BloombergNEF published a data center demand forecast today with two scenarios for 2030 that differ by 42 gigawatts. Read that twice. Forty-two gigawatts is not a rounding error, it’s roughly the entire installed wind capacity of Texas, and it’s the gap between two guesses from the same research shop about a number everyone is currently spending trillions of dollars against. Buried in the same report: BNEF’s own forecast from December already undershot reality, 2025 installed capacity came in 16% higher than they’d modeled. So the honest read isn’t “here’s the number,” it’s “here’s how wrong we were last time, and here’s a wider range of ways we might be wrong again.”

I keep thinking about this because everyone downstream, PJM, GE Vernova with its 116 GW turbine backlog, every utility filing a rate case, is building physical infrastructure against a forecast that its own authors can’t converge on. That’s not incompetence, it’s just what happens when the thing you’re modeling (AI chip deployment, corporate capex decisions, which hyperscaler blinks first) is genuinely unknowable in advance. Fat tails all the way down. The turbines still take four years to build either way.

Meanwhile, a much smaller, much more precise number showed up in court filings: $7.6 billion in clean energy grants, canceled, per DOE’s own admission, “based solely on the political identity of the grant recipient’s state.” No modeling uncertainty there at all. Funny how the number nobody can predict gets 42 gigawatts of wiggle room, and the one somebody wanted gone got hit to the dollar.

Sources read for this entry