Texas lost 33 gigawatts at sunset. Nothing happened.

On the evening the Texas grid set its all‑time demand record, the entire solar fleet switched off. Four years ago that is the sentence that ends with people being asked to turn off their air conditioning. This time the lights stayed on, and the reason is a machine that barely existed in 2021.

At 6 PM on Wednesday 22 July 2026, the Electric Reliability Council of Texas was serving 91,075 megawatts. That is the most electricity Texas has ever used, beating the previous record of 85,432 MW set in August 2023 by almost six gigawatts.

Records get set in heatwaves. What makes this one worth a look is what happened next. Over the following four hours, while demand was still within a few per cent of its all‑time high, the largest single source of supply on the system went to zero.

The sun clocks off

Solar peaked at 33,484 MW at 1 PM — 39.2% of everything Texas was consuming. By 10 PM it was producing 1 megawatt. Not one gigawatt. One megawatt, which is roughly a large office building.

The totals undersell how fast it goes. In the single hour between 7 and 8 PM the solar fleet lost 12,355 MW — that is 206 megawatts every minute, sustained, for sixty minutes.

That is not a failure. It is just what solar does, every single day, and the grid has always known it was coming. The interesting question is never whether the sun sets. It is what fills the hole, and how quickly.

Something had been eating the surplus

Go back to the morning. At 10 AM, with solar climbing through 29 GW and demand still relatively modest, the battery column in the data reads −10,002 MW.

Negative, because it is a load. Ten gigawatts of batteries were pulling power off the grid and putting it into cells, at exactly the moment when Texas had more sunshine than it could use.

Eleven hours later, at 9 PM, the same column reads +11,316 MW. The batteries were the single biggest thing standing between a record evening peak and a problem.

The grid didn’t get bigger. It got a memory.

The whole day, hour by hour

This is the shape of it. Every figure below is measured, straight out of EIA’s Hourly Grid Monitor. Battery figures are signed: negative is charging, positive is discharging.

Three things are worth staring at in that table. Gas ramps hard, from 33.6 GW at midday to 47 GW by 9 PM — it is still doing most of the heavy lifting, and pretending otherwise would be silly. Wind quietly triples after dark, which is the long‑standing Texas pattern. And nuclear, which is not in the table because there is nothing to see, moved a grand total of 16 megawatts across the entire day: 4,952 to 4,968. It is the flattest line in the dataset.

What would have happened instead

“Nothing happened” is true, and it is a deeply unsatisfying place to stop. The question anyone actually has is what the alternative looked like — and that is not in the file. It has to be modelled, so here is the model, in the open.

Hold the same per‑megawatt utilisation and shrink the fleet to the size it was at the end of 2023. The 9 PM discharge of 11,316 MW becomes 2,774 MW. That is a 8,542 MW hole at nine in the evening on the hottest day of the year.

Two things stop that being a blackout headline, and both are checkable. The gap lands at the hardest hour gas worked all day — 47,031 MW at 9 PM, its daily maximum. And gas still had roughly 22,000 MW of nameplate headroom above that. So the honest statement is the size of the hole and who would have had to cover it, not that the lights go out. There is a gate in the build script that fails if either of those two facts ever stops being true.

Four thousand to seventeen thousand

None of this was possible three years ago, and you can prove that from the generator inventory rather than from anyone’s press release.

Take every operating battery in ERCOT, group by the year it came online, and add it up. At the end of 2023 — when Texas set its previous demand record — the state had 4,174 MW of grid batteries. Today it has 17,025 MW. That is a 4.1× increase in thirty months, and it happened almost entirely without anyone outside the industry noticing.

Two things this article nearly got wrong

The build script for this piece runs eight checks against the raw files and refuses to produce anything if one fails. Two of them caught real errors in the first draft, and both are worth naming, because they are the kind of mistake that makes a great‑sounding number.

1. The 2024 zero that wasn’t

The obvious comparison is “on the last record day the batteries did nothing.” The data appears to support it: on 20 August 2024, ERCOT’s previous record day, the battery column reads exactly zero at the peak hour.

It reads zero because EIA did not break ERCOT battery storage out into its own column until 23 October 2024. Before that date the number is not zero, it is absent. Publishing that comparison would have been a fabrication dressed as a dataset. The year‑on‑year figure in this piece therefore runs 2025 to 2026, both fully reported: peak summer discharge went from 6,990 MW to 11,316 MW, up 62%.

2. It is a Texas record, not an American one

“The biggest demand any grid in America has ever served” is a much better sentence, and it is false. ERCOT is the third largest balancing authority in the United States by peak demand. In the same month, PJM peaked at 162,648 MW and MISO at 118,745 MW. Both are in the same file, four lines away. There is now a permanent check that fails the build if a future data pull ever makes ERCOT look like the biggest.

3. Six power plants that weren’t in Texas

The map is a silhouette of the state, so a tower standing outside that silhouette reads as a bug whether or not the coordinate is correct. Two things cause it: ERCOT serves a handful of sites just over the line, and any outline is a simplification of a river border. Every generator is now tested against the polygon that actually gets drawn, and anything outside is dropped — 6 of 1,887, or 0.3%. The capacity totals are recomputed afterwards, so the map and the numbers can’t disagree with each other.

The part that generalises

The useful idea here is not about Texas, and it is not really about batteries either. For a century, a grid could only ever spend electricity at the instant it made it. Every design decision followed from that: keep enormous machines spinning against a peak you might hit four days a year, and hope.

What changed on 22 July is that a meaningful share of Texas’s evening ran on power generated that afternoon. The system did not need more capacity to survive its own record. It needed somewhere to put the middle of the day.

Sources: EIA-930 Hourly Grid Monitor, balancing authority ERCO, six-month bulk files 2023–2026 · EIA-860M Preliminary Monthly Electric Generator Inventory, June 2026 · ERCOT public fuel-mix dashboard (used only to reconcile the fleet, never as a source of content). Every figure reproduced from the raw files by data/ercot_sunset.py.