The AI infrastructure build-out is bottlenecked on power, not chips
Everyone is talking about GPU shortages. The real bottleneck is megawatt-scale power delivery. We pulled 18 months of data on interconnection queues and the picture is grim.
GPU supply was the bottleneck of 2023-2024. It's not anymore. NVIDIA is shipping H200 in volume. B200 and GB200 are ramping. Lead times for new orders are measured in weeks, not quarters.
The new bottleneck is power.
The interconnection queue data
We pulled 18 months of interconnection queue data from the four largest US ISOs (PJM, ERCOT, MISO, CAISO) and a few smaller ones. The picture is consistent across all of them:
- →PJM: 250+ GW in queue. Median time-to-interconnect for a 100+ MW project: 5+ years.
- →ERCOT: 200+ GW in queue. Median: 4+ years. Fastest in the country.
- →MISO: 350+ GW in queue. Median: 6+ years.
- →CAISO: 400+ GW in queue. Median: 7+ years.
That's roughly 1,500 GW of projects stuck in queues across these four ISOs alone. For comparison, total US generating capacity today is about 1,200 GW.
Why this matters
If you want to build a 100 MW AI compute hall today, you cannot get a grid connection in any major US market in less than 4 years. Some markets, effectively never. This is the binding constraint on US AI capacity.
Three responses
The industry is responding in three ways:
- →Behind-the-meter generation. Gas turbines (the conventional answer). Batteries (where they make economic sense).
- →Siting in uncongested regions. Wyoming, West Virginia, parts of Oklahoma — places with cheap power and available grid capacity, even if the network latency to major data center markets is bad.
- →Co-location with existing generation. Some hyperscalers are buying retired coal plants and running on-site. We think this is short-term.
We chose option 1 with batteries. The economics work when grid prices are volatile (which they increasingly are), and federal customers can't touch gas.