We are seriously discussing putting data centres in orbit because building enough of them on Earth is getting difficult. I find that more revealing than another benchmark.

The meme in my latest LinkedIn post was the easy part. The harder part is deciding how seriously to take the projects behind it. Terafab is proposing to manufacture chips at an extraordinary scale. Companies building orbital compute want to take the electricity problem somewhere with a better view of the Sun.

There is a sensible argument underneath the spectacle. AI can improve on a software schedule while the infrastructure it needs moves on a construction schedule. Eventually, those two calendars collide.

The model ships. The power connection doesn’t.

A better model can reach users through an API update. More physical capacity requires equipment, a site, financing and permission to connect it all. If one of those arrives late, the rest can sit idle.

The numbers are already large enough to make that uncomfortable. The IEA’s 2026 update puts global data centre electricity consumption at 485 TWh in 2025 and projects roughly 950 TWh in 2030. Those figures cover all data centres, rather than AI alone. Its projection for AI-focused facilities grows faster, roughly tripling over the same period.

Global data centre electricity consumptionThe IEA estimates 485 terawatt-hours in 2025 and projects approximately 950 in 2030. The second bar is a projection, not an observation.The electricity bill keeps growingAll data centres worldwide · TWh per year2025485IEA estimate2030~950IEA projectionSource: IEA, Key Questions on Energy and AI (2026)
The outline marks a forecast. These are annual energy figures, not the power rating of a single facility.

A global total can hide the awkward part. A cluster needs power in a particular place, at the right time, with the right reliability. Spare capacity somewhere else does little for an installation waiting on its own connection.

Efficiency helps, of course. But lower cost per task can also make it worth running more tasks. I wouldn’t plan a large AI product on the assumption that efficiency gains will quietly solve every capacity problem.

Terafab is a manufacturing bet

Terafab’s stated plan brings logic, memory and advanced packaging under one roof, serving both terrestrial applications and space-based compute. That combination matters: a processor without the memory and packaging needed to use it is unfinished capacity.

The headline ambition is output at a terawatt scale per year. I would handle that number carefully. It is an announced manufacturing ambition expressed in terms of compute power capacity; it is not a measurement of chips already being delivered, nor an annual electricity-consumption figure. Putting it next to a TWh demand chart without explaining the units would make a very impressive, very misleading graphic.

What interests me is the attempt to control more of the path from chip design to deployment. That could reduce dependence on somebody else’s expansion schedule. It also means taking responsibility for more things that can go wrong.

For a project like this, I want to see production yield, usable output and delivery dates. A factory announcement tells us where someone wants to go. Deliveries tell us how fast they are getting there.

Orbit changes the engineering problem

The attraction is easy to understand. Put solar-powered compute in an orbit chosen for favourable illumination and you may reduce dependence on terrestrial power connections, land and local weather. Whether that pays off depends on the full system you have to build and launch.

I keep getting stuck on the cooling argument. Space is cold, but a vacuum gives you no surrounding air to carry heat away. NASA’s thermal-control guidance describes the distinction: heat moves through the spacecraft’s materials and thermal systems, and is exchanged with its surroundings through radiation.

The GPU still produces heat. You still need to move that heat to radiators, and the radiator area, operating temperature and orientation still matter. Sunlight and radiation from Earth also enter the thermal budget. A cold background is useful; it doesn’t make the hardware disappear.

That is why I’m wary of treating cooling as a free bonus in a launch-cost spreadsheet. The thermal design becomes part of the payload you must deploy and keep working.

Then there are the questions I would ask before buying any compute service. How much usable work does it deliver? How does the data get there? What happens when hardware fails? How often will I need newer chips? In orbit, the answers have to survive the cost and difficulty of servicing or replacing the equipment.

What has flown, and what is still a date on a roadmap?

It would be wrong to dismiss all of this as a rendering. Starcloud reports that Starcloud-1 launched in November 2025 with an NVIDIA H100 GPU and subsequently trained a small language model in space. That is a real step towards testing the idea. It doesn’t yet tell a customer what sustained commercial capacity will cost.

There is also a correction worth making to the short version in my LinkedIn post. The 2027 demonstration and 2028 prototype dates come from Orbital’s published roadmap. They are that company’s targets, not a shared timetable for everyone working on orbital data centres.

Three different kinds of evidence
ProjectMilestoneHow I read it
Starcloud-1H100 launch in November 2025; small-model training reported afterwardsA company-reported demonstration
Orbital pathfinderAI inference demonstration targeted for 2027A planned mission
Orbital-1Multiple GPU nodes and commercial inference targeted for 2028A planned prototype service
TerafabLarge-scale integrated chip manufacturingAn announced production ambition

I would keep those distinctions visible even in an optimistic forecast. Launching a processor, running a model and offering a dependable service are separate achievements. Each deserves its own evidence.

What I would actually plan around

For someone building AI products, a 2027 or 2028 target is close enough to watch. It is still too uncertain to become a dependency in a delivery plan.

I would use the capacity available today and keep the product portable enough to benefit if new supply becomes attractive. I’d measure the cost of a completed job, including retries, rather than celebrate a cheaper token price while the total bill rises. And I would ask which tasks genuinely need a larger model, longer reasoning or an agent that keeps working after the first answer.

Those are ordinary product decisions. They become more consequential when a service’s appetite for computation grows faster than the infrastructure behind it.

I’m optimistic about people trying difficult engineering. A working orbital compute service would be remarkable. Before treating it as the answer to AI’s energy problem, though, I want the unglamorous details: useful output, uptime, replacement cost and an invoice a customer is willing to pay.

That’s the update I’ll be looking for after the next launch video.