Films Why AI companies are moving into the power business

  • AI
  • Energy

Why AI companies are moving into the power business

Published 31 Oct 2026 · Last fact-checked 17 Oct 2026 · 16:29 · 13 claims, 24 sources

In two sentences

Some AI companies now sit beside turbines, sign twenty-year power contracts, or buy the companies that build power plants. This film follows real arrangements, from Mississippi to Louisiana, to show what "their own power" usually means: who owns the plant, who pays for it, and who controls when the electricity arrives.

Chapters

What we know and what we don'tWhat we know

Established

Uncertain or limited

Sources, by chapter

“In Southaven, Mississippi, SpaceX's AI division—now branded SpaceXSI—says it is building something that looks nothing like AI.”

As checked: MZX Tech LLC, a subsidiary within SpaceX's AI business, holds Mississippi's construction permit for a Southaven project that the AI division, now publicly branded SpaceXSI, describes as a 1.2 GW permanent power plant with 41 turbines under construction.

Our conclusion from the sources

“And it is not alone.”

As checked: The common strategic objective across these different arrangements is greater control over time-to-power.

“Technology companies are signing decades-long power contracts, financing reactors, placing data centers beside power stations, and even buying energy developers.”

As checked: Google has moved into direct energy development by acquiring Intersect, including multiple gigawatts of co-located projects in development or construction.

“Technology companies are signing decades-long power contracts, financing reactors, placing data centers beside power stations, and even buying energy developers.”

As checked: Microsoft's 20-year PPA supports Constellation's proposed restart of the 835 MW Crane reactor.

“Technology companies are signing decades-long power contracts, financing reactors, placing data centers beside power stations, and even buying energy developers.”

As checked: Amazon is financing X-energy and enabling a 320 MW first phase with Energy Northwest, but Energy Northwest is to construct, own, and operate the reactors.

“Technology companies are signing decades-long power contracts, financing reactors, placing data centers beside power stations, and even buying energy developers.”

As checked: Power-plant co-location does not necessarily mean common ownership or permanent behind-the-meter service.

“Because the best computer in the world is still useless if its power arrives years late.”

As checked: AI/data-center expansion increasingly collides with slower grid connections and energy-infrastructure development.

“US data centers used about 192 terawatt-hours of electricity in 2024. In the Department of Energy's reference case, that rises to 649 terawatt-hours by 2030.”

As checked: Electricity use by data centers is rising rapidly in the United States and globally, with AI-focused facilities a major driver.

Also used in this chapter

“And who carries the risk if construction runs late, costs explode, or the customer disappears?”

As checked: Large-load utility tariffs and negotiated service agreements can allocate generation, transmission, interconnection, minimum-payment, exit, and stranded-asset risk differently between the large customer and other ratepayers.

“And who carries the risk if construction runs late, costs explode, or the customer disappears?”

As checked: Long-term PPAs, offtake rights, equity investment, and development funding can improve project financeability by reallocating revenue and development risk, but do not guarantee completion or remove schedule, technology, fuel, construction, and regulatory risk.

“It can finance a project before committing to its output. It can place computers beside an existing plant. Or it can pay a utility that builds generation for the wider system.”

As checked: Entergy's first approved Meta-related Louisiana package contains three combined-cycle plants totaling about 2,260 MW; Entergy later announced plans for seven additional plants totaling more than 5,200 MW under a separate expansion agreement. Meta does not own the plants.

Also used in this chapter

This chapter states no new facts of its own.

Also used in this chapter

This chapter states no new facts of its own.

Also used in this chapter

What you are seeing

Every scene of the film is one of four kinds, and the picture always says which.

  • evidence

    Real evidence

    A documented fact, attributed and dated: a regulator, a record or a company says.

  • illustrated

    Reconstruction

    A real case redrawn in our style: no logos, no real buildings.

  • explanation

    Explanation

    The idea itself, drawn with our own pictures.

  • scenario

    Scenario

    A possible future: always labelled, never shown as fact.

Transcript

0 · A power plant for a computer

  • In Southaven, Mississippi, SpaceX's AI division—now branded SpaceXSI—says it is building something that looks nothing like AI.
  • Not a chip factory. Not a model.
  • A power plant.
  • State records name MZX Tech LLC as the permit holder. Federal records place it inside SpaceX's AI business.
  • SpaceXSI says the project will contain forty-one turbines and produce 1.2 gigawatts.
  • That is power-station scale—being built to help feed computers. Hungry computers.
  • The permit authorizes construction; it does not mean the plant is already operating.
  • Public records do not reveal who legally owns every turbine.
  • But the category collision is real. A software business is reaching backward into fuel, steel, permits, and generation.
  • And it is not alone.
  • Technology companies are signing decades-long power contracts, financing reactors, placing data centers beside power stations, and even buying energy developers.
  • They are not all building their own plants. They are buying control over when electricity becomes available.
  • Because the best computer in the world is still useless if its power arrives years late.

