ERCOT Found Four Data Center Clusters That Can Each Dump 5,000 MW Off the Grid. That's 14 Times Its Largest Power Plant.

Five thousand megawatts. That's the demand a single cluster of Texas data centers could rip off the grid during a voltage disturbance, according to an internal ERCOT document dated May 21, 2026, first reported by Reuters, and four such groups now sit on the Texas grid, each capable, if it disconnected simultaneously during a grid fault, of erasing demand equivalent to Boston's entire electricity consumption.

ERCOT's largest single generation unit, one of the two reactors at Comanche Peak Nuclear Station southwest of Fort Worth, produces about 1,400 MW, and the entire grid planning apparatus is built around the assumption that losing a unit of this size is the worst plausible supply-side contingency. Grid planners build around losing it. Called N-1 contingency, the principle is simple: survive the loss of your biggest generator without rolling blackouts. But 5,000 MW of data center load vanishing in seconds isn't a generation loss. It's a demand-side crater. And it's 14 times larger than the biggest supply contingency Texas was designed to handle.

No steel mill, aluminum smelter, or chemical plant in the hundred-year history of electrification has ever presented this kind of risk to a grid operator. For a century, every major industrial customer from the copper smelters of Arizona to the paper mills of Wisconsin stabilized the grid just by existing on it, their synchronous motors acting as distributed shock absorbers that no utility had to install, maintain, or pay for. Data centers broke that compact, and ERCOT now has the numbers to prove it.

Why Every Other Industrial Load Helps the Grid

Foundries, paper mills, water treatment plants, automobile factories: all packed with synchronous motors. Massive spinning rotors coupled directly to the alternating current supply, their physical mass acting as a flywheel that absorbs transient disturbances before operators even know something happened. When grid frequency dips, those rotors resist the change, releasing stored kinetic energy back into the system and buying operators precious seconds to respond. Engineers call this rotational inertia, and for the entire history of electrification, grids got it for free from their customers.

Data centers, by contrast, contribute none of the rotational inertia that electrical grids have relied on since the first power stations were built. Their entire power chain runs through solid-state inverter-based electronics, from the utility feed through uninterruptible power supplies to the server racks. Australia's grid operator already classifies them as "large inverter-based loads." Up to 95% of a data center's energy passes through equipment with no rotating mass at all. When frequency drops, nothing pushes back.

It gets worse. Data centers are engineered to disconnect at the first sign of voltage trouble, which is a perfectly rational decision for any individual operator whose customers are paying for five-nines availability and whose SLA penalties for a thirty-second outage can exceed the cost of the electricity itself. Protection systems island the facility instantly. What data center engineers call "protecting the equipment," grid operators call "cascading abandonment during a system emergency." Since 2023, ERCOT has logged 26 separate events where data centers or cryptocurrency operations abruptly disconnected during grid disturbances.

December 2022: The Warning Nobody Heeded

A transformer failed in west Texas. Four hundred cryptocurrency miners and data centers disconnected within seconds, erasing 1,700 MW of demand. The grid, suddenly oversupplied, saw frequency spike dangerously. ERCOT scrambled to force 112 MW of generation offline to prevent equipment damage across the interconnected system, a manual intervention that took precious minutes during a window where automated protections were already fighting a surplus nobody had planned for. Generators designed to run at exactly 60 Hz were suddenly pushing power into a grid that didn't need it.

Traditional loads fail differently. A steel mill losing power just stops drawing. The grid notices the reduced demand and adjusts. But when hundreds of inverter-based loads disconnect in concert, the surplus hits all at once, frequency climbs, and generators trip on overvoltage protection. The correction creates its own cascading crisis, because the emergency response to surplus power requires shutting down generators that were moments ago keeping the lights on. In December 2022, the cascade involved 1,700 MW. ERCOT's current testing reveals that single clusters on the grid today, already interconnected and drawing power, could dump three times the 2022 amount in a single correlated disconnection event, turning what was a manageable nuisance into a system-threatening contingency that existing reserve margins cannot absorb.

The Original Calculation: Stability Cost per Megawatt

No grid anywhere in the world currently prices the stability liability that data centers impose. Consider two facilities drawing 500 MW each. An aluminum smelter at that scale provides rotational inertia through its pot lines, rides through voltage sags as its motors absorb the disturbance, and ramps consumption in predictable patterns that operators can model hours in advance. A data center at 500 MW provides exactly none of these services, may disconnect without warning during the moment the grid most needs stable load, and concentrates enough capacity in a single site to create a demand-side contingency that N-1 planning never contemplated.

Both facilities pay identical rates per megawatt-hour, yet the stability costs they impose on the shared grid are wildly different.

