AI Infrastructure’s Real Bottleneck Is Power Procurement
Everyone is counting chips. The constraint that decides where the next generation of AI capacity actually lands is the interconnection queue.
Every AI infrastructure conversation starts with chips and ends with power. The GPU shortage got the headlines. The power shortage gets the consequences.
Eighteen months of buildout has settled the question. The binding constraint on AI infrastructure is not compute supply. It is electricity access and the lead time to get it. Chips arrive in quarters. Megawatts arrive in years.
The demand curve explains why this does not fix itself. Data centres are already around 3 per cent of US electricity demand, and RMI and EPRI both put them between 5 and 9 per cent by 2030. Efficiency gains are real. They are smaller than the demand. Grid capacity, not model efficiency, is what actually scales the industry.
Chips were the easy part
GPU supply was the crisis of 2024. It is not the crisis of 2026. Foundry capacity, packaging and high bandwidth memory all caught up, and frontier models need less compute per query than they did two years ago. Nobody sane prices a multi-year bet on chip scarcity anymore.
The real backlog is on the grid. More than 2.6 GW of generation and storage sat in US interconnection queues by the end of 2024, according to the Lawrence Berkeley National Laboratory's annual queue survey. The median project takes about five years from queue entry to commercial service, and that is before transformer supply and local planning are in the picture. A campus signed against today's substation is a 2030 commissioning event.
That reorders the whole industry. The scarce asset in AI infrastructure is no longer the chip allocation. It is the substation. It is the power purchase agreement with a real energisation date. It is grid capacity held while a neighbour still works out the paperwork.
The hyperscalers are buying like utilities
Watch the buyers, not the keynotes. Constellation is restarting Three Mile Island to feed Microsoft's data centres. Microsoft separately committed to gas fired generation in Ireland after the grid queue ran dry. Google signed 500 MW of small modular reactor capacity with Kairos. Amazon backed X-Energy's SMR roadmap and bought a 960 MW campus next to Talen's Susquehanna nuclear plant. OpenAI parked its Abilene build on Constellation's gas fleet at gigawatt scale.
These are procurement decisions wearing tech costumes. The companies are not betting on the physics of the reactor. They are buying power projects with dates, because the interconnection queue is a lottery with a five year lead. When the largest buyers on the planet buy generation instead of waiting in line, the line is the problem.
What this means for AI infrastructure operators
Power procurement has moved from a cost line to the core function. The change shows up in the numbers before it shows up in the narrative. Five rules follow.
Site on the grid, not the map. A region with a fully subscribed queue is a trap whatever the headline tariff says. Live generation nearby, or a zone the grid is already upgrading, beats a greenfield site with a cheap study.
Make energy elasticity loadable. A facility that can shed and hold load when the grid tightens gets priority in a crunch, and in some markets gets paid for it. I argued in Beyond Megawatts that performance depends on more than the peak number: https://brawlersguide.com/post/beyond-megawatts-ai-infrastructure. The operating layer for that is dynamic energy control: https://brawlersguide.com/post/dynamic-energy-control-ai-energy-elastic-infrastructure. Interruptible load is not a compromise when it is priced properly.
Hold storage on site. A few hours of battery turns an intermittent connection into a scheduled one. In constrained markets that is the difference between a PPA that energises and one that idles.
Buy power years ahead. The projects that energise on time are the ones whose agreements were signed while the site was an empty paddock. The power purchase agreement is the new GPU order. Treat it accordingly.
Finance the power, not the silicon. Lenders are underwriting the offtake before the hardware. AI infrastructure that cannot show a locked power supply is a building with an opinion.
The round is decided at the grid
The buildout was billed as a fight for chips because chips are easy to see. It is a fight for electrons. The operators that win hold grid capacity, sign long dated power, and design everything else around the electricity.
The GPU invoice is table stakes. The substation is the fight.