Is space a sensible place to put a data center? SpaceX went public in June 2026 at close to two trillion dollars, with orbital compute part of the story it told investors, and Musk has floated launching more AI capacity each year than currently exists on the ground. That makes the question worth working through properly. Not whether this is possible, because it is, but what would have to change before anyone should build it.

TLDR: the physics is harder than it looks, the economics are the actual gate, and the problems standing in the way have surprisingly little to do with rockets.

What this research covers

• Why cooling in space is harder than cooling on Earth
• The launch price that would make orbital compute competitive (and how far off it actually is)
• How much compute low Earth orbit can physically hold
• Why frontier model training does not work up there, and what does
• The blockers worth watching and whether anyone knows how to solve them
• Who is actually building this

What to expect

• A threshold you can check launch prices against as they fall
• A clear line between the orbital workloads that work now and the ones that do not
• Why the two most promising hardware fixes will land in terrestrial data centers first
• Where we believe the real companies are

What is a space data center?

A constellation of satellites, each carrying somewhere between a few dozen and a few hundred GPUs, talking to each other over laser links, and working as a distributed cluster in low Earth orbit.

Why move compute to orbit?

Space is cold. About three degrees above absolute zero, if you measure it by the faint radiation left over from the Big Bang. So the intuitive case for orbital data centers more or less writes itself. Take the machines straining every power grid and water table on the planet, and move them somewhere with free sunlight and infinite cold.

Half of that is exactly right. The sunlight is free, it is stronger above the atmosphere than below it, and in the correct orbit it never stops. Put a satellite on the terminator, the moving line that divides day from night on Earth's surface, and its solar panels see the sun continuously while never passing into Earth's shadow. The cold is the half that fails, and understanding why is most of the story.

Can you cool a data center in space?

Cooling is considerably harder in orbit, and the reason is that cold is not a thing you can use.

On Earth a processor sheds heat into water, which carries it away. In vacuum there is nothing to carry anything. Every watt must leave as infrared radiation, roughly a hundred times less effective per unit of surface than water in a cold plate. Vacuum is not a coolant; it is an absence.

One high-end AI rack draws about as much power as a hundred homes. Shedding that heat takes something like a hundred square meters of panel, and powering it takes five times more solar array again: roughly a fifth of the International Space Station's entire solar wingspan, for one cabinet of GPUs. That geometry is why an orbital data center cannot be a single facility. The panel area forces each unit to stay small, and the cluster to be a swarm.

What would launch costs need to fall to?

Around two hundred dollars a kilogram. Today the figure is closer to three thousand, which is why nobody has built this. At current prices the transport bill for a real constellation runs into the trillions, less a financing problem than a category error. Cut launch cost tenfold and the same deployment costs a year or two of the sector's total revenue, which capital markets can absorb. Starship's stated target sits well under a hundred dollars a kilogram; add healthy margins on third-party launches and you land almost exactly on two hundred, which suggests their published target is less an aspiration than a derivation.

We also worked out how many of these satellites low Earth orbit can physically hold, building up from radiator area and orbital mechanics without reference to anyone's filings. The range we arrived at brackets what SpaceX has actually filed for, at the same power per satellite.

Why will the best solutions reach Earth before space?

Two technologies would materially improve the economics above: proton-gated transistors, which yield prototypes an order of magnitude smaller and less power-hungry, and heterogeneous computing, which roughly doubles the useful work extracted from accelerators that otherwise sit idle. Neither is a space technology. Both improve terrestrial data centers just as much, and chip designers and hyperscalers will deploy them on the ground first, because upgrading a building is far easier than upgrading a constellation.

What has to be true

Orbital compute does not need a single breakthrough. It needs four, and only one involves rockets.

Launch has to get roughly ten times cheaper. Servicing has to become routine, which means robots that do not exist yet, with fine motor control in zero gravity and the software to orchestrate them. Interconnect has to close the hundredfold gap between a laser link and what a single cabinet achieves on Earth, or model architectures have to stop needing it. And the hardware has to get dramatically more capable per watt, which points back at protonic devices and heterogeneous compute.

Each of those looks to us less like a feature and more like a company. On-orbit servicing robotics, protonic and sub-threshold compute, heterogeneous-compute orchestration, memory-sparse architectures are the four areas we are researching. Whether this technology race ends in space or on the ground is genuinely open, and it is the very thing we are watching. Read the paper for the full analysis.

‍

‍

_{DISCLAIMER: All views expressed are Hivemind’s own views. The information provided herein has been produced and issued by Hivemind Capital Partners UK LLP and/or Hivemind Capital Partners LLC (“Hivemind”) and is being provided for informational purposes only. This document is not to be distributed or reproduced in any way. This document does not constitute or contain an offer to purchase or sell securities. This document is confidential and intended for the person to whom this was delivered. If you have not received this document from Hivemind you are hereby notified that you have received it from a non-authorized source and you are prohibited from reading, using, retaining, disseminating or copying this material without the prior express written consent of Hivemind. Neither Hivemind nor any of its affiliates or representatives makes any representation or warranty, express or implied, as to the accuracy or completeness of the information contained herein or any other written or oral communication transmitted or made to the recipient. The information contained in this document is current as of the date indicated, and Hivemind undertakes no obligation to update, modify or amend this document or to otherwise notify a reader in the event that any matter stated herein changes or subsequently becomes inaccurate. This document has not been compiled, reviewed, or audited by an independent accountant. Past performance should not be construed as an indicator of future results, and there can be no assurance that historical trends will continue. This document does not include information regarding each investment or investment strategy pursued by the Funds. References to investments included herein should not be construed as a recommendation of any particular investment .Certain information contained herein may constitute “forward-looking statements,” which can be identified by the use of forward-looking terminology such as “may,” “will,” “should,” “expect,” “anticipate,” “project,” “estimate,” “intend,” “continue,” or “believe,” or the negatives thereof, other variations thereon or comparable terminology. All such forward-looking statements are solely statements of opinion, and there is no assurance that they will be predictive of actual events.}