Stranded compute is filling up partially complete data centers. The GPUs are arriving faster than the power infrastructure needed to run them. Elon Musk is right on the bottleneck: the issue is no longer chips. It is transformers, cooling, and permits.
Morgan Stanley revised U.S. data-center power demand up 43 percent to 97 GW through 2028. Even after behind-the-meter generation the shortfall is still roughly 32 GW (TFTC on Morgan Stanley). High-voltage transformers now sit at lead times above 160 weeks (EMSNow / Wood Mackenzie). You cannot announce a new campus on the same schedule as a new transformer.
Most people still treat “compute” as something abstract that happens inside a laptop or a phone. It is not. Compute is the processing capacity that runs the models, the recommendation engines, fraud detection systems, and the enterprise layers that now sit on top of almost every digital service. You do not see it any more than you see the oil derivatives in your clothes or the plastic parts in your car. You just notice when it is missing.
Oil never left the economy just because politicians started talking about reducing dependence on it. It moved into places people do not think about. Compute is doing the same thing, only faster. The difference is that compute has an immediate and non-negotiable input: electricity. Without reliable power the GPUs do nothing useful. They become expensive metal in an unfinished building.
That is why the portfolio has been accumulating energy producers. Not as a thematic bet on “the energy transition,” but because electricity is the baseline currency for compute. Constellation Energy is an active holding for this exact reason; it bypasses the narrative and owns the physical constraint. The unbundled platforms we use to hold the energy book sit under our brokers and accounts. The common thread among these companies is not ideology. It is that they produce or control the thing the new demand actually needs.
The people who locked in GPU capacity just to resell it are now discovering that a chip without a grid connection and a cooling system is not a business. It is inventory. You can announce new data centers. You cannot announce new high-voltage transformers or new nuclear plants on the same timeline.
We are not going to stop using compute. But the market is still pricing the software while ignoring the grid. The alpha isn’t in the chips; it’s in the power required to turn them on.