AI data centers: the energy bill nobody can close

You ask ChatGPT a question and the answer arrives in seconds. It seems to cost nothing. But somewhere in the United States, a warehouse full of computers burned electricity to write that answer. These warehouses are AI data centers, and in the last week of September 2026 three pieces of news showed they are running out of outlets.

Oracle sent a "force majeure" notice, the document that allows delays without penalty, about one of the country's largest projects. Crusoe, which builds these warehouses, walked away from a $1.25 billion contract for turbines, the machines that burn gas to generate electricity. And Cloudflare's chief said robots could become a thousand times more numerous than people on the internet.

American sites covered each story separately. This article connects all three, explains the difficult terms and shows who pays the bill.

Quick summary: the answer in 30 seconds. AI needs a lot of electricity, and the American power grid cannot deliver on schedule. That is why companies started generating power on their own land, almost always with natural gas. This plan now runs into environmental permits, turbine shortages and opposing residents. Meanwhile, demand only grows: more than half of website visits, the so-called traffic, already comes from programs, not people.

Index

• What happened in the last week of September?
• Why do AI data centers consume so much energy?
• What is Oracle's force majeure notice on Project Jupiter?
• Why did Crusoe give up Boom Supersonic's turbines?
• Will bots really be a thousand times more numerous than humans?
• Who pays the AI energy bill?
• What does this change for you?
• Frequently asked questions
• What to watch in the coming months

What happened in the last week of September?

In three days, three companies showed the same problem from different sides. On the 24th, news broke that Oracle invoked a protection clause on a giant New Mexico data center. On the 25th, Crusoe walked away from the turbines it had ordered. On the 26th, Cloudflare projected an internet taken over by robots.

The connection between them is a three-link chain. Robots and AI assistants make more and more requests. Each request is served by computers, which sit in data centers. And each data center needs a power plant to function. When the last link jams, the other two feel it.

Why do AI data centers consume so much energy?

AI data centers consume so much energy because they gather thousands of powerful computers running day and night, without pause. A single cluster of these buildings can demand as much electricity as a large city. In the United States, the power grid takes years to meet a request of this size, and companies do not want to wait.

A data center is a building that houses computers instead of people. Think of a cold storage facility: from the outside it is a common warehouse, but everything inside depends on power running all the time. When you talk to an AI, it is in a place like this that the answer is assembled. Without it, the app on your phone does nothing.

The unit used in this conversation is the gigawatt, abbreviated GW. It measures power, that is, how much energy something can deliver or spend at the same time. The entire Itaipu plant, with its 20 turbines, has 14 GW of installed capacity. That is the ruler we will use from here on.

Stargate, OpenAI's data center program with Oracle and SoftBank, was announced as a commitment of $500 billion and 10 GW. That is almost three quarters of an Itaipu just to run artificial intelligence. One of the chosen locations is in Doña Ana County, New Mexico. It is Project Jupiter, which appears in the next section.

"Behind the meter" energy: generating on your own land

Generating "behind the meter" is producing electricity within your own land, without going through the public grid. It is like a building that installs a generator in the basement because the utility cannot turn on the lights in time. The difference is scale: here the generator is the size of a power plant.

This way out became the rule. Wired magazine reported that, in some places, the queue to connect a data center to the grid takes years. Building the source next to the building cuts the line. And the fastest source to erect, today, burns natural gas.

The numbers come from Global Energy Monitor, an organization that tracks energy projects worldwide. According to the survey released in January 2026, gas plant projects in the United States nearly tripled in 2025 and reached almost 252 GW, the equivalent of 18 Itaipus. More than a third of that was planned to feed data centers on site.

In one sentence: AI demands energy faster than the grid can deliver, and the solution found was to bring the power plant inside the data center.

That is why energy, not the chip, became the bottleneck. You order a chip, it leaves the factory and arrives by truck. A plant needs land, fuel, heavy equipment and government authorization, and each item has its own queue.

The gas shortcut solves the wait for the grid. But it trades one problem for three: getting the gas, getting the machine and getting the permit. Oracle ran into the permit. Crusoe, as we will see, into the machine.

What is Oracle's force majeure notice on Project Jupiter?

It is a document in which Oracle warns it may delay payments if the data center is not ready by 2028, citing reasons beyond its control. The company is not leaving the project. The information came from Bloomberg and was repeated by TechCrunch. Oracle says the work remains on schedule.

Force majeure is a contract clause that frees one party from fulfilling the agreement when something happens that it could not prevent. Imagine you rented a hall for a party and a flood closed the street. You did not cancel the party, but it also makes no sense to pay a penalty. The clause exists for cases like this.

In Project Jupiter, Oracle is the tenant. The one building the cluster of buildings is a company linked to Blue Owl Capital, which manages investors' money, and Oracle will occupy the space to serve OpenAI. A lease of this size starts being charged on a set date. The notice serves to say: if the delay is not my fault, the clock does not run.

