Every new AI model has to be trained somewhere, and that somewhere runs on electricity at a scale the grid was never asked to deliver before. Data center electricity has quietly become the binding constraint on the entire AI build-out, which is why Jon Najarian’s Infinite Power presentation for Banyan Hill spends so much of its runtime on megawatts and transmission lines rather than chips and software. Here is the bottleneck in plain terms.
The Demand Shock
A modern AI training cluster draws power on a scale that would have been unthinkable for a data center a decade ago. Where a traditional server room might run on a few megawatts, the largest training facilities now talk in hundreds of megawatts, enough to power a small city. Every hyperscaler is racing to add capacity, and every unit of new compute needs a matching unit of electricity before it can do anything useful.
That is the demand side of the story, and it compounds quickly. The more capable the models, the more compute they need to train, and the more electricity each new generation consumes. The constraint is not the chips; the chips are being manufactured as fast as anyone can buy them. The constraint is getting enough current to the building.
A Grid That Was Never Built for This
The supply side is where the problem lives. The American Society of Civil Engineers graded US energy infrastructure a D+ in 2025, and its age statistics are stark: 70% of power transformers are 25 years or older, 60% of circuit breakers are 30 years or older, and 70% of transmission lines are 25 years or older. The University of Texas Energy Institute estimates that bringing the grid fully up to date would cost near $5 trillion.
Worse, the Department of Energy has warned that blackouts could rise 100 times by 2030 under certain scenarios as demand growth collides with that aging hardware. The grid was built for a world of steady, predictable load. Data center electricity is neither, and the mismatch is the bottleneck. Our full Infinite Power teardown covers how Najarian frames these same statistics.
The Memphis Shortfall
The concrete example that grounds the whole pitch is xAI’s Colossus supercomputer in Memphis. The facility needed 150 megawatts to come online. Memphis Light, Gas and Water, the local utility, could spare about 8. A full Tennessee Valley Authority approval and transmission build-out would have taken roughly three years.
Rather than wait, xAI trucked in about 35 gas-turbine “power pods,” each a turbine and generator mounted on two trucks, and stood up more than 400 megawatts in roughly 122 days. The gap between 8 and 150 is the bottleneck in one number, and it explains why data center operators have stopped treating utilities as their only option. For the demand story from another publisher’s angle, our Elon’s Shadow Trust review is a good companion read.
How Data Centers Are Responding
The response is behind-the-meter generation: building power on site so the electricity never has to cross a congested, aging grid. Gas-turbine pods are the fastest route, online in under 12 hours each, and they can be scaled by adding more pods rather than waiting on transmission approvals. The same approach is now scaling at Southaven, Mississippi, where a second, larger site is coming together.
That shift turns power from a utility service into a supplier market, and it is the entire investment thesis Najarian is selling. The equipment vendor becomes more interesting than the utility, because the vendor ships the pods that make the workaround possible.
The Scale of the Fix
The scale of the problem is hard to overstate. The University of Texas Energy Institute puts a full grid modernization near $5 trillion, and the Department of Energy has warned that blackouts could rise 100 times by 2030 in some scenarios. Those are not growth forecasts a supplier can capture quickly; they are the reason data centers are building their own power rather than waiting for the utility to catch up. The bottleneck is not going away, which is exactly what makes electricity the scarce input of the AI build-out.
The Honest Read
The data center electricity bottleneck is real and well documented, and the Memphis numbers check out against public reporting. The investor question is narrower than the trend: which suppliers capture the spending as data centers self-generate, and at what price. Najarian’s answer sits behind the $79 Market Rebel Report trial, and the concentration risk, 96% of the teased supplier’s 2024 Power Pod revenue from one customer, is the counterweight any buyer should weigh against the demand story.
Ready to see the research? Click here to access Jon Najarian’s report.
NewsletterVetter is an independent publication. We receive compensation from some of the services we review through affiliate links. Nothing on this site is investment advice. Always do your own research.