AI, As I Go

Energy

Power from space, and the nuclear backup plan

AI needs more electricity than the grid can add quickly. Elon Musk and Jeff Bezos both say the answer is in orbit, where the sun never sets. Here is what each of them actually said, what is really being tested, how much power space could bring, and why nuclear is the nearer-term bet.

Three ways to power AI A hand-drawn sketch of three ways to power one AI campus. Grid and gas: today's default; Entergy plans 10 gas plants for one Meta campus. Space solar beamed down: today's demo satellites send about 1 kilowatt each; NASA's 2050 reference designs deliver about 2 gigawatts each, but at 12 to 80 times the cost of power from ground solar and wind. Or put the AI chips in orbit instead, the idea Musk, Bezos and Google are pursuing. Nuclear: one restarted reactor, Crane, adds 835 megawatts, targeted for 2027; fusion is still unproven. All of it feeds one AI campus, which can need up to 2.7 gigawatts. three ways to power AI Read: The power problem ↓grid + gastoday's default10 gas plantsbeing built for oneMeta campus Read: How much power, really? ↓space solar, beamed downdemo satellites today:~1 kW eachNASA 2050 design: ~2 GW,power 12–80× pricier Read: Musk's version ↓or put the AI up there(Musk, Bezos, Google) Read: Nuclear: the nearer-term bet ↓nuclearone restarted reactor:835 MW(Crane, target 2027)fusion: still unproven Read: How much power, really? ↓one AI campus:up to 2.7 GW Three ways to power AI A hand-drawn sketch of three ways to power one AI campus. Grid and gas: today's default; Entergy plans 10 gas plants for one Meta campus. Space solar beamed down: today's demo satellites send about 1 kilowatt each; NASA's 2050 reference designs deliver about 2 gigawatts each, but at 12 to 80 times the cost of power from ground solar and wind. Or put the AI chips in orbit instead, the idea Musk, Bezos and Google are pursuing. Nuclear: one restarted reactor, Crane, adds 835 megawatts, targeted for 2027; fusion is still unproven. All of it feeds one AI campus, which can need up to 2.7 gigawatts. three ways to power AI Read: The power problem ↓grid + gastoday's default10 gas plantsfor one Meta campus Read: How much power, really? ↓space solar, beameddemo satellites today:~1 kW eachNASA 2050 design: ~2 GW,12–80× pricier power Read: Musk's version ↓or put the AI up there(Musk, Bezos, Google) Read: Nuclear: the nearer-term bet ↓nuclearone restarted reactor:835 MW(Crane, target 2027)fusion: still unproven ↓ all to feed ↓ Read: How much power, really? ↓one AI campus:up to 2.7 GW
Three ways to get power to one AI campus, with the honest numbers for each, explained in the sections below. Click or tap a title or picture in the sketch to jump to its section.

A single large AI campus can draw hundreds of megawatts to multiple gigawatts. One project in Cheyenne, Wyoming, which Google acquired from developer Crusoe in November 2026, has grown to 2.7 gigawatts, nearly triple the state’s current electricity use. Globally, data centers used about 485 terawatt-hours in 2025, up 17% in a year, and are on track for roughly 945 terawatt-hours a year by 2030, according to the IEA.

Today that power mostly comes from the grid, plus new gas plants built for the purpose. Entergy plans ten of them for Meta’s Hyperion campus in Louisiana (three approved so far, seven awaiting a vote). The full picture of what a campus consumes is in What it takes to keep AI running.

Meta (META) — demand Entergy (ETR) — utility

Every option below is in the same unit, gigawatts (GW) of steady power, so you can compare them directly. What we need comes first, then a couple of examples of who can supply what, from what's real today to what's still on paper. Each bar shows that option as a share of what data centers worldwide will need by 2030, about 108 GW — roughly double the 47 GW they use today. A bar you can barely see means the option is tiny next to that need.

Option and whoPowerShare of AI’s 2030 need (108 GW)When and the catch
What AI needs
All data centers worldwide, 2030IEA projection 108 GW 100% Needed by 2030945 TWh a year, a little more than double today
One giant AI campusGoogle, Cheyenne, Wyoming 2.7 GW 2.5% PlannedNearly triple Wyoming's current electricity use
Real today, or signed and being built
Grid + 10 new gas plantsEntergy, for Meta's Hyperion campus 7.5 GW 6.9% Being built3 plants approved, 7 awaiting approval
Power from a running reactorTalen, for Amazon 1.92 GW 1.8% Signed, through 2042Susquehanna plant, Pennsylvania
Planned for the 2030s
Upgrades + new reactorsVistra, Oklo, TerraPower, for Meta 6.6 GW 6.1% Signed, by 2035“Up to”; most new reactors land in the 2030s
Small modular reactorsX-energy, for Amazon 5 GW 4.6% Planned, by 2039“More than 5 GW”; first 320 MW in Washington state
Unproven, or on paper
AI computing in orbitSpaceX (plan) 100 GW 93% Plan: per year of launchesNeeds 1M tonnes launched a year; the world launched ~3,000 tonnes in 2025
Space solar, beamed downToday's demo satellites 1 kW <0.01% Demos, 2026About 1 kW each: one campus would need ~2.7 million
Space solar, beamed downOverview Energy, for Meta 1 GW 0.93% Agreement, 2030“Up to”; first orbital demo planned for 2028, unproven
AI computing in orbitBlue Origin (vision)“gigawatt”no number yet10–20 yearsBezos’s own timeline; Project Sunrise filed for ~51,600 satellites

