AI’s Power Problem Moves Off Earth

A megawatt of computing power in orbit would need radiators covering about four tennis courts, Karol Szubstarski of OTB Ventures said, and enough modules to fill a Falcon 9 by mass. On Earth, as one participant said, a megawatt is nothing. That gap ran through the Ripple on AI's power problem. Several investors at the table thought engineers would close it in time. None could say who would pay for the years before they do.

Patrick Tucci of Vsquared Ventures began by separating edge compute, a GPU on one satellite or a computing satellite shared by a constellation, from a full data centre in orbit that would compete with training and inference on Earth. Many start-ups, Patrick said, raise money on the data-centre story while building edge compute.

Edge compute already works. The largest cluster in orbit, Karol said, is ten satellites carrying 40 GPUs, launched by Kepler a couple of months earlier. It runs real workloads, and it is tiny. It suits radar imaging, where files are so large that downlink bandwidth limits them and filtering on board gets results to disaster response or defence faster.

A moving target

The step from there to a data centre is large. With no air to carry heat away, a data centre in orbit must radiate it, and radiators weigh three to five kilograms per kilowatt. Running hotter would shrink them, but GPUs are not designed for high temperatures. Satellites cannot easily be maintained in orbit, so redundancy adds more mass, and AI accelerators are replaced every two to three years. For the sums to work, Karol said, launch would have to fall from around $3,000 a kilogram to perhaps $200.

Meanwhile better chips, memory and energy sources, fission and fusion among them, are being developed on Earth, so the bar that orbital compute has to clear keeps moving. Karol called the engineering problems solvable, as SpaceX's early problems proved to be. What mattered was the scale and launch price at which the unit economics start to work.

There are, in the end, so many levers that you can pull to make data centres in space viable. The question is: what's the timeline for that?

The links between satellites drew doubt. Data centres grew large, a participant said, because chips need high bandwidth between them, and satellites of a few kilowatts each would have to talk by laser. Karol agreed there is no solution yet. The nearest is a Chinese system of 12 satellites working as one. The participant saw a twist. If distributed computing solved the problem in orbit, it would also work on Earth, with inference running on phones, and then data centres would be needed in neither place.

Patrick takes the bull's side

Patrick then took the bullish side on purpose and invited a challenge. Compute in orbit is already growing, so the line between edge compute and a data centre will blur. Radiators will improve and launch costs will fall. "There are, in the end, so many levers that you can pull to make data centres in space viable. The question is: what's the timeline for that?" Patrick admitted it was a bullish stance, without promising to put money on it.

A participant who works with space-infrastructure start-ups described the strongest version. Earth cannot build power fast enough, so orbit is the only place left. They found a gradual path more likely, with compute added alongside the growing fleet of communications satellites. One of the hosts replied that fixing the bottlenecks on Earth would probably be faster.

Who pays until then

The money question came from a private-credit investor. Even the largest terrestrial data-centre deals in the US are starting to exhaust the capital market, they said, because funds can put only so much into data centres, and space needs far more. They accepted that operating in orbit might one day be cheaper. The trouble was the distance between that future and the capital steps needed to reach it.

A venture investor said their firm asks the same question. Many European start-ups in the field package NVIDIA GPUs for space and launch on someone else's rocket. The investor wondered whether that is a venture case at all or an infrastructure one. Infrastructure and debt investors, though, want unit economics that are secured and contracted, and these are not. One of the hosts asked whether incumbents such as SpaceX already hold the best positions in the value chain.

Launch was the other dependency. Starlink has registered about 1.6 million satellites with the ITU, one participant said. If SpaceX launches Starlink at that volume plus data-centre modules, Karol said, it will use up its own launch capacity, which makes Isar Aerospace's recent launch important for Europe. Karol found it embarrassing that Europeans must ask Elon Musk for a ride to orbit. European start-ups often struggle to get SpaceX slots, an investor said.

Karol expected the attempt to pay off anyway. Thermal management, radiation hardening and links between satellites carry over to other systems in orbit, and there will be buyers for sovereign and military-grade compute in space. The unit economics of full data centres, though, speak against them "at least for the next few years". One of the hosts called launching rockets close to commoditised. An investor looking for multi-billion outcomes disagreed. In European space today they would still back rockets and launch, along with vertically integrated satellite makers, because without launch Europe cannot do much. Before anyone builds a data centre in orbit, Europe has to be able to get there.

This Ripple was hosted by Patrick Tucci (Vsquared Ventures) and Karol Szubstarski (OTB Ventures) at The Drop 2026 on 16 September.