories simultaneously. Actual demand may be lower in aggregate but more concentrated in the zones where projects ultimately land.
How does Virginia’s 39-57% data-center share of state demand affect grid reliability?
A grid where a single industry class may represent more than half of peak demand faces unusual volatility risk. AI training workloads are bursty and can shift significantly between data centers. Virginia grid planners must account for demand correlations that do not exist in more diversified load profiles.
What is the difference between IT load capacity and grid power draw?
IT load capacity is total compute power that installed servers can consume at peak. Grid power draw is what is actually purchased from the utility, which is lower due to sub-100% utilization rates and behind-the-meter generation. Bloom Energy’s 80 GW IT load estimate against Goldman Sachs’s 31 GW grid draw implies significant on-site generation across the installed base.
Facilities expecting to energize in the 2026-2027 window should be audited now against their utility interconnection dates and transformer delivery schedules. Both are documented in Articles 1 and 5 of this series.
For capacity planners inside large enterprises, the 60 percent upward revision in the EPRI estimate between 2024 and 2026 justifies escalating data-center power as a board-level infrastructure risk. Standard IT planning cycles are too slow for this problem.
Track the Grid Strategies overcount figure quarter by quarter. When RTO planners publish updated load forecasts, the delta between submitted interconnection requests and actual commercial operation dates will reveal where demand is real and where it is hedging. That intelligence has direct value for site selection.