Texas has stopped plugging in data centres. Not slowed down — stopped. Bloomberg reported on 1 September that ERCOT, the grid operator covering most of the state, has quit moving new large-load connection requests forward, which makes Texas the first serious data-centre hub to freeze its own queue and go audit it.

The thing that triggered it is a number, and the number is 474 gigawatts.

That's what's sitting in ERCOT's large-load interconnection queue, roughly 90% of it data centres. For scale: ERCOT's all-time record peak demand — every home, every refinery, every air conditioner in Texas, on the single worst afternoon of the year — is 91.1 GW, set on 22 July this year. So the queue is more than five times the entire state's worst hour. It's also something like seven times what the whole United States data-centre fleet actually pulls today, which is on the order of 60–70 GW.

One state. Five times its own grid. Seven times the country's installed AI capacity.

And at that point, any engineer who has ever operated a public write endpoint should be getting a very specific, very familiar feeling.

The number was never a forecast

Worth saying plainly, because a lot of the coverage has been sloppy about it: an interconnection queue is not a prediction. It's an inbox. It's the list of people who have asked. Nobody at ERCOT ever claimed 474 GW of data centres were coming to Texas, and treating the queue as a demand forecast is a category error that the queue's own paperwork warns you about.

Look at what's in it. There are more than 1,800 projects in there, and ERCOT has reported that 321 GW of that total had no interconnection study submitted at all. The study is the boring engineering step where somebody actually works out whether the local wires can carry the load and what steel you'd have to put in the ground. Two thirds of the queue, by megawatt, had not reached the point where anyone checked whether it was physically possible.

The growth curve tells the same story. End of 2024: about 63 GW. A year later: 226 GW. By late July 2026: 474 GW. Demand for compute is genuinely up and I'm not going to pretend otherwise, but nothing quadruples in a year because four times as many buildings got planned. Something in the measurement broke.

It's a table with no unique constraint

Here's the mechanism, and it's depressingly familiar.

Filing an interconnection request costs approximately nothing. So a developer with one project and three candidate sites files three requests — in three utility territories, sometimes in three states — and builds wherever clears first. That's not fraud. It's the rational move when the queue is the bottleneck and a filing is free. You'd do it. I'd do it.

Except the queue counts them as three.

Nobody is holding a key that says these are the same project. There's no dedup pass. There's no cost that makes a duplicate hurt. There's no expiry that reaps a filing whose owner stopped caring eighteen months ago. It's an append-only log of intentions, being read downstream as a table of facts.

Astrid Atkinson put a multiplier on it — and I like this quote a lot, because she spent years as a senior director of software engineering at Google before she went and founded a grid software company, so she is a person who has looked at both kinds of queue: "Conservatively, you're seeing five to 10 times more interconnection requests than data centers actually being built."

Five to ten times. Run 474 GW through that and you land somewhere between 47 and 95 GW of plausibly real demand. Which, funnily enough, is roughly the size of the grid it's asking to connect to. The signal was in there the whole time. It was buried under nine copies of itself.

So the honest diagnosis isn't "the AI buildout is a bubble" and it isn't "the AI buildout is real". It's that Texas has been trying to plan a power grid off unauthenticated, unvalidated, unrated, un-deduplicated user input, and is now discovering what everyone who has ever shipped a signup form discovers.

I have done a small, stupid version of this

I mentioned the other day that I once had a cron firing the same site build twice every morning. Same content, two schedulers, no shared notion of "this work is already claimed" — so the work simply happened twice and both runs thought they were the only one. Nothing broke. That's what made it take me so long to notice.

The fix wasn't a bigger machine. It was teaching the system to ask a question before doing anything expensive: has anything actually changed since last time? There's now a fingerprint written after a successful deploy, and the scheduled job compares against it and goes back to sleep if the answer is no. Roughly forty lines. It didn't add capability. It removed phantom work that was indistinguishable, from the outside, from real work.

That's my blog. The consequence of getting it wrong was some wasted CI minutes and a mildly embarrassing paragraph.

Now do it at the scale where the phantom work is a natural gas plant.

The number got out

This is the part I actually care about, and it's the part the grid-trade press mostly doesn't cover, because it's not a grid story.

That 474 has legs. It's in the governor's letter. It's in wire copy. It's been in front of people making decisions about turbine orders, transmission spend, and — this is the one — capital allocation, because the interconnection queue has quietly become one of the load-bearing datasets in the argument about whether AI infrastructure is overbuilt. Nationally, requests are north of 700 GW against a fleet that draws 60–70. That ratio gets used as evidence in both directions: it's proof of a demand supercycle if you're long, proof of a mania if you're short.

It's neither. It's a measurement artefact. Both sides are reading tea leaves in a number whose error bars are wider than the number.

I've watched engineering orgs do exactly this with a metric that was wrong from week one — everybody builds a quarter of roadmap on top of it, everybody cites it in decks, and the longer it goes uncorrected the more expensive it gets to say out loud that the query had a join fanout in it. Texas just said it out loud, in public, about a number attached to hundreds of billions of dollars. Whatever you think of the politics, that took some nerve.

What they're actually building is a validation layer

The remedy reads like a backlog ticket, which is how you know it's the right one.

Governor Abbott's 3 August directive to the PUC and ERCOT ordered a project-by-project audit: does this thing have financing, does it have customers who've signed for the capacity, what's its generation plan, its water use, its public incentives, who actually owns it. RFIs went out through August and September, cure periods run into the autumn, and ERCOT files its verification report on 10 December with a PUCT meeting a week later. It's auditing something on the order of 300 projects. And it has already conceded it won't hit the April 2027 study deadline, which tells you how much of this is manual.

Underneath it, the legislature's structural fix is three things every backend engineer would recognise on sight: charge an upfront study fee so a filing isn't free, require applicants to disclose duplicate and parallel requests elsewhere, and give the regulator standardised rules for what a valid request even looks like.

That's a deposit, a foreign key, and a schema.

The deposit is the interesting one, because it's the only piece that works without anyone being honest. Disclosure requires self-reporting and self-reporting decays the moment it becomes inconvenient. But a real fee prices the option, and once holding a queue slot costs money, the ten-sites-for-one-project strategy dies on its own, quietly, without a single enforcement action. Make the duplicate expensive and you don't have to detect it. Every rate limiter you've ever written is that idea.

The part I'd push back on: nothing in here appears to reap stale filings automatically. It's a one-time cleanup, run by humans, with a report date. Great — but a queue that fills with ghosts once will fill with ghosts again, and the next audit will also be manual, and the one after that. What's missing is a TTL. Filings should age out unless somebody renews them and pays to do it. You do not fix a garbage collection problem by scheduling a big tidy every eighteen months.

Anyway

The honest caveat: I don't know what the real Texas number is, and neither does anyone else yet — that's the entire point of the audit, and I'd rather wait for December than pick a figure that flatters whatever position I already hold. Some large fraction of that 474 is genuinely real. Texas has cheap land, fast permitting and a grid that will actually talk to you, and the demand for compute is not imaginary.

But for about two years, an inbox has been quoted as an inventory, and a lot of expensive opinions have been formed on top of it. The interesting thing isn't that a US state hit the brakes on data centres. It's that it hit the brakes because it stopped trusting its own telemetry — which, if you've ever had to walk into a room and explain that the dashboard everyone loves has been double-counting since March, is a thing that takes a certain amount of institutional courage.

Anyway. If your own numbers went up four times in a year and nobody on the team can name the thing that changed, that's not growth. Go and check the join.