1 · The two clocks

  • Why would a technology company wander into the energy business?
  • Because an AI data center runs on two different clocks.
  • On the digital clock, models improve, capital moves, and campuses rise quickly.
  • On the electrical clock, utilities must approve projects, secure permits, order equipment, build lines and substations, then test everything before the first rack draws power.
  • Those clocks are drifting apart. And you cannot download a substation.
  • US data centers used about 192 terawatt-hours of electricity in 2024. In the Department of Energy's reference case, that rises to 649 terawatt-hours by 2030.
  • It is a projection, not a promise—and includes every kind of data center, not AI alone.
  • But the pressure is concentrated. A giant campus does not ask the country for a little electricity. It asks one region for a block, on one date.
  • The industry calls the problem time-to-power: how long until dependable electricity is available where the computers will stand?
  • If the answer is later than the data center, companies have two choices.
  • Wait for the grid—

2 · "Their own" is doing too much work

  • The word “own” sounds simple until electricity gets involved.
  • Picture a server plugged into a power plant. Now pull on the cable.
  • The plug becomes a meter. The meter becomes a utility tariff. Behind that sit transmission lines, permits, fuel contracts, lenders, and agreements. It is a very long cable.
  • So ask three separate questions.
  • Where is that machine—beside the data center, across the grid, or still on a drawing board?
  • And who carries the risk if construction runs late, costs explode, or the customer disappears?
  • Those questions create six arrangements. A technology company can buy an energy developer. It can host generation on-site. It can promise to buy electricity through a long-term power purchase agreement.
  • It can finance a project before committing to its output. It can place computers beside an existing plant. Or it can pay a utility that builds generation for the wider system.
  • One project can use several at once.
  • Think of them as a ladder of control. The first rung requires no turbine, reactor, or smokestack on a single technology company's balance sheet.
  • It begins with a promise: if you build the power, we will buy it for decades.

3 · Buy twenty years of tomorrow

  • Start with the promise.
  • Microsoft has agreed to buy power under a twenty-year contract.
  • Microsoft does not own the reactor. Constellation does. And the plant is not operating yet; the restart remains under federal review.
  • But Microsoft's commitment matters because power projects consume money long before they produce electricity.
  • Developers must finance engineering, equipment, construction, and fuel.
  • A long-term buyer makes future revenue less mysterious. It gives lenders and investors something to model.
  • Amazon is using a different lever in Washington State. It invested in reactor developer X-energy and is helping advance a first 320-megawatt project with Energy Northwest.
  • Yet Energy Northwest—not Amazon—is expected to construct, own, and operate the reactors.
  • Even that description is ahead of the machinery. The project remains in feasibility, design, and due diligence, with a Nuclear Regulatory Commission construction-permit application still to come.
  • So a technology company can influence which power plants return or get developed without owning the steel.
  • A contract controls future output. Early capital buys influence and options.
  • But there is another shortcut that looks almost comically obvious.
  • If the electricity cannot reach the computer quickly enough, move the computer next to the electricity source. Every cat knows this trick. If the warm spot won't come to you, go to the warm spot.

4 · Move the computer to the power

  • That is the logic behind Amazon's campus beside Pennsylvania's Susquehanna nuclear station.
  • From above, the arrangement looks like the world's most expensive extension cord: reactor here, data center there, problem solved.
  • Except the reactor belongs to Talen, not Amazon.
  • And being neighbors did not make the data center a private grid island.
  • The original arrangement placed the campus behind the meter.
  • When a proposed expansion ran into federal regulatory resistance, Amazon and Talen rewrote the structure.
  • Talen says the transition to a front-of-the-meter agreement occurred in April 2026.
  • In plain English, the data center remained beside the plant, but its commercial and electrical relationship moved through the grid-facing side of the meter.
  • Nothing important moved on the map. Almost everything important changed on paper.
  • That is why co-location is not ownership, and “beside a power plant” does not mean “plugged directly into a private reactor.”
  • Geography can reduce some transmission and scheduling friction, but regulators, markets, and grid rules still exist between the machines.
  • So companies climbed another rung.

