SpaceX is building data centres in orbit, and the numbers are serious

The biggest problem with AI data centres is not the chips. It is the heat. A hyperscaler data centre can spend up to 40% of its total energy just moving heat around, cooling the servers that run the models you talk to. SpaceX's answer to that problem is to put the servers in space, where the vacuum does the cooling for free.

That is the core logic behind the AI1 Compute Satellite. SpaceX has detailed a solar-powered orbital data centre designed specifically for AI inference workloads, running in low Earth orbit with 150kW peak power output and 120kW average compute capacity. There are no cooling towers, no chillers, no water loops. Heat radiates passively into the void. Near-constant sunlight in LEO means the power problem is largely solved by the Sun.

The Reflection deal and what it signals

This is not conceptual. SpaceX signed a computing deal with Reflection, an open-source AI startup, worth up to $6.3 billion. SpaceX has also filed with the FCC for a constellation of up to one million AI1-class satellites. That last number is ambitious to the point of sounding absurd, but SpaceX's existing Starlink constellation, now over 7,000 satellites, once sounded the same way.

Elon Musk has claimed space-based data centres will be more cost-effective than ground-based equivalents within two to three years. That timeline depends on launch costs continuing to fall as Starship becomes operational, manufacturing scaling for the compute hardware, and the downlink infrastructure working at the required throughput. SpaceX already has the last piece: Starlink provides the ground-to-orbit connection that makes the data centre commercially usable.

Why orbit actually makes sense for inference

Training an AI model requires moving enormous amounts of data between thousands of chips during the training run. That is hard to do with any meaningful latency. Inference, answering your prompts, is different. Each request is relatively self-contained, and the result can tolerate a small amount of round-trip latency without the user noticing. For an AI assistant responding in under two seconds, a few hundred milliseconds of additional orbital latency is manageable.

The economics of terrestrial AI infrastructure are also under real pressure. The demand for AI compute is growing faster than hyperscalers can build and cool new data centres. Power is constrained in many regions. Grid connections take years to secure. A satellite that can be manufactured, launched, and operational within months sidesteps most of those bottlenecks, at least in theory.

Competition is already here

A startup called Orbital filed FCC paperwork to deploy up to 100,000 data centre satellites targeting 10 gigawatts of compute. Meta is exploring selling excess AI compute capacity from its own infrastructure. The orbital AI infrastructure race is moving faster than most people expected.

Intel and SpaceX are also partners on Terafab, a planned $119 billion manufacturing facility in Austin slated to open in 2029, which would produce the chip hardware for AI satellites at scale alongside Tesla and SpaceX satellite production. That vertical integration, controlling the chip, the rocket, and the satellite, is what gives SpaceX a structural advantage that a pure-play data centre company cannot replicate.

What could go wrong

The obvious risks are failure rates and replacement costs. A ground-based server that fails costs a technician an hour. A satellite that fails costs a replacement launch. The redundancy model depends on having enough satellites in the constellation that individual failures do not matter, which is the same logic Starlink uses for its broadband service.

The latency question will also limit some use cases. Real-time AI applications where sub-100ms response is required are not candidates for orbital inference today. The applications that make sense are longer-horizon inference tasks, batch processing, and workloads where the marginal cost per computation matters more than round-trip speed.

If orbital compute reaches the cost curve Musk is projecting, it reshapes who can afford to run frontier AI models and where that compute physically lives. Concentrating that much infrastructure in the hands of the company that also owns the rockets is a question worth keeping in view.