U.S. data centers consume 176 terawatt-hours annually—4.4% of all electricity . Goldman Sachs Research projects that figure will surge from 31 gigawatts in 2025 to 66 GW by 2027, more than doubling capacity in two years . The AI boom isn't a distant forecast anymore. It's showing up in grid operator warnings, utility earnings calls, and a scramble to redesign data centers before the power runs out.
Rampant power consumption threatens to expand AI's carbon footprint and create political backlash, with industry leaders warning of a basic constraint: the limits of power generation . "Very soon, maybe even later this year, we'll be producing more chips than we can turn on," Tesla CEO Elon Musk said earlier this year . The crunch is forcing a rethinking of how data centers are designed, built, and powered—and who pays for the grid upgrades to support them.
Can the Grid Handle What's Coming?
In the first quarter of 2026 alone, nearly 200 proposed data centers and other large energy users requested to join the Texas grid . Data centers and high-energy users are seeking a combined 438 gigawatts of power in Texas—more than five times the electricity used to power the entire state during record-breaking demand . Much of that is speculative, but even conservative estimates point to a collision.
ERCOT will vote Tuesday on a new evaluation process for data centers, known as "Batch Zero," as the grid operator tries to understand which projects are actually feasible and what infrastructure is needed to build them . Data centers will have to make a financial commitment to be evaluated before joining the grid, with the goal of filtering out speculative projects .
The pressure isn't limited to Texas. The EIA expects U.S. electricity use to grow by 1% in 2026 and 3% in 2027, marking the first time since 2007 that power demand has risen for four years in a row and the strongest four-year growth period since 2000 . The driving factor is increasing demand from large computing centers, with natural gas supply critical as electricity demand rises through 2027, driven largely by data centers, according to EIA Administrator Tristan Abbey .
Natural gas is filling the gap. The EIA forecasts natural gas consumed by the U.S. electric power sector will average 43.7 billion cubic feet per day during summer 2026, rising 6% to 46.1 Bcf/d in summer 2027—surpassing the previous record set in 2024 by 3% . Record-high consumption in 2027 is primarily driven by increasing electricity sales to commercial and industrial sectors in the West South Central and Mid-Atlantic regions, with demand growing from new data centers and large manufacturing facilities, particularly in Texas and Virginia , according to market data.
Who Builds the Power When the Grid Won't?
The wait times are untenable. Developers in Texas responding to a Bloom Energy survey typically expected to be connected to the grid about a year before the utility expected to have that power available, with the disconnect between power providers and data center developers only widening as demand increases .
So the industry is going around the grid. Fluence Energy shares soared more than 40% Monday after the company said it collaborated with Siemens and Nvidia to develop a design for AI data centers, with the blueprint incorporating a 136 megawatt facility that would include Fluence batteries . The reference design outlines a facility with total capacity of 136 MW and an IT load of 100 MW, covering the path from utility connection at 34.5 kV through medium-voltage distribution and modular low-voltage power blocks to the rack interface .
Fluence's Smartstack platform is intended to support functions including voltage and frequency ride-through, black start, demand response, and AI load smoothing, helping data centers maintain operations and manage power quality while scaling up . The design is a bet that battery storage can smooth the wild power swings AI workloads create. A single server's power load might fluctuate from 15kW to 30kW within seconds, scaling up to 50-100MW swings at the data center level—going from 20% of provisioned power to 150%—and traditional grid infrastructure can't handle that kind of swing, according to Bloom Energy tests .
Fuel cells are emerging as another workaround. Data center leaders expect approximately 30% of all data center sites to use some onsite power as a primary energy source supplemental to the grid by 2030 . Bloom Energy can deliver 50 MW of fuel cells in as little as 90 days and 100 MW in 120 days if gas supply and permits are in place, enabling operators to align power availability with aggressive project timelines . Fuel cells convert natural gas directly into electricity without combustion, achieving 15% to 20% higher efficiency than most open-cycle gas turbines or reciprocating engines .
The appeal is speed. A hyperscaler that waits seven years for a grid connection loses seven years of revenue from a 100+ MW AI training campus, which can generate $500 million to $1 billion in annual revenue at current market rates, with fuel cells that deploy in 90 days to 12 months unlocking that revenue stream years before grid-dependent alternatives come online .