Today’s worldwide figure is the yearly total spread evenly over the year (415 TWh ÷ 8,760 hours ≈ 47 GW), my arithmetic. Contract figures are what was announced, often “up to.” SpaceX’s 100 GW is computing it hopes to add per year, not power sent to Earth.

Above the clouds, a solar panel in the right orbit gets sunlight almost all the time. Google’s researchers estimate it can be up to 8 times more productive than the same panel on the ground.

There are two ways to use that. The old idea, from the 1960s onward, is to beam the power down: collect sunlight on a huge satellite, turn it into microwaves or a laser, and catch it with a big antenna on the ground. The new idea is to skip the beam: put the AI chips in orbit next to the panels, and send only the answers down. That second idea is the one Musk and Bezos are talking about.

On February 2, 2026, SpaceX confirmed it had acquired xAI, Musk’s AI company. Around the same time, SpaceX asked the U.S. FCC for permission to launch up to 1 million satellites for orbital data centers. Its math: 1 million tonnes of satellites a year, at 100 kilowatts of computing per tonne, adds 100 gigawatts of AI computing a year.

“The sun always shines in space!” Musk said, adding that “ultimately, there is a roadmap to launch a terawatt” a year. At Davos in January 2026 he went further: “the lowest cost place to put AI will be space and that will be true within two to three years, three at the latest.”

SpaceX’s own IPO filing is more careful. It says the orbital AI plans involve “unproven technologies” and “may not achieve commercial viability.” Note what Musk is not proposing: beaming power to the ground. His plan uses the sunlight in orbit and moves the computing to it.

SpaceX (SPCX) — launch + orbital compute

Bezos has pitched this for years. In 2019 he said, “We’ll move all heavy industry into space,” leaving Earth zoned for homes and light industry.

At Italian Tech Week in October 2025 he applied it to AI: “we’re going to start building these giant gigawatt data centers in space.” His reason: “These giant training clusters, those will be better built in space, because we have solar power there, 24/7.” His timeline was “10+ years,” and he bet on “not more than 20.”

In March 2026 his rocket company, Blue Origin, filed with the FCC for Project Sunrise: about 51,600 data center satellites in sun-synchronous orbits, launched on its New Glenn rocket. Like Musk, Bezos is talking about using the power in orbit, not sending it down.

Blue Origin — private

Beaming power down. Caltech’s SSPD-1 launched in January 2023 and became the first to beam power wirelessly in space and point a detectable beam at Earth, onto a rooftop in Pasadena. It was funded by more than $100 million from Donald and Brigitte Bren. Japan’s OHISAMA demo is a 180 kg satellite with a 2 m² panel, meant to send about 1 kilowatt to the ground. UK start-up Space Solar has a letter of intent to deliver 30 megawatts to Reykjavik Energy in Iceland by 2030. Much bigger on paper: in April 2026 Meta signed for up to 1 gigawatt from Virginia start-up Overview Energy, which plans to beam sunlight down as near-infrared light onto existing solar farms so they keep generating at night, with a first orbital demo planned for 2028 and commercial power as early as 2030. Aetherflux (now Cowboy Space), founded by Robinhood co-founder Baiju Bhatt, put its laser demo satellite, Reason-1, into orbit on October 1, 2026, on SpaceX’s Transporter-18 rideshare. As of mid-November, no result from its first attempts to beam power down to a ground receiver had been reported.

China’s plans are the biggest on paper. Rocket scientist Long Lehao has described a kilometer-wide solar station in geostationary orbit, built with the coming Long March-9 rocket; one outside analyst expects a roughly 1 megawatt demo around 2030. Europe looked hard and paused: ESA’s SOLARIS study judged the technology “still relatively immature” for a demo mission.

Computing in orbit. Starcloud launched a 60 kg satellite carrying one Nvidia H100 GPU in November 2025. It first pitched a 5 gigawatt orbital data center; it now aims for 20 gigawatts of computing spread across 88,000 satellites, with no date set. Its next step is much smaller: two Starcloud-2 satellites with 8 kilowatts of power each, planned for rideshare launches in 2027, after a $250 million raise in August 2026 at a $2.3 billion valuation. Google’s Project Suncatcher put its first prototype satellite into orbit on October 1, 2026: built with Planet, launched on a SpaceX rocket, carrying four of Google’s TPU AI chips. On October 30, Google said the chips were up and running: an onboard Gemma model wrote a haiku, the first AI response from a data-center-class TPU in orbit. Radiation and heat tests come next, and two more satellites are planned for 2027 to test how they link up. Cowboy Space has joined in too: the upper stage of its planned Big Iron rocket is meant to stay in orbit as a megawatt-class data center carrying five racks of Nvidia Vera Rubin GPUs, with the first flight targeted for 2028 and a constellation of up to 20,000 of them awaiting FCC approval.

