The AI race just hit a wall made of electricity

The AI story everyone tells is about chips. The one that has quietly become just as important is about electricity. A modern AI data center can draw as much power as a small city, and the grid that has to feed it is, across much of the country, decades old and already running near its limits. The bottleneck has moved from the chip to the wall socket, and the clearest proof isn't a forecast. It's the order books at the unglamorous companies that supply the power gear.

One maker of the power modules that feed AI chips, Vicor ($VICR), recently saw its one-year order backlog climb sharply, with new orders running at more than twice what it shipped. A maker of the electrical gear that powers data centers, Forgent Power Solutions ($FPS), roughly doubled its revenue and booked record orders at its last update, and management raised its outlook. When orders move like that, the demand has already arrived.

Step back and the scale of the problem is striking. A single large AI campus can need on the order of a gigawatt of power, roughly what a million homes use, and the companies building these sites want that power in a year or two, not the decade a major grid upgrade can take. The U.S. grid, meanwhile, is a patchwork that's decades old in places, and two things are short at once. There isn't enough generation in the right spots, and the gear that moves power and steps it down, the high-voltage lines, the transformers, and the switchgear, is badly backordered. A large transformer can take years to get, and that lead time, not anyone's willingness to build, is what sets the pace.

You can see the strain in the system's own behavior. PJM, which runs the largest U.S. grid, is moving to line up tens of gigawatts of new generation for data-center load, and it has sought emergency authority to curtail data centers during extreme weather to protect everyone else on the system. The policy response is now pulling the same way. In June, federal regulators at FERC directed every U.S. grid operator to write faster rules for connecting big new loads like data centers, and to make those loads help pay for the grid upgrades they trigger. That does two useful things at once: it clears a path for demand that had been stuck waiting in line, and it routes more of the bill for new transformers and substations to the data centers themselves, which only adds to the order books at the companies that make that gear.

It's also a race. China, by contrast, is widely estimated to be heading toward hundreds of gigawatts of spare capacity by around 2030, and recent reporting suggests it has added more power in just a few recent years than the U.S. has managed over far longer stretches. That gap, between how fast we can build power and how fast AI wants it, is the whole opportunity, and the whole risk.

The mistake is shopping for "the one power stock." There isn't one. It helps to follow the electricity itself, in three stages, because different companies own different stages and they carry very different risks and timelines.

Generation is where power gets made. Natural gas is the bridge that can scale now, nuclear is the certain but slow answer for later, and renewables plus storage fill in around them. This is the stage with the biggest headlines and the longest lead times, since a new power plant, a nuclear one especially, is a multi-year project.

Transmission is moving that power across the grid and stepping it down to a usable voltage. This is the world of high-voltage lines, the transformers that change voltage, the switchgear (the heavy-duty switches and breakers that route and protect power), and the mechanical hardware that holds it all up on the towers. It's unglamorous, it's backordered, and it's exactly where the FERC push lands.

Delivery is the last stretch, inside the data center itself. It's the power electronics that take the building's power and hand it to each chip at the precise voltage it needs, the batteries that ride the site through interruptions, and the cooling that keeps it all from melting. As AI racks pack more power into less space, this layer is being redesigned, which is part of why the order books here are turning first.

Generation gets the attention, but it builds slowly. The gear that moves and delivers power is what's binding right now, it ships on shorter cycles, and the order books show the demand is already here. That's where I spend most of my attention.

There's a second path worth understanding, because it's growing fast. If you can't wait years to connect to the grid, you build your own power on-site. Natural-gas turbines and engines, fuel cells, and on-site battery storage can stand up power for a data center in months, on its own meter, without waiting in the interconnection queue at all. The industry calls this "speed to power," and it has become a real strategy: some of the largest AI infrastructure deals of the past year have been for gigawatts of on-site generation, contracted now rather than promised for later in the decade.

The new FERC rules actually reward this. The fastest grid connections are being offered to loads that can flex and curtail, which turns a data center's own storage and backup into a way to get connected sooner, not just insurance for when the grid wobbles. So the same bottleneck that strains the grid is also pulling money toward the companies that help data centers route around it. I'll name the specific on-site names in Sunday's issue, but keep the category in mind, because it's one of the clearest near-term demand signals in the whole theme.

There's a single material under this whole theme, and it gets confused, so it's worth a beat. Copper is losing one race and winning another. For moving data, copper is hitting its physical limits, which is the whole reason optics is taking over inside the chips. For moving power, there is no substitute, and demand keeps climbing. Analysts estimate a grid needs several times more copper than the data centers it feeds, see data-center copper use climbing toward hundreds of thousands of tonnes a year by the end of the decade, and have the market already in deficit. While we wait for newer technology to come online, the wiring in between is copper, and more of it every year.

The companies that solve this are not glamorous. They make switchgear, batteries, power chips, grid hardware, cooling, and fuel cells. But a GPU with no power and no way to shed its heat doesn't run slowly, it shuts off. That makes this gear as essential to the buildout as the chips themselves, and it's why the order books are turning before the headlines catch up.

The specific companies I hold at each stage, the on-site names I'd only watch, and the nuclear, rare-earths, and copper bets I treat as a long-horizon play rather than a near-term one, are what I'll walk through this Sunday.

If picking individual names feels like a lot, you can get exposure through funds. The most theme-specific is the Defiance AI & Power Infrastructure ETF ($AIPO), the first fund built around exactly this intersection, which holds AI-and-power names across grid equipment, decentralized energy, data centers, and AI hardware. Beyond that, there are broader grid and infrastructure ETFs (tickers like $GRID and $PAVE) for the transmission and equipment names, uranium and nuclear funds (like $NLR and $URA) for the generation side, and copper funds (like $COPX and $COPJ) for the materials angle.

A few things to understand before you go that route. A broad infrastructure fund gives you the theme diluted, since it will also hold utilities and industrial names with little to do with AI. A theme-specific fund like AIPO is more targeted, but it's relatively new, and like any thematic fund it blends the strong names with weaker ones and can be concentrated in whatever is in favor at the moment. And the smaller, more direct names often sit at tiny weights or aren't in these funds at all. None of that is good or bad on its own. It's a tradeoff between owning the whole theme loosely and owning a few parts of it deliberately, and that choice is yours. As always, look under the hood at a fund's actual holdings and weights before deciding.

Sunday's issue is where I name them. I'll cover the company I'd build this basket around, and why I treat it as a steadier hold than the flashier names; the grid-hardware and power-electronics makers riding the buildout; and the on-site "skip the grid" names that the new FERC rules just made more valuable. I'll also lay out how I'm playing copper, where nuclear and rare earths fit as a long-horizon bet, and which of these I own versus only watch.

My first book, Make Your Own Alpha, is on its way. It's the full version of the approach behind these issues: how to build your own portfolio deliberately enough to make work optional, without a finance background or a windfall to start. I'll share a launch date once it's set.

If you want first word on that date, plus the special launch bundle and pricing, get on the launch list.

— Koh

Disclaimer: Nothing in this newsletter constitutes investment advice or a recommendation to buy or sell any security. Numbers and observations are as of publication. I may hold positions in companies discussed above. Always do your own research and consult a licensed financial advisor before making investment decisions.