Here’s a chart you almost never see in electricity circles: inflation adjusted average cost of electricity from the very beginning. We’ve been flat for 56 years since 1970. Utter failure. We must unlock cheaper electricity.

An opportunity exists to add 75 GWs of additional US inference compute, quickly, affordably, and in a manner local communities will celebrate rather than fight. A terrawatt of solar projects are stuck in the bureaucratic morass contraining our economy known as interconnection queues. A ton of these projects will withdraw from the queue and abandon their plans simply because they can't wait any longer. It costs money to sit there and do nothing. If retooled for 95%+ uptime off-grid AI data centers, these projects could support something like 150 GW of compute. Cut that in half because a bunch of projects are fake or shitty. Still, 75 GW of compute. THAT'S A LOT OF COMPUTE, EXTREMELY FAST, INCREMENTAL TO ALL CURRENT PLANS. Then hook em' up to the grid a few years down the line to trade all of that overbuilt solar and battery capacity (which will lower people's bills) for getting to 99.99% uptime at the data center. The solar resource is there. The land is buildable. The permitting can be done. And the grid is there, just eventually. And these data centers would alievate all of the typical local objections: - they don't raise your electric or gas bill - they actually will lower it in the near future - they don't pollute your local air via a combustion power plant - they don't make a ton of noise via a combustion power plant - they don't need to use abnormal amounts of water - they wont be gigantic 1 GW collosus concrete buildings that upon merely observing you become depressed. it'll be just a bunch of solar farms with modular compute boxes scattered throughout. sheep will graze the glass under the panels. it'll be chil. "But Duncan, there's no way you could build enough storage. It would have to be so expensive." If you design this from first principles rather than copy typical banker-style solar project designs (direct current, dense & minimal solar racking, modular compute, etc.) the LCOE can be < $150/MWh. That's fine. Deals are clearing at that level already. "But Duncan, everyone wants 1 GW sites. A ton of 25 MW compute sites simply wont due." Not for inference. It's only training that needs giant coherent clusters. Historically hyperscalers have wanted all compute to be fungible between traning and inference, but that is quickly changing as they witness the massive wave of inference headed their way. "But Duncan, 95-99% uptime isn't good enough! Inference must be 100% uptime!" Nah. This is all incremental, so 95% is a hell of a lot better than 0%. And its just 95% for a few years while you wait to hook up to the grid. For now if you must get 100%, then dedicated your GW scale traning site to supply the remaining 5% on a handful of winter nights. And inference demand shapes are biased toward the daytime anyway. It'll be fine. "But Duncan, I have no imagination. I can't think beyond what has already been proven. Who will the offtakers be? Will it be bankable? I'm feeling anxious!!" We're already short on inference compute and AI has touched less than 1% of the economy. If you can't see where things are going I can't help you. Solar and batteries are the only massively scalable option we have. The grid bottleneck is here. The gas prime-mover bottleneck is here. The EPC bottleneck is here. We need a new way. And keep in mind, the above only describes retooling existing in-development solar projects for a quick win. There would be many many more GWs of opportunity beyond this if we develop new additional solar projects on top of that.

Wow, less than 10 years ago dats centers wanted power for $30/MWh. Now they’ll take it for $150 in some instances.