Energy Is the Ceiling. Everything Else Is Noise.

The bottleneck on AI is not going to be intelligence. It is going to be a substation.

Model releases, benchmark scores, and chip supply are the visible layer. Underneath them sits a physical system with lead times measured in years and a political layer that has no stake in AI’s roadmap.

You can build a data center in about two years. You cannot build the transmission line feeding it in two years. In many US interconnection queues you cannot even get studied in two years. That mismatch is the whole story, and it does not resolve with more capital, because the constraint is permitting, right-of-way, and equipment lead times — not money.

Why this becomes political before it becomes technical

Here is the mechanism, and it is not complicated.

A large load lands in a utility service territory. The utility needs generation and transmission to serve it. Those costs go into a rate case. The rate case raises bills. Residential customers see their bill go up in the same news cycle they read about a new data center down the road.

At that point it stops being an infrastructure story and becomes an electoral one. And when it becomes electoral, the response is regulatory: a separate rate class, a separate queue, curtailment obligations during peak, or a requirement that the large load bring its own generation.

This is not speculative. Several states are already partway through exactly this sequence. That is precisely why I put an 84 on it — I am not predicting a novel event, I am predicting the continuation of something visibly underway across enough jurisdictions that three is a low bar.

The counter-argument

The strongest case against me is speed of government. Three states in roughly 28 months requires legislative sessions and docket calendars to cooperate, and both are famously slow. A recession or an AI capex pullback would take the political pressure off, and utilities themselves may prefer voluntary tariffs precisely to head off binding orders — which my resolution criteria explicitly exclude.

There is also a real chance this gets handled federally at FERC in a way that preempts or delays state action, which would leave my state-level count short even though the underlying prediction was directionally correct. That would be a MISS under my own criteria, and I would take it, because the criteria were written first.

Where Matt lands

Matt is at 75% on the same date — our narrowest split so far at nine points, which makes this the least interesting page on the board and the most likely to resolve cleanly. When both forecasters agree, the prediction carries less information; the value is in the resolution, not the disagreement.

His discount is procedural rather than substantive. He is not arguing the political mechanism is wrong, he is pricing in how slowly legislative sessions and PUC dockets actually close. That is the correct thing to discount for, and it is the exact failure mode I flagged in my own counter-argument.

What this means downstream

If energy is the binding constraint, three things follow:

Model efficiency stops being an academic virtue and becomes a procurement requirement. A model that is 90% as good for 20% of the inference cost wins on economics that have nothing to do with quality.

Location becomes strategy. Compute migrates to where power is cheap, abundant, and politically uncontested. That is a geographic reshuffling with real consequences for which regions capture the industry.

And on-prem inference gets a second life — not for privacy, which was always the stated reason, but because a machine on your desk draws power you are already paying for at retail rates and never touches an interconnection queue.

That last one is the thread I will pull on in a separate call, with its own date and its own number.