Here's the math on total exposure. ERCOT's four high-risk groups each threaten 5,000 MW in simultaneous trip risk, yielding 20,000 MW of aggregate demand-side contingency. Against ERCOT's approximately 150 GW of total installed generation capacity, that demand-side risk from data centers alone represents 13% of system capacity, concentrated entirely in loads that provide zero inertia and have repeatedly demonstrated willingness to disconnect without coordination. Grid planning has no precedent for demand-side risk at this scale, because the entire discipline was built around supply-side failures: a generator trips, a transmission line goes down, a fuel pipeline is disrupted, and the system compensates by dispatching reserves or shedding load in a controlled sequence that operators have rehearsed for decades. Generation-side contingency reserves, typically sized at 3,000-4,000 MW for ERCOT, are not designed to absorb surplus from a 5,000 MW demand cliff.

Ireland at 21%: Where This Road Ends

Ireland is already living in the future this trajectory creates. Data centers now consume 21% of the nation's total electricity generation, according to an analysis reported by the Wall Street Journal on June 7. In Dublin and County Meath, the share exceeds 50%. Housing estates in Dublin's northern suburbs experience blackouts during storms because the grid is stretched thin keeping data centers online.

Dublin's response was a "Bring Your Own Power" mandate requiring new data center operators to self-supply a portion of their electricity, effectively acknowledging that the shared grid model cannot absorb unlimited inverter-based load without fracturing. The White House has signaled interest in similar provisions. Maine went further, imposing a blanket moratorium on any new data center exceeding 20 MW until November 2027, giving its grid operator and legislature time to develop interconnection standards that account for the stability costs that existing rules ignore entirely, a regulatory pause that would have been unthinkable five years ago when states were competing to attract data center investment with tax breaks and expedited permitting.

These are not NIMBYism or anti-tech sentiment dressed up in safety language. They are engineering admissions, arrived at independently on three continents, that the shared electrical grid was designed around loads that contribute stability in proportion to their consumption, and that when the fastest-growing load class only takes without giving back, the social contract of shared infrastructure buckles under the asymmetry.

The Strongest Counterargument

Data centers offer genuine flexibility that traditional industrial loads cannot. Amazon, Google, and Microsoft all run demand response programs that voluntarily curtail consumption during grid emergencies. Many facilities operate onsite battery energy storage systems and diesel backup generators capable, in theory, of providing power back to the grid during shortfalls. The industry's investment in co-located BESS is accelerating. A data center equipped with grid-forming inverters could theoretically provide synthetic inertia, replacing the rotational inertia that its servers don't generate mechanically.

The qualifier is "if required." Current interconnection standards were written for a world where loads inherently stabilized the grid. They don't mandate that new loads replicate what old loads provided for free. Fixing this gap requires regulators to classify data centers as what they empirically are: a fundamentally novel load class with grid impacts that traditional industrial frameworks were never designed to address.

What This Analysis Did Not Prove

ERCOT's full internal report remains non-public. This analysis relies on Reuters' and Data Center Knowledge's reporting of the May 21, 2026 document; the four groups that failed voltage ride-through testing are unnamed. Grid inertia calculations require system-specific dispatch data that ERCOT does not publish in granular form, so the 13% demand-side exposure figure uses installed capacity as the denominator, not real-time dispatch, which would make the percentage higher during low-generation periods. Ireland and Texas operate structurally different grids: ERCOT is electrically isolated from the rest of North America, while Ireland interconnects with the United Kingdom via subsea cables. Direct comparison of destabilization dynamics between the two requires caution.

What You Can Do About It

If you pay an electricity bill in ERCOT territory, you are subsidizing grid stability costs that data centers create but don't pay for. The mechanism is indirect: when data center disconnections force emergency generator curtailments or trigger ancillary service deployments, those costs are socialized across all ratepayers. Support ERCOT's push for stricter voltage ride-through requirements and advocate for differential stability pricing that charges inverter-based loads for the inertia they don't provide.

If you work in data center operations, push your interconnection team toward mandatory VRT compliance and grid-forming inverter technology. Facilities that can demonstrate synthetic inertia provision will face fewer permitting obstacles as regulators tighten requirements. Being early is a competitive advantage, not a cost center.

If you're evaluating AI infrastructure investments, know that 30-50% of planned U.S. data centers are currently delayed or cancelled. The bottleneck is not silicon, not capital, not even land. It is transformers, grid interconnection agreements, and the growing realization among grid operators that the electrical system was not built for loads that fight back.

The Bottom Line

For a century, every major industrial customer acted as a shock absorber for the electrical grid. Spinning motors smoothed frequency, rode through voltage sags, gave operators time to respond. Data centers broke that symmetry. They draw power through inverters that contribute zero inertia, disconnect at the first voltage dip, and concentrate enough load in single clusters to create demand-side contingencies 14 times larger than what the grid was designed to survive. ERCOT tested it. The answer came back ugly. The grid isn't running out of electricity. It's running out of loads that play nice.