Inside Crusoe’s energy-first approach to data centers

Start with the constraint, because it drives the rest. Asked directly whether power or chip supply limits AI today, Lochmiller says power, and describes the bottleneck moving. A year earlier the scramble was for chips. Once the existing data center capacity was absorbed, US vacancy ended 2023 below 2% and was below 1% at the time of recording, which for a commercial real estate asset class is effectively nothing. What now limits new capacity is access to power, and he lists the specific chokepoints rather than leaving it abstract: grid interconnection queues, high-voltage transformers, switchgear and backup generation.

If power is the scarce input, the siting logic inverts, and one property of AI makes that possible. Training workloads are far more tolerant of latency than most computing, so the load can be positioned geographically in a way that, say, a trading system or a consumer app cannot. That is the whole basis of the energy-first approach. It sends Crusoe to places that are not traditional data center markets: West Texas, where wind and solar are heavily built out and there is substantial curtailment and negatively priced power; a former Alcoa factory in upstate New York attached to a large hydro dam Lochmiller describes as massively underutilized; and Iceland, where geothermal and hydro are cheap and the climate makes cooling easier. Alongside those brownfield sites, the company partners with independent power producers on new generation, taking load behind the meter, using the existing substation infrastructure and keeping the grid connection as a backup rather than as the primary supply. His analogy is that aluminum smelting already made this trade, shipping ore across an ocean to do the electricity-intensive step in Iceland. Training a model is the better version of the trade, because what moves is data over subsea cables rather than physical material.

The demand argument is where he separates himself from the industry’s public posture, and Lacey names the contrast: large tech companies concede privately that this will use a lot of power while hedging in public. Lochmiller does not hedge. He thinks demand is under-forecast, and supports it with a thought experiment rather than a model: if every American adult ran half an H100 as a copilot for daily work and social interaction, that alone would require 250 gigawatts, which he says is orders of magnitude larger than the forecasts on offer. He adds positive reflexivity, meaning that as the models get more useful people use them more, which drives more compute, which draws more power. The reason he frames this as an opportunity rather than a threat is that the load is steerable, so it can pull net new clean generation into existence where it would not otherwise be built: batteries alongside West Texas wind and solar as costs fall, or new gas paired with carbon capture and sequestration, an area he says Crusoe spends a great deal of time on. Underneath that sits a value judgment he states plainly, that power consumption is not bad in itself if it advances things worth advancing, and an expectation that AI will invent solutions to sustainability problems. That last step is the load-bearing one and it is asserted rather than demonstrated.