Alphabet (GOOGL) — Suncatcher Planet Labs (PL) — satellites Nvidia (NVDA) — GPU in orbit

Here is the honest gap. Today’s demo satellites move about 1 kilowatt. The 2.7 gigawatt Cheyenne campus would need about 2.7 million of them (my arithmetic). Space Solar’s 30 megawatt pilot, if it arrives in 2030, would cover about 1% of that campus. Overview Energy’s deal with Meta, up to 1 gigawatt by 2030, would cover about 37% of it (my arithmetic), if the technology works.

NASA studied full-size versions in 2024. Its two reference designs each deliver about 2 gigawatts to the grid, roughly one campus, with a ground receiver up to 6 km wide, assuming operation from 2050. The cost came out at $0.61 to $1.59 per kilowatt-hour, versus $0.02 to $0.05 for wind and solar on the ground: 12 to 80 times more expensive. Launch was over 70% of the cost, at an assumed $1,000 per kg.

Computing in orbit avoids the beam, but it has its own math. Google says launch prices need to fall below about $200 per kg, which it expects around the mid-2030s, before orbital AI roughly matches ground costs. SpaceX’s 100 gigawatt-a-year plan needs 1 million tonnes of satellites a year; the whole world launched about 3,000 tonnes in 2025, more than 300 times less (my arithmetic). And heat is hard to get rid of in a vacuum: by one estimate, radiating away 1 megawatt takes about 1,200 m² of radiator, about four tennis courts.

So space power is real physics with real demos, but this decade it is kilowatts, not gigawatts.

Nuclear runs around the clock, which is exactly what an AI campus wants, and the big tech companies are already signing deals. The fastest power comes from reactors that already exist.

Restarts and existing plants. Microsoft signed a 20-year deal in 2024 to restart Three Mile Island Unit 1, renamed the Crane Clean Energy Center: 835 MW, which Constellation still expects to start in 2027 (one grid analysis suggests a connection could slip to 2031). Amazon agreed in June 2025 to buy up to 1,920 MW from Talen’s Susquehanna plant through 2042.

New small reactors, mostly in the 2030s. Amazon and X-energy aim for more than 5 GW by 2039, starting with four reactors (320 MW) in Washington state, expandable to 960 MW. Google and Kairos Power aim for 500 MW by 2035; the first 50 MW plant in Tennessee is due in 2030. In January 2026 Meta announced deals for up to 6.6 GW by 2035, with Vistra (433 MW of upgrades to existing plants), Oklo (up to 1.2 GW in Ohio, as early as 2030) and TerraPower (up to eight Natrium reactors).

The numbers are the same order as an AI campus, which is why tech companies like nuclear. The catch is time: apart from restarts and upgrades, most of this lands in the 2030s.

Constellation (CEG) — restart Talen (TLN) — supply Vistra (VST) — supply Oklo (OKLO) — reactors Microsoft (MSFT) — demand Amazon (AMZN) — demand Alphabet (GOOGL) — demand Meta (META) — demand

Fusion is the bet with the biggest payoff and the least proof. In 2023 Microsoft agreed to buy power from Helion, which broke ground in July 2025 on a 50 MW plant in Malaga, Washington, targeted for 2028; a Helion executive has said full output might not come until 2030, and some physicists doubt the approach. In June 2025 Google agreed to buy 200 MW from Commonwealth Fusion Systems’ first ARC plant in Virginia, planned for the early 2030s. CFS is still building SPARC, the test machine meant to show its design can make more energy than it uses.

Treat both as unproven until a plant actually sends power to the grid.

Microsoft (MSFT) — Helion buyer Alphabet (GOOGL) — CFS buyer Helion, CFS — private

Now to 2027: whether Cowboy Space’s Reason-1, already in orbit, beams power down, and whether Japan’s OHISAMA does too; how Google’s first Suncatcher satellite performs, and its two follow-ups in 2027; Starcloud’s two Starcloud-2 satellites in 2027; Starship launch prices; the Crane restart.

2028 to 2030: Helion’s first fusion power; Overview Energy’s first orbital demo in 2028, the test behind Meta’s 1 GW deal; Kairos’s 50 MW plant; Oklo in Ohio; Space Solar’s 30 MW pilot.

2030s: SMR fleets from Amazon, Google and Meta; launch costs near $200 per kg, the point where orbital AI starts to make sense on paper.