Where the project got stuck: gas and permit

Project Jupiter depends on natural gas, which arrives by pipeline, and on a permit that has not yet been issued. The energy plan changed midway. In January, Oracle spoke of gas turbines. In April, it announced the switch to Bloom Energy fuel cells, with up to 2.45 GW installed.

A fuel cell is a device that transforms gas into electricity by chemical reaction, without flame. It works like a battery that never runs out as long as gas keeps coming in. In Jupiter, thousands of them would form a 2.45 GW plant, a little over one sixth of Itaipu. For neighbors, the advantage is less smoke and almost no water use.

Oracle itself gave the numbers for the swap in a letter to the state environmental agency. Nitrogen oxide emissions, gases that worsen air quality, would drop about 92%. Carbon dioxide emissions, the main cause of global warming, would drop 21%. The system continues using natural gas, and the letter says so in so many words.

The air quality permit is missing. The public hearing was scheduled for October 19, and New Mexico's environment department has until November 23 to decide, according to the Source New Mexico site. Without it, the cells cannot be turned on.

What Oracle and Blue Owl say

Both companies deny delay. In a statement to CNBC, Oracle stated that Project Jupiter "remains on the planned schedule" and that it maintains its commitment to New Mexico. Blue Owl declared that "this notice does not change the financial commitments" of the project.

Notice the first two words from Blue Owl: "this notice." They confirm the document exists. Neither published the document, and the known content comes from Bloomberg's anonymous sources. What is confirmed is that the notice exists. What it allows in practice, only the contract says. On the day of the news, Oracle shares fell more than 3%.

Why did Crusoe give up Boom Supersonic's turbines?

Neither company explained the reason in detail. Crusoe only says the partnership is not the most suitable at this time and that it will continue using turbines, just from other manufacturers. The agreement called for 29 gas turbines from Boom Supersonic, for $1.25 billion, and ended before the first delivery. The end was announced by Boom's president on September 25, as TechCrunch reported.

First, who is who. Crusoe is one of the largest AI data center builders in the United States. It was the one that erected the Abilene, Texas cluster used by OpenAI. Eight days before the turbine news, the company had announced a $3.9 billion investment. Investors valued the company at $30.9 billion.

Part of this money goes to Crusoe Spark, small data centers, mass-produced and assembled wherever there is energy. It is the logic of the prefabricated house: instead of building everything on site, the pieces arrive ready. Crusoe claims this shortens the deadline and reduces cost compared to common construction.

The deal that did not get off the ground

Boom Supersonic is known for another project: a passenger plane faster than sound. In December 2025, it announced a parallel business. It adapted the plane's engine to stay still on the ground, burning gas and generating electricity. It named the machine Superpower and presented Crusoe as its first customer.

A gas turbine is, at bottom, an airplane engine bolted to the ground: the gas burns, spins a shaft, and the shaft spins a generator. Each Superpower delivers 42 megawatts (one gigawatt has a thousand megawatts). The 29 units would total 1.21 GW. With their profit, Boom intended to finance the plane.

On the 25th, Boom's president, Blake Scholl, wrote on X that the two companies would not go forward together. According to TechCrunch, the first version of the text said the turbines had left Crusoe's short-term plans, and that sentence was deleted later. Scholl stated that Boom will deliver about 250 megawatts to other customers in 2027. That gives six turbines, against the 29 in Crusoe's order.

Crusoe replied to TechCrunch that its energy plans "have not changed." It chooses the source for each location, among turbines, wind, sun, batteries and the grid. According to the company, the first Abilene cluster, of 1.2 GW, uses the grid and keeps turbines for emergencies. A second, of 900 megawatts, will run on turbines on site.

The gas turbine queue

The case exposes a bottleneck few people see: there are not enough turbines for everyone. The Global Energy Monitor survey shows that two thirds of gas plant projects under development worldwide still have no manufacturer defined. Three companies, GE Vernova, Siemens and Mitsubishi, concentrate more than 75% of orders that already have a supplier.

It was this queue that opened space for an airplane manufacturer to sell generators. And that is why losing the first customer weighs on Boom, even with other buyers on the list. For those building data centers, the lesson is that money buys the land and the chips, but does not speed up the turbine factory.

In one sentence: Crusoe has billions in cash and still wants gas; what changed is from whom it buys the machine.

So far we have talked about energy supply. We still need to look at the other side: why demand for computing keeps rising.

Will bots really be a thousand times more numerous than humans?

It is a projection, and the author himself warns he got the previous ones wrong. Matthew Prince, Cloudflare's president, said in an interview with The Verge that, in five years, automated traffic could be a thousand times larger than human traffic. The confirmed fact is another: more than half of internet traffic already comes from programs.

A bot (robot) is a program that visits sites on its own. The oldest is Google's, which reads pages to build search. The new ones are AI agents: programs that browse on your behalf. You ask "find the best camera up to 3,000 reais" (in Brazilian reais), and the agent opens stores and compares prices for you.

