AI Industry Structure, Layer 3: Power

Have you seen the news that Microsoft signed a direct power supply deal with a nuclear power plant? Why would a tech company sign a contract with a nuclear operator? The answer is simple — AI data centers now use about as much power as a decent-sized city. Let’s start by covering why power, the last essential piece of Layer 3 of the AI industry structure, suddenly became so important.

Revisiting Power

Running tens of thousands of GPUs non-stop takes about as much power as a decent-sized city. Even with servers and cooling equipment fully in place, a data center is useless if there isn’t enough electricity to run it.

That’s why securing power itself has recently emerged as the new bottleneck determining how fast data centers can be built.

Why Power Became a Bottleneck

Servers and GPUs can be bought relatively quickly, as long as you have the money. But the power grid is different. Building a new power plant or laying transmission lines takes years. Power companies never expected AI data center demand to explode this suddenly, and today’s power infrastructure simply wasn’t designed to handle this level of demand from the start.

As a result, some regions are seeing data center construction delayed because “the land and funding are there, but the power isn’t.” The core of this problem is that it’s a bottleneck money can’t solve.

Why a Finished Power Plant Still Can’t Be Used: The Interconnection Queue

There’s one more lesser-known specific bottleneck here. Even after a power plant is fully built, actually connecting it to the grid means getting in a separate line. In the U.S., the generation capacity stuck in this queue totals 2,600 GW — more than double the country’s entire installed generating capacity. The average wait is 5 years, and in regions packed with data centers like Northern Virginia, it can take up to 7 years.

This is the real reason behind installing gas turbines and SMRs right next to the data center, as mentioned earlier. Since even the largest power plant faces a 5-7 year wait just to connect to the grid, companies bypass the queue entirely with a “behind-the-meter” approach — generating and using electricity right on-site without ever going through the grid.

Why Big Tech Chose Nuclear

Renewable energy like solar and wind is cheap, but generation stops when the sun sets or the wind dies down. AI data centers, on the other hand, have to run non-stop, 24/7, without a single moment’s break. That’s why nuclear power, which generates electricity steadily regardless of weather, has become such an attractive option.

As a result, big tech companies like Microsoft, Amazon, and Google are signing direct, long-term power supply deals with nuclear operators. Constellation Energy, the largest nuclear operator in the U.S., frequently appears as the key partner in these deals.

Another thing worth noting is gas turbines. Since nuclear plants take a long time to build, more companies are installing gas turbine generators right next to data centers as a stopgap. Companies like GE Vernova, which supplies these turbines and grid equipment, are also benefiting from this trend.

Smaller, Faster SMRs Have Also Emerged

Large nuclear plants typically take more than 10 years from approval to completion. That’s why an alternative called SMR (Small Modular Reactor) has been rapidly gaining ground recently. Because components are pre-manufactured in a factory and just assembled on-site, SMRs can be built far faster and cheaper than large nuclear plants.

Oklo has already signed a deal with Meta to build a 1.2 GW nuclear campus in Ohio, and struck a 500 MW deal with Equinix as well. NuScale Power is also pursuing a large-scale SMR deployment in the Tennessee Valley. SMRs aren’t as proven as large nuclear plants yet, but big tech’s interest is rapidly gravitating toward them, since they’re a dedicated power source sized just right for a single data center.

What’s Actually Powering Things Right Now

Nuclear gets so much buzz that it might seem like the dominant choice, but in reality, close to half of U.S. data center power still comes from natural gas. By IEA figures, natural gas is the overwhelming #1 at over 40%, renewables sit at 24%, and nuclear is still around 20%. Nuclear and SMRs are the glamorous story, but gas is what’s actually holding things up right now.

In other words, big tech doesn’t put all its eggs in one basket. They run a “portfolio strategy” — using gas to put out the immediate fire, renewables and batteries to lower costs, and nuclear and SMRs to prepare for 10 years down the road, all at the same time.

What Investors Should Watch For

Power has become a variable that shapes AI competitiveness just as much as securing GPUs. No matter how many GPUs a company buys, they sit idle without power.

The thing to check is the scale and duration of the power purchase agreements (PPAs) this company has signed with big tech. These deals typically run for decades, so once one is signed, it tends to lock in stable revenue for a long time.

Closing Thoughts

Power is the least visible layer in the AI industry structure, but it’s actually the first one to get gridlocked. No matter how many GPUs a company secures, they’re useless without the electricity to run them. That’s why big tech is playing a complicated game: preparing for 10 years down the road with nuclear and SMRs, working around the immediate bottleneck with gas turbines, and running multiple power sources at once, all while factoring in the physical limit of the grid interconnection queue.

What makes this layer interesting from an investment angle is that it’s not a technology race — it’s a game of “who lands the contract first.” Companies like Constellation Energy, GE Vernova, and Oklo aren’t selling flashy new technology; they’re selling already-proven power generation methods to big tech through long-term contracts. That’s why stocks in this layer tend to react more sensitively to data center groundbreaking news and power-deal announcements than to AI model performance releases.

In the next article, we’ll take a close look at the final topic in Layer 3: networking and optics. We’ll cover why InfiniBand and Ethernet are going head-to-head in the technology that ties tens of thousands of servers together.