---

The Energy Problem Behind AI - and a Plan to Solve It in Orbit

Every business is being told to adopt AI. Far fewer are talking about what AI actually runs on: electricity, and a lot of it. The data centres behind the tools you're being encouraged to roll out are now among the largest new sources of power demand on the planet, and the grid is starting to feel it. In a recent conversation with his engineering team, Elon Musk laid out SpaceX's answer to that problem — and it's not a bigger power station. It's moving the data centres off Earth entirely.

It's an audacious idea. But the reasoning behind it says something useful about where AI is heading, and why energy — not algorithms — may be the real constraint on how far it can scale.

Why energy is becoming AI's ceiling

Musk frames the whole thing against the Kardashev scale, a way of ranking a civilisation by how much energy it can harness. Type I captures all the power available on its planet, Type II the power of its star, Type III the power of its galaxy. By that measure we barely register — humanity currently uses less than a trillionth of the Sun's output.

That sounds abstract until you connect it to AI. Every leap in AI capability has come with a leap in compute, and every leap in compute comes with a leap in power draw. The industry is on track toward roughly 100 gigawatts a year of new AI compute. At some point you stop being limited by chips or clever models and start being limited by how much energy you can physically generate and cool. That's the wall this plan is built to get around.

Why the answer points to space

The Sun pours out an almost incomprehensible amount of energy, but only about half a billionth of it reaches Earth's cross-section — and most of that is unusable, because roughly 70% of the planet is water and much of the rest is ice cap or desert. To capture a meaningful share of the Sun's energy, Musk argues, you have to leave the planet.

Space also quietly solves AI's other physical headache: heat. Instead of building enormous cooling systems on the ground, a satellite can radiate its waste heat straight into the vacuum. Put those together — abundant uninterrupted sunlight and free cooling — and orbit starts to look less like science fiction and more like the obvious place to put compute that needs vast power.

The practical plan: three things that have to scale

The pitch isn't hand-waving. SpaceX breaks the challenge into three concrete bottlenecks, each with a plan attached:

• Getting hardware up there. You need to lift millions of tonnes to orbit — the job of Starship, designed for full, rapid reusability. SpaceX already launches an estimated 85–90% of all mass Earth sends to orbit; the goal is to climb from around 2,500 tonnes a year toward a million tonnes a year within about three years.
• Power and cooling. Huge solar arrays paired with double-sided radiators to shed heat into space.
• Chips. At first, existing hardware — reference designs cover Nvidia's GB300 and upcoming Rubin chips, plus Google's TPUs. Longer term, a proposed "Terafab" chip plant of roughly 100 million square feet (about ten times the size of Tesla's Gigafactory Texas) to push output toward a terawatt a year.

What an orbital data centre actually looks like

Counter-intuitively, the team says an AI satellite is simpler to build than a Starlink one. Strip away the complex antennas and you're left with mostly solar cells, a radiator, and laser links. The draft first version — call it "AI1" — targets around 150 kW peak power, which is roughly what a single Nvidia GB300 rack of 72 GPUs draws. In effect, each satellite is a rack of compute in orbit, linked to others and back to the ground at low latency through the Starlink network. Crucially, it reuses solar technology already flying on current Starlink hardware — not a leap into the unknown.

The timeline (with a grain of salt)

Musk is upfront that these are targets, not promises: an annualised rate of about 1 gigawatt of space-based AI compute by the end of next year, scaling roughly tenfold each year — 10 GW in about two and a half years, 100 GW in three and a half, and eventually toward a terawatt. For context, a terawatt is roughly twice the entire current electricity consumption of the United States. Beyond even that, the long-range vision involves manufacturing on the Moon and launching satellites into deep space with an electromagnetic mass driver.

Why this matters if your business runs on AI

You don't need a rocket programme to take something from this. The signal for everyday businesses is that AI's costs are increasingly energy costs, and energy is finite, contested and rising in price. That feeds directly into what you'll pay for AI services, how reliable they are, and which providers survive the squeeze. The companies thinking hardest about AI right now aren't just asking "what can it do?" — they're asking "what does it cost to run, and is that sustainable?" Whether or not data centres ever reach orbit, that's the right question to be asking on the ground.

Source: "JUST RECORDED: Elon Musk Announces SpaceX Plans," Brighter with Herbert, YouTube, 8 June 2026. Figures and quotes are drawn from the published transcript of the conversation.