AI’s Power Problem Just Became a Contract Business

For most of the AI boom, the scarce thing was the chip. Investors paid whatever it took to own the companies that made them and the companies that bought them. This month the scarce thing changed, and you can see it in the contracts.

On October 6, Google signed a power agreement with Constellation covering 3.6 gigawatts. Of that, 890 megawatts is new nuclear capacity that Constellation will squeeze out of 11 existing reactors at six sites in Illinois, Pennsylvania and New Jersey, under a 20-year deal. Constellation is spending more than $4.3 billion on the upgrades. A second agreement covers 2,700 megawatts for 15 years.^1, 2

Ten days later Microsoft did something similar in Indiana. It agreed to take roughly 1,100 megawatts from NiSource for a data center campus for 20 years, running from 2028 to 2048, and to pay for all of the generation and transmission needed to deliver it.³

The details of that Indiana deal matter. Microsoft is not just buying electricity at whatever the market charges. It is paying for the generation and the transmission lines themselves, and NiSource has said it will give more detail on its earnings call on October 28.³ When the customer pays to build the supply, the supplier carries far less of the risk.

Twenty-year contracts are not a hype signal

A press release about a chatbot can be forgotten by Friday. A 20-year power purchase agreement cannot. The companies with the most to spend on artificial intelligence are now committing to pay for electricity into the late 2040s. That tells you how long they expect the demand to last, and how worried they are about getting enough of it.

A press release about a chatbot can be forgotten by Friday. A 20-year power purchase agreement cannot.

Microsoft alone spent $116 billion on capital expenditures in fiscal 2026.³ That money buys buildings, chips and servers, but every one of those servers has to be plugged into something, every hour of every day.

✓ Trusted partner One research publisher has put out a briefing on a profitable American energy company it says sits right where AI’s power demand and Washington’s energy push meet →

The government’s own numbers point the same way. In its October outlook, the Energy Information Administration forecast that U.S. power use will reach 4,288 billion kilowatt-hours in 2026 and 4,356 billion in 2027, after a record 4,195 billion in 2025. It expects commercial demand, the category that includes data centers, to set a new record as well, and it sees wholesale power prices rising 11%.⁴

Texas is where the strain shows first

Texas runs its own grid, and it is setting records in months that used to be quiet. On September 12, ERCOT hit a September peak of 86,647 megawatts, 6,698 megawatts above its September 2025 peak. The grid operator also reports 9,456 megawatts of large loads already approved to connect.⁵

Large load is grid language for facilities like data centers. Each one is a customer that needs power around the clock and often wants to lock in supply for years. For a company that already produces or moves energy in Texas, that is a very different customer from the one it had five years ago.

Washington picked a side

Policy is pushing in the same direction. On August 27, the White House declared a national emergency over foreign-made power grid equipment and gave the Energy Department 120 days to write new rules. Energy Secretary Chris Wright had put the logic plainly in May: “What’s the most important source of energy for AI? Electricity.”⁶

Then on October 21, President Trump went further on Truth Social, writing that “New and large… Data Centers are mandated by me to make their own electricity,” and telling the industry to “MAKE DEALS, and get them built!!!” The post is not a law or an executive order. But it tells data center builders where the political pressure is headed: away from the shared grid and toward dedicated supply, signed and paid for.⁷

Who gets paid

Put the pieces together and a pattern appears. Technology companies have the cash. Power producers have the electrons. Washington wants the two to deal directly. The result is a new kind of contract, long, large and paid by some of the richest companies in the world, landing on the books of energy companies.

✓ Trusted partner See the research behind this story →

It also shifts the bargaining power. For years, power producers had to compete for customers. Now the largest buyers of computing in the world are competing for the same limited supply, and they are willing to sign for decades to secure it. In a market like that, the seller of the scarce input gets to set more of the terms.

That changes what an energy business is worth. A producer selling into a spot market lives and dies with the price of the day. A producer with a multi-year agreement from a top AI company has something much closer to a revenue stream it can plan around. Constellation’s deal with Google is the headline example, but the same logic applies to smaller producers that sign agreements of their own.

The market does not always catch that shift quickly. Investors chasing AI have spent the last few years staring at chipmakers and software companies. The businesses that generate the power, move the fuel and own the ground around those data centers get far less attention, even when they are already profitable and already paying shareholders. When a company like that signs a long contract with an AI giant, the question is whether its share price has noticed.

There are real risks. Power prices can fall, projects slip, and deliveries under Google’s deal start as early as 2028.¹ A political post can be walked back. And if the AI spending wave slows, every company counting on it will feel it.

But the direction is hard to miss. In the space of six weeks, two of the largest technology companies on earth signed up for decades of electricity, the federal forecast called for new records, Texas broke one, and the President told data centers to bring their own supply. AI has a power bill now. The companies that get to send it are worth knowing.