Did Cheap Power Just Become THE AI Trade?

Here’s something that I didn’t think I’d be telling you this summer: American wholesale electricity just got cheaper.

Cheaper… full stop. 

How, you ask? 

Well, the EIA’s latest outlook has power averaging $45 a megawatt-hour this summer, down 8% from last year.

Now here’s the sentence sitting right next to it in the same week’s news: spot prices in Virginia hit $600 a megawatt-hour during a heatwave, and New England prices jumped 243%.

Both of those are true. At the same time. About the same grid.

Trust me, that contradiction isn’t a simple data error. 

This is the AI story, and it’s critical that you understand why it’s about to decide the fate of hundreds of billions of AI dollars. 

This Gas Glut Has a Map

First let’s start with WHY power got cheaper. 

You know as well as I do that the answer isn’t because AI demand cooled off. 

Rather, we can turn our eyes toward one thing — natural gas.

Look, record U.S. natural gas production is clearly outrunning demand growth, and Henry Hub has settled around $3.60 per million BTU — that’s cheap by any recent standard. 

In fact, gas-fired plants have set the price in most American power markets, which makes sense since cheap gas flows straight through to cheap electricity.

But the thing is, it doesn’t flow evenly…

In California (of all places) wholesale power is forecast to fall 30% this summer to just $23/MWh. 

Breaking it down a little more, we’ll find that it’s down 27% to $28/MWh in the southwest, and down a stunning 46% in the northwest, where power prices were greatly helped by a big jump in hydropower. 

Natural gas delivered to power plants in these regions is simply cheaper — and in California’s case, aggressive solar buildout piles on top of that, thus flooding the midday grid with generation and dragging the average down further.

Meanwhile, PJM — the largest grid in America, covering Virginia, Ohio, and the Mid-Atlantic — is forecast at $69/MWh this summer, while ISO New England sits at $64/MWh. 

Both are higher than last year due to the fact that pipeline bottlenecks make gas more expensive to deliver into the Northeast. Of course, neither region has the hydro or solar cushion the West enjoys.

Keep that map in your head, because it matters more than any national average ever could.

You see, an AI data center isn’t a normal power customer. It doesn’t ramp up in the evening and idle overnight like your house does. We’re talking about a gigawatt-scale, 24/7 industrial load — and there’s no question that a lot more of them are coming.

U.S. data center power demand is projected to hit roughly 76 gigawatts in 2026, up from about 50 gigawatts just two years ago. 

Globally, AI data center electricity use jumped 50% last year to around 485 terawatt-hours, and it’s on pace to approach something close to 1,000 terawatt-hours globally — in the neighborhood of Germany’s entire annual electricity consumption, running through server racks instead of factories and homes.

You know that the money behind this AI buildout is almost incomprehensible, too. 

Remember, the five biggest hyperscalers — Amazon, Alphabet, Meta, Microsoft, and Oracle — are set to spend roughly $725 billion on AI infrastructure this year alone. 

To put a little more perspective on that number, it’s more than the GDP of Switzerland.

And every dollar of it assumes the power will be there to plug into.

Nadella’s Warehouse Problem

This is the point when things stop being just theoretical.

A few months back, Microsoft CEO Satya Nadella told Wall Street something that should have been a bigger headline than it was. 

Essentially, he said that he’s got the chips (the most sought-after hardware on the planet), but he’s sitting in inventory that he can’t plug in. 

That’s his problem today.

His CFO, Amy Hood, was even blunter on their latest earnings call — hardware hasn’t limited Azure’s growth; it’s been the power and data centers we’ve been talking about that have restricted things. 

The result is an $80 billion backlog of unfulfilled Azure orders, which is real customer demand Microsoft simply cannot serve. That’s not because the chips aren’t ready, mind you, but because there’s nowhere with enough electricity to switch them on. 

In some of the country’s most established data center markets, interconnection queues now stretch four years or longer.

Don’t get me wrong, this isn’t a Microsoft problem. 

It’s an industry problem, and it’s the same story — it’s just wearing a different hat.

In other words, AI’s bottleneck has moved. 

It used to be about chip supply, and now it’s all about the immense power needed to run things. 

That shift changes where this $725 billion in AI spending actually lands.

Think about it like this… 

Cheap land used to decide where a factory got built. 

Today, cheap available power decides where a data center gets built — and “available” is doing most of the work in that sentence. So, it doesn’t matter how much cash a hyperscaler is sitting on if the regional grid operator tells them the earliest interconnection slot is 2030.

Makes sense, right? The market is still pricing this as two separate stories — an energy story on the commodities desk, a chip story on the tech desk. 

The truth is that it’s actually one story: a power demand story that the investment herd isn’t positioned for yet.

That’s why my readers and I have been tracking the next stage of the AI buildout— the developers building where the power actually is, the players feeding the buildout, and the behind-the-meter plays letting hyperscalers generate their own supply instead of waiting in a four-year line.

The chips were never the bottleneck. 

The power is.