The forecast was right for one day
Texas, 3 August. The governor ordered an audit of every data centre in the ERCOT interconnection queue before any more advance: 474 GW requested, about 90% of it data centres, more than five times the grid's record peak. ERCOT reports back on 10 December. One forecaster's estimate of what gets built on the same grid by 2030 is around 8 GW.
I have a table like that of my own.
In November 2020 I finished a paper at Texas A&M on Bitcoin price. Six models. Two of them, the tree models, I reported as overfitted, and I still think that was the honest part of the paper. The one I trusted was a vector autoregression, once it had the best RMSE of the set, and I used it for the only falsifiable thing in 21 pages: a 15-day forecast, 11 to 25 December, with a 99% interval. Table 8, page 19.
Day one, 11 December: forecast mean $18,015, close $18,057. Forty-two dollars off. I remember being pleased with that, which should have been the warning. On 16 December, day six, the price crossed $20,000 for the first time in its history; the upper bound for that day was $19,332. On 25 December, the last day of the forecast, the close was $24,665 against a ceiling of $20,295. A 99% interval, broken on day six, finishing 21% above its own top. I never published the comparison. This is it.
What broke is more useful than that it broke. The introduction named three pillars — confidence in the protocol, digital scarcity, social interest — and called social interest "maybe the most important factor." The variable I used for it was the Google Trends index for the word Bitcoin. In December 2020 the buyers were an insurer putting in $100 million and a London fund putting in 2.5% of £20 billion. Those buyers don't search. Search volume that month was well below the 2017 peak, when the price was lower. The pillar was right. The sensor on it was sampling a population that wasn't the one moving the price.
That is the whole finding, and it took me six years to write it down: a forecast is only as good as the population its inputs sample, and nobody tells the model when the population changes.
So, the 474 GW. The verb is requested. Not built, not contracted, not metered — filed. Counted by whoever files, and until this summer filing cost nothing. What the queue measures is how many people would like a free option on Texas power. That is a real quantity; it is just not the one a load forecast needs. Give it its comparisons, once a naked number is worth nothing: the record peak is about 91 GW; one build-out estimate through 2030 is about 8 GW. Between what was filed and what one house expects to see built, roughly 60 to 1. The audit is Texas asking its instrument what it actually measures. The number to read on 10 December is not how many projects passed; it is requested megawatts against megawatts with money posted behind them.
I don't know how ERCOT's own forecasts weight the queue. My guess is they discount it, and that the discount was tuned on a period when the queue was mostly generators and the requests were mostly serious. Same failure as mine, other direction. My instrument didn't see the buyer; theirs sees loads that don't exist.
Machines do this too, and I learned it there first. An accelerometer on a bearing housing measures the structure's response, not what is exciting it. On a hydro unit in 2012 that was enough to tell me the machine was shaking and nothing about why; the reading that carried the diagnosis came from a proximity probe on the shaft, and the probe belonged to the client's consultants. I borrowed it. Most of the predictive maintenance I get shown now is Google Trends: a proxy that tracked the thing until the population changed. A drill line gets replaced, the load cell is recalibrated, and the baseline is still.
What I'd bet on
We don't publish a lead time from a backtest as if it were caught live. We have models on a fleet of engines that, run against real events, would have warned days ahead. All of them were built after the events they now predict. The number goes in the deck labelled backtest, in smaller type, and it costs us the best slide we have.
Also this month
- The EIA's short-term outlook has US electricity demand at records in 2026 and 2027, mostly data centres.
- RAND ran the US interconnection queue through completion rates and losses: 1,086 GW in, 151 GW out.
Diego Mercadal started as a commissioning engineer on high-voltage motors and generators, joined an offshore drilling contractor as a rig hand, worked several positions in the drilling crew, and ended up running its AI/ML function. He is now co-founder and CEO of Wonder DataLabs.