The Grid That AI Built—And Broke

Four big data-center and crypto projects failed Texas grid voltage tests, exposing how fast AI demand is outrunning ERCOT's stability safeguards. The loss was big enough to rival a city the size of Boston.

The Electric Reliability Council of Texas said in a report dated May 21 that unnamed large electricity users abruptly disconnected during a test of how they would handle routine voltage disturbances. When a facility trips offline during a grid fault, it doesn't just go dark—it destabilizes the entire system. When large customers abruptly cut their power use, it can knock the grid off balance and trigger wider outages.

The failures arrive at an awkward moment. Global electricity consumption for data centers is projected to double to reach around 945 TWh by 2030 in the IEA's Base Case, representing just under 3% of total global electricity consumption.

The US data center sector is moving from roughly 180 TWh today toward 400-600 TWh around the end of the decade. That's not a forecast—it's a collision already in motion. And the grid, most of which was built between the 1950s and 1970s, with approximately 70% approaching the end of its life cycle, wasn't designed for it.

Can Natural Gas Fill the Gap Fast Enough?

The scramble for power has turned into a scramble for turbines. The price to build a new combined cycle gas turbine power plant has risen from less than $1,500 per kilowatt of generating capacity in 2023 to $2,157 last year, according to a report.

What's more, it now takes 23% longer to complete a new facility. The bottleneck isn't just construction—it's equipment. By the end of this year, prices for gas turbines, which constitute up to 30% of a new power plant's cost, are expected to be up 195% over 2019 prices.

Waitlists are stretching into the early 2030s.

Natural gas has become the default answer to AI's power hunger. With a share of over 40%, natural gas is currently the biggest source of electricity for data centers in the United States. As demand growth is particularly rapid over the next five years, natural gas is the largest source of additional supply, adding over 130 TWh of annual generation until 2030. In early 2026, the trend accelerated. In January, Pacifico Energy's GW Ranch in West Texas became the largest approved gas power project in the country when the Texas Commission on Environmental Quality granted an air permit of up to 7.7 gigawatts of generation by natural gas turbines to power a private grid supporting data centers.

NextEra also secured approval for two large natural gas plants in Texas and Pennsylvania with a combined 10 GW of power to be connected to their respective regional transmission grids.

But gas infrastructure has its own constraints. Current filed plans for data centers are around 125GW in the US—representing a staggering 50% or more increase in gas demand that our distribution grid is not ready for. The pipes can't keep up with the chips.

Could AI Save More Energy Than It Consumes?

The irony is sharp: the same technology straining the grid could be the thing that fixes it. JLL's Hank platform analyzes occupancy and external data to optimize heating, ventilation and air-conditioning, cutting energy use by 20% while supporting comfortable conditions for building users.

A meta-analysis reveals that reinforcement learning achieves highest energy savings at 22.3% in smart buildings, followed by hybrid methods at 28.1%.

By accelerating high-efficiency and net-zero buildings, AI could cut energy and emissions by 40-90% by 2050 combined with adequate policies, according to research published in Nature Communications.

The question is whether those savings can scale fast enough to offset the data centers themselves. Buildings in the EU account for nearly 40% of energy consumption and approximately 36% of greenhouse gas emissions, according to a 2020 European Commission study. If AI can optimize even a fraction of that footprint, the net energy equation starts to look different. But the timeline matters. Connecting a new facility to the power grid can take 4-10 years, while AI data centers are typically planned and built within two to three. The infrastructure lag is the real constraint.

Some researchers are exploring flexibility instead of brute capacity. Peter Zhang, assistant professor at Carnegie Mellon's Heinz College, is exploring whether dynamic workload adjustments would help stabilize the electricity demand profiles of data centers, and what it would take to incentivize a shift to an overnight schedule. He hopes nocturnal data centers could reduce AI's strain on the country's aging energy grid.

In 2026, data centers will play a more active role in stabilizing the grid and mitigating cost increases by securing strategic investment and promoting load flexibility via load shedding or curtailment.

The U.S. Department of Energy is betting on simulation. The DOE has launched a test platform called Agora to simulate one of the biggest infrastructure collisions in the country: hyperscale AI campuses connecting to an already strained electric grid. The platform replicates the electrical behavior of large data centers, including the volatile, high-density power demands reshaping utility planning across the US.

DOE announced 26 challenges, including using AI to improve power grid planning and interconnection at up to 20-100 times faster decision-making speeds.

What Changed This Week

Texas moved from theory to enforcement. ERCOT's board approved two landmark sets of rules that would shape the future of data centers in the state if finalized. One package establishes new criteria and processes for bringing big electricity users onto the grid by reviewing them in batches. The other rules require data centers and cryptocurrency-mining facilities to stay online during brief grid disruptions.

Since 2023, ERCOT has identified at least 26 events in which data centers or crypto mining facilities have abruptly disconnected from the grid because they could not handle disturbances in the flow of electricity. The voltage ride-through failures are no longer edge cases—they're a systemic risk heading into summer peak demand.

What to Watch

ERCOT's Batch Zero process will determine which projects get power first and which wait years for transmission upgrades. ERCOT expects about 100 GW's worth of projects to meet the approved criteria to be part of Batch Zero. The Texas Public Utility Commission will review both the batch process and the voltage ride-through requirements in the coming weeks—decisions that could set national precedent. Watch for EIA's pilot studies on data center energy demand, expected later this summer, which will provide the first granular federal data on AI electricity consumption by region. And track natural gas turbine delivery timelines: if waitlists stretch past 2030, the entire AI infrastructure buildout timeline shifts with them.