Cloudflare can measure this because more than 20% of the world's websites pass through its network. The company works like a doorman: it stands in front of the site, lets in who should enter and blocks attacks. From the door, you can count who is people and who is a program.

From 20% to more than half of traffic

The turn was quick. Before conversational assistants like ChatGPT, bots were about 20% of traffic, according to Prince. In March 2026, he predicted at the SXSW festival they would pass humans in 2027. Two months later, in May, the team announced it had already happened.

The reason is the difference in appetite. Prince gives an example: a person researching a camera visits about five sites. Their agent visits a thousand times more, something like 5,000. In a post published on September 30, Cloudflare reports that daily requests from AI agents on its network grew more than 1,700% in one year, that is, became 18 times larger.

The forecast for the entire internet is another calculation: it comes from stretching this growth curve for five years. Prince cautions that people will not disappear. Human traffic stays more or less the same, and robot traffic explodes. The number is worth treating as an order of magnitude: he himself admits having gotten every prediction he made on the subject wrong.

Advertising does not work with robots

Each robot visit costs money to the site owner, who pays for the computer that hosts the pages and for the volume of data sent. With people, the account closed with advertising. With robots, it does not close, because robots do not click on ads. Prince sums up the problem like this: if traffic grows a thousand times, someone has to pay for the infrastructure that serves that traffic.

Cloudflare's bet is to charge at the door. There has existed on the internet, since the 1990s, a response code called 402, which means "payment required" and was almost never used. The company started returning this code to robots: the program pays a fraction of a cent and receives the page. The system is still in testing.

For our story, the point is another. Each of these visits, paid or not, is processed in a data center. If agent traffic grows close to what Prince imagines, the queue for energy that stopped Oracle and Crusoe only increases.

Who pays the AI energy bill?

For now, those paying are the companies and their investors, with the promise that neighbors will not feel it in their electricity bill. The environmental cost, that one stays for everyone.

In New Mexico, Oracle promises to bear all the energy of Project Jupiter, with its own grid, separate from the public one, so as not to affect residents' rates. It also promises $360 million in direct payments to the county and $50 million for water systems. That is the price of being accepted.

Even so there is resistance. CNBC cites local opposition and environmental groups' concerns among the project's obstacles. The reason is simple to understand: a 2.45 GW plant next to your house, with flame or without flame, consumes gas every day for decades.

On climate, the bill is collective. If all gas projects under development in the United States come to fruition, the country's gas plant fleet grows almost 50%, Global Energy Monitor calculates. Swapping turbine for fuel cell cleans the neighborhood air, but, by Oracle's own number, cuts only 21% of carbon dioxide.

What does this change for you?

In the short term, little: the AIs you use remain online. The effect appears in three places, depending on your case. Those who have a website will receive more robots than people. Those who pay for AI depend on projects that may delay. And those concerned about climate gained a number to track: how much carbon dioxide each project truly cuts.

In Brazil, the subject still seems distant, but the logic is the same anywhere: without spare energy, there is no data center. To follow the other sides of this race, our artificial intelligence page gathers the site's articles. The fight over chips, for example, is in Nvidia buys Hugging Face.

Frequently asked questions

What is force majeure in a contract?

It is the clause that suspends an obligation when something beyond the company's control prevents the agreement, such as a natural disaster or a denied permit. It does not cancel the contract. In Oracle's case, according to Bloomberg, it serves to delay payments if the data center does not open in 2028.

What is OpenAI's Stargate?

It is OpenAI's AI data center program with Oracle and SoftBank. The announced commitment is $500 billion and 10 GW of capacity in the United States. Project Jupiter, in New Mexico, is one of the locations.

How much is 2.45 GW?

It is a little over one sixth of Itaipu's installed capacity, which has 14 GW. That is the maximum size planned for Project Jupiter's fuel cell plant, all of it dedicated to a single data center cluster.

Did Crusoe give up using gas?

No. The company says it continues using gas turbines, just not Boom Supersonic's. It also uses wind, sun, batteries and the power grid, depending on the location.

Are bots already the majority on the internet?

Yes, according to Cloudflare. The company states that, in 2026, for the first time, more than half of internet traffic was not human. The measure applies to sites that pass through its network, which are more than 20% of the total.

Will AI energy consumption make electricity bills more expensive?

It depends on how the data center is connected. In Project Jupiter, Oracle promises to pay all energy with its own grid, with no effect on residents' rates. When the data center uses the public grid, the discussion is different, and it was not addressed by the sources in this article.

What to watch in the coming months

Three dates tell whether the account closes. On October 19, New Mexico holds the public hearing for Project Jupiter's permit. By November 23, the state decides. And 2027 is the year Boom promises to deliver its first turbines to other customers, proof that the product exists off paper.

The deadline that matters most is 2028, when Jupiter should turn on the computers. If it delays, the force majeure notice stops being paper kept in a drawer. AI data centers were born from a race for chips. The next phase is decided in a much less glamorous place: the power plant.