5 · Move the power to the computer

  • In New Mexico, Oracle is pursuing that idea at utility scale.
  • Project Jupiter's plan proposes up to 2.45 gigawatts of Bloom Energy fuel cells in a behind-the-meter microgrid for its data-center campus.
  • Oracle says it will bear the energy costs.
  • But the word “own” remains slippery. Bloom supplies the fuel cells. Yucca Growth Infrastructure and BorderPlex are development partners.
  • Public documents do not establish that Oracle owns the generating assets.
  • And this is still a proposal moving through regulation, not 2.45 gigawatts humming in the desert.
  • New Mexico listed air permit 10883 as pending on October 17, 2026. A temporary court stay was lifted in September, but no final permit had appeared.
  • Behind the meter does not mean off-grid. A microgrid can interact with the wider grid, operate independently in some conditions, or combine both modes, depending on its design.
  • Oracle's fuel cells and Southaven's turbines are different technologies. Their strategic logic is similar: reduce dependence on the grid's delivery timetable.
  • But every shortcut brings its own homework—capital, permits, fuel, maintenance, redundancy, emissions rules, and local impacts. And nobody gets to copy the answers.
  • To gain more control, a company can keep assembling partners.

6 · Buy the company that builds the power

  • Google did.
  • This was not merely a contract for electricity. Google bought the organization that assembles sites, permits, capital, and generation around data-center construction.
  • The transaction included Intersect's team and multiple gigawatts of energy and data-center projects in development or construction.
  • That sounds like the top of the control ladder—and at the developer level, it is.
  • But even here, the footnotes matter. IPX Power was carved out during the sale, retaining a large portfolio of solar and battery assets in Texas and California.
  • So “Google bought Intersect” does not mean “Google owns every power plant Intersect ever developed.” Owning the developer also does not reveal who holds title to every generator.
  • What Google unquestionably acquired was capability: people, projects, expertise, and a development pipeline designed to make computing and energy arrive together.
  • That is when “moving into the power business” becomes literal. No metaphor required.
  • Yet buying an energy developer is expensive, unusual, and unnecessary for most companies if someone else can build the megawatts on schedule.

7 · Let the utility build it

  • In Louisiana, Meta chose the utility route.
  • Its Hyperion data-center campus led Entergy Louisiana to an initial package: three combined-cycle gas plants totaling about 2,260 megawatts, plus transmission infrastructure.
  • State regulators approved it in 2025.
  • Entergy began construction that December on two Richland Parish plants; a third is planned later for Waterford.
  • Then the scale grew. In March 2026, Entergy announced plans for seven additional gas plants totaling more than 5,200 megawatts, plus transmission and batteries.
  • They were proposals, not part of the 2025 approval.
  • Meta did not become a power company. It became a very large customer. Entergy—not Meta—would build, own, and operate them, integrating them into its system. Meta pays through utility agreements.
  • That is the difference between cause and ownership.
  • For Meta, the arrangement summons gigawatts without operating turbines. For Entergy, it creates a customer and infrastructure for the broader grid.
  • But utility assets can outlive contracts. Forecasts can be wrong. Eligible costs can enter regulated rates.
  • So the final question is not who owns the plant.

8 · Who pays for speed?

  • Every rung on the control ladder moves risk. None makes it disappear.
  • Long contracts stabilize revenue, but bind buyers.
  • Development capital can launch a reactor project, but cannot manufacture a license or a working plant.
  • On-site generation reduces dependence on grid timing, then creates obligations for permits, fuel, maintenance, backup capacity, and community acceptance.
  • Utility construction moves risk into public regulation.
  • For Louisiana's first approved Meta package, the settlement includes protections.
  • Entergy must obtain parent guarantees and collateral.
  • Early departure triggers a termination fee, and the commission reviews how it is used.
  • Material project-only transmission upgrades stay off other customers' rates.
  • Entergy must report construction, and regulators can review the costs.
  • But the generators remain utility system resources. Eligible costs can enter Entergy's rates. Protections reduce exposure, not all risk.
  • And money is only one ledger. Depending on the technology and location, speed can bring emissions, water use, land demands, noise, fuel dependence—or new capacity and local investment for host communities.
  • Speed has a price.

9 · The new AI moat

  • Return to Southaven.
  • Most companies do not need a power-plant deed. They need confidence that enough electricity will exist, in the right place, when their servers are ready.
  • That confidence can come from contracts, finance, co-location, on-site generation, utility agreements, or acquired developers. Each controls part of the schedule.
  • Chips and models are no longer the entire AI stack. Land, permits, fuel, generation, financing, and utility relationships are competitive infrastructure too.
  • Software's trick is abstraction: hide machinery and make complexity feel instant.
  • It must be generated somewhere, carried through something, approved by someone, and paid for by somebody.
  • So AI companies are not all building their own power plants.
  • They are moving into the power business because advantage may belong to whoever can make physical infrastructure meet a digital deadline.
  • The prize is not simply owning the plant.

How this film was made

Researched, written, drawn and checked by our team. We use AI tools throughout: research, writing, code-drawn animation, narration and music. Every claim is linked to its source.

Corrections

No corrections to date.

Spotted an error? Write to [email protected] with the film's title.Spotted an error? Email us ↗