The hard limit on AI is not chips. It is power. And the grid cannot carry the load, so it is being pulled forward as an infrastructure and energy race.
The power problem is real and wide
Power shortages hit data centers, all manufacturers, and anyone trying to bring new manufacturing online this century. Drawing power from the local grid can hurt the community you move into. The way forward is to build "behind the meter" energy campuses that make their own power. You cannot just build a thousand-acre data center and plug it in. That model does not work going forward.
Innovation campuses, not bare data centers
A California family office builds large infrastructure campuses it calls innovation hubs. Each site generates all its own power because the grid cannot support it. The sites use no outside water. They recapture all water from every use, plus rainwater, hold it in large detention and retention ponds, and filter it through many systems to reuse on site. The goal in any new community is to bring more value than you take.
Much of the public protest against data centers comes from people who do not understand what the centers are for but do see them causing trouble, so they push to restrict them. Generating all power behind the meter removes that harm and sets an example.
The Texas site shows the wider plan. It is a 700-acre site under construction. A government facility being built there will make critical components for the US government. The site also brings in magnet manufacturing, 3D printing, drone manufacturing, and advanced energy research. The hope is about 5,000 permanent jobs, instead of a short job boom that then flattens to two or three hundred. So this solves the energy problem and the community relationship problem at once.
What big tech is doing about its own energy
Big tech is in a straight competition for power. There is only so much behind-the-meter power available, and it starts with natural gas. Making power from natural gas needs gas turbines, and those turbines are in short supply. When extra turbines are announced, Meta and SpaceX grab them right away.
Bloom Energy is a standout. Its fuel cells are highly innovative. For sites in non-attainment areas or under very tight emission limits, Bloom's fuel cells give off very little emissions and get around many of the hurdles in the large-power game.
The money is staggering
The spending on building out AI infrastructure is shocking. The cost runs into the tens of billions of dollars per megawatt. At a 100-megawatt-plus or a gigawatt campus, the CapEx matches the GDP of smaller countries.
The world needs these large language models. This is a matter of national security, a tech race and a tech war among the greatest powers, and the US is at the front of it. The aim is to make enough energy so Meta, SpaceX, and Anthropic get the power they need.
A gigawatt of power sounds simple on paper. Building it in the real world is a major effort: pipelines, natural gas supply, turbines, NOx emissions, gear, and air permits all go into it. These campuses take years to develop.
Timelines that do not match
Many tech companies can only forward-look about 6 to 8 months. Site developers spend 2 or 3 years getting these sites ready, so big tech can come in and use them on its own timeline.
Where the opportunity is: nuclear
We are in a power and energy race. The most interesting space now is nuclear. Going forward, the grid will not be able to support all the models. The only path will be MMRs and SMRs (small modular reactors) placed right next to data centers. That gives behind-the-meter power for the data center and lets the campus sell power back to the grid.
The current grid is weak and vulnerable. The Texas site will have a grid connection, but not to take power. It exists to give power back, especially when the grid goes down or struggles. These large campuses can act as a bridge the grid needs in hard times, which is part of the national security logic behind them.
The next shift
At some point the story shifts back to software efficiency, doing more with less power, rather than only building out more AI capacity.
