EnergyMindAI

By Saad Iqbal

At around 4:00pm on July 25th, a single power line failed just outside Washington, DC. Lines fail somewhere on the American grid most days, and the system is built to shrug them off in the time it takes to blink. This one didn’t. More than three gigawatts of data centers detected the disturbance and, exactly as their safety systems are designed to, simultaneously cut themselves loose from the grid. Lights flickered as far away as the suburbs, and it took operators more than eleven minutes to restore a steady state — an eternity, in grid time, for something that should resolve in seconds.

Nobody lost power for long, and nobody made the evening news outside the trade press. But if you want to understand what artificial intelligence is doing to the American energy system this year, those eleven minutes are a better guide than any keynote. The grid was built for slow-moving, predictable demand. It’s now sharing space with a customer that can appear or vanish by the gigawatt in the time it takes a circuit breaker to trip. This week brought three concrete signs of how seriously the industry is taking that mismatch.

The Emergency Auction

Three days after the DC-area incident, PJM Interconnection — the grid operator covering 67 million people from Illinois to Virginia, and operator of “Data Center Alley” in Northern Virginia — announced it will run an unprecedented emergency capacity auction. The trigger: PJM’s mid-July auction for the 2028/2029 delivery year cleared 138,318 megawatts against a reliability requirement near 145,000, a shortfall of 6,831 megawatts, or roughly seven nuclear reactors’ worth of missing generation. It was the first time in PJM’s history the entire region has fallen short.

The new “Reliability Backstop Procurement” launches in September 2026, covering the delivery year beginning June 2028 — and its costs will be billed specifically to data centers rather than spread across all ratepayers. That structure echoes a proposal floated earlier this year by the White House and a bipartisan group of governors, who suggested letting tech companies bid directly on 15-year capacity contracts to fund the plants their AI ambitions require. PJM’s own filing names the cause: this year’s forecast added roughly 2,000 megawatts of peak demand, driven largely by “the continued trend of the addition of large data center loads.”

When the Grid Blinked

The July 25th event matters beyond one bad afternoon near the capital. Data centers currently account for roughly 3% of PJM’s total load; by 2040, PJM projects that could reach 24%. Every one of those facilities is built to snap onto backup power the instant it senses grid instability — sensible for a single building, but a real hazard when thousands of megawatts do it in the same second. Grid engineers have started calling these synchronized trip-offs “the canary in the coal mine,” and say events like it are becoming more frequent.

The fixes emerging are as interesting as the problem. ON.Energy is installing roughly 3 gigawatts of campus-wide uninterruptible power supply systems across four data center sites — letting a facility absorb a grid fluctuation internally, covering servers and chillers alike, rather than disconnecting and dumping the disturbance back onto the grid. Regulators are moving too: ERCOT, the Texas grid operator, is now writing “ride-through” requirements into its rules, obliging large loads to stay connected and weather a disturbance rather than bail at the first sign of trouble.

Oilfield Services Cash In

If grid operators are playing defense, the energy-services giants are playing offense — and getting paid for it. Baker Hughes reported second-quarter earnings this week that beat expectations across the board: 64 cents a share against a 50-cent estimate, on revenue of $6.74 billion, with adjusted EBITDA margins expanding to a record 18.3%. Leadership was explicit about why, telling investors that the rapid buildout of AI and data centers is producing a “step change” in global electricity demand and turning power generation into one of the firm’s biggest long-term bets.

The company backed the talk with action: it closed its acquisition of Chart Industries, expected to deliver $325 million in annual cost synergies within three years, booked $1.8 billion in new LNG equipment orders, and struck a deal with Mantle Reach Power for up to 500 megawatts of geothermal development — geothermal being an increasingly popular answer to the round-the-clock power profile AI data centers demand. Baker Hughes isn’t alone; rival SLB has spent the year building modular “AI Factory for Energy” infrastructure with NVIDIA. The companies that spent a century pulling hydrocarbons from the ground are repositioning, in real time, as the contractors who keep AI’s data centers electrified.

Put the three stories side by side and a single shape emerges: a grid operator forced into an emergency fix it’s never needed before, a physical demonstration of why that fix is overdue, and an industry retooling to profit from the gap between the two. None of this is a crisis — the lights stayed on in Washington, and PJM’s backstop auction is still more than a year from delivering power. But it’s a preview. AI’s most consequential infrastructure story this decade may not be written inside a data center at all. It will be written in substations, capacity auctions, and turbine orders — the unglamorous machinery straining to keep pace with a demand curve nobody quite planned for.