The AI Infrastructure Gold Rush: Who Pays for the Electricity?

For thirty years, we thought of digital technology as weightless. AI is reminding us that intelligence at scale is an industrial activity, built on concrete, copper, cooling water and electricity.
In September, Microsoft signed a twenty year agreement with Constellation Energy to restart Unit 1 of the Three Mile Island nuclear plant in Pennsylvania, to supply power to its data centres. In March, Amazon bought a data centre campus directly adjacent to a nuclear plant operated by Talen Energy for $650 million. In October, Google signed an agreement to buy electricity from small modular reactors to be developed by Kairos Power.
When the largest technology companies in the world start contracting nuclear power, it is a sign that the constraint on AI is no longer only chips. It is energy.
The physical constraint
The International Energy Agency estimated earlier this year that data centres consumed around 460 terawatt hours of electricity in 2022, and that this could more than double by 2026. In Ireland, data centres already account for roughly a fifth of national electricity consumption, and the grid operator has effectively restricted new connections in the Dublin area. In several parts of the United States and Europe, the waiting time for a new grid connection for a large facility is measured in years.
This year we also saw how fast capital can move when it wants to. Elon Musk’s xAI built a cluster of 100,000 GPUs in Memphis in a matter of months, partly powered by on site gas turbines while grid capacity caught up. That is a story about engineering speed, but it is also a story about how quickly physical infrastructure has become the bottleneck.
A new set of investment opportunities
For investors this opens a set of opportunities that rarely appeared in technology funds: power generation, transmission and grid modernisation, energy storage, cooling technologies, electrical equipment, and specialised real estate.
Private markets have moved quickly. Blackstone, which acquired QTS in 2021, agreed this year to buy AirTrunk, one of the largest data centre platforms in Asia Pacific, in a deal valued at around A$24 billion. Infrastructure funds that once focused on toll roads and airports are now competing for data centre platforms and power assets.
The risks that come with heavy assets
But this kind of investment comes with risks that software investors are not used to.
Data centres and power plants are long lived assets financed with significant debt. Their returns depend on long term contracts with a small number of very large tenants. If AI demand grows more slowly than expected, or if efficiency improvements reduce the compute needed per unit of output, there could be overcapacity. The telecom fibre build out of the late 1990s is the cautionary tale everyone in this market should know.
There is also a social and regulatory dimension. Communities are starting to ask who pays for grid upgrades, how much water is used for cooling and whether data centres compete with households and industry for affordable power. Energy is becoming a political question, and AI infrastructure is now part of it.
From digital to industrial
The big picture is that the AI revolution is starting to look less like a software cycle and more like a major industrial investment cycle, closer to railways or electrification than to the app economy. That changes the profile of the winners. Some of them will be technology companies. Many will be utilities, equipment manufacturers, construction firms and infrastructure investors.
Investment question: Will the next generation of AI investment winners be found in software, or in the physical infrastructure that makes AI possible?
