AI Data Centers in Space: What Musk's Orbital Bet Says About AI Infrastructure

In March, Elon Musk stood in front of a friendly crowd and announced that SpaceX, now merged with his AI company xAI, intends to put data centers into orbit around the Earth. His reasoning was simple: on the ground, power is the bottleneck, but in space, he said, "it's always sunny." Google has a quieter version of the same idea underway, called Project Suncatcher, and a smaller company, Starcloud, has already flown a single Nvidia chip into orbit to prove the concept works at all.

Strip away the rocket footage and the pitch is really an admission. The biggest AI companies on the planet are running out of room to plug in more machines, and their answer is to leave the planet rather than fix the underlying problem. Global data-center electricity demand is expected to roughly double to nearly 1,000 terawatt-hours by the end of the decade, and executives are openly warning that chips will sit idle in warehouses for lack of power to switch them on.

That is the exact pressure Aphelion AI was built to relieve, just not by leaving Earth. Aphelion is a private enterprise AI platform built to deploy AI agents inside infrastructure a business already owns or exclusively controls, sized to that business's actual workload rather than to hyperscale ambitions. The orbital story is a useful, if extreme, illustration of what happens when AI strategy depends on ever more centralised infrastructure, and why a smaller, private, right-sized approach avoids the whole problem.

A Power Problem Too Big to Ignore

The scale of the constraint is hard to overstate. The International Energy Agency expects data-center power consumption to nearly double this decade, and that demand is concentrated in a handful of hyperscale regions already straining local grids. Some operators are building dedicated gas turbines. Others are turning to nuclear restarts. Philip Johnston, CEO of Starcloud, put the timeline bluntly: within six months, he said, companies will simply be leaving chips in warehouses because there is no power left to turn them on.

That is the environment space-based data centers are meant to solve. It is also the environment that makes any AI strategy tethered to shared, centralised infrastructure fragile. If your organisation's AI runs on capacity you rent from a provider who is themselves fighting for grid connections, your access to that capacity is only as secure as the provider's next power negotiation.

Why Orbit Looks Appealing, on Paper

The logic of space is genuinely attractive at first glance. Solar power is continuous, there is no grid to negotiate with, and land is not a constraint. Musk has proposed launching upward of a million satellites in polar orbit, including a first-generation "AI Sat Mini" with solar arrays roughly 180 metres across. Google's approach is more conservative: an 81-satellite cluster flown in tight formation with imaging partner Planet, with two prototypes due to launch in early 2027.

But the napkin math hides the engineering. The International Space Station, the largest power-producing structure currently in orbit, generates around 100 kilowatts from solar panels roughly half the size of a football field. Replicating a modest 100-megawatt terrestrial data center in space would require a facility 500 to 1,000 times that size. Heat is an equally hard problem: space is a vacuum, so a satellite cannot shed heat the way a building on Earth can, and every watt of compute needs a matching radiator to dump the resulting heat somewhere.

• Launch cost: putting a satellite into orbit currently costs around $1,000 per kilogram, and Google itself believes that figure needs to fall by a factor of five before space-based compute is economically sensible.
• Latency: smaller satellite constellations reduce the heat and power problem per unit, but they then need to exchange huge volumes of data over laser links between satellites, and even light-speed transfer between orbiting nodes adds delay that slows computing down.
• Maintenance: a typical terrestrial facility like DataBank's 144,000-square-foot IAD1 site in Virginia has vendors on-site every single day, installing servers and fixing hardware. Getting a technician to a satellite is not an option, which pushes far more of the burden onto software and pre-launch testing.

Even the people building these systems are candid about the odds. A Carnegie Mellon researcher who specialises in satellite computing called Musk's two-to-three-year timeline "an optimistic interpretation," and an MIT astronautics professor put a working orbital data center of any real scale at further out than three years, "certainly."

The Case for Staying Grounded, and Private

An enterprise does not need a satellite constellation, or even a share of a hyperscale campus, to run AI well. Most organisations' actual AI workload, a private assistant handling internal knowledge, customer queries and day-to-day operations, is a rounding error next to the frontier training runs driving the industry's power crisis. That gap is exactly what makes a private, right-sized deployment practical today rather than speculative for 2027 and beyond.

Aphelion's private AI agent runs on infrastructure the business already has or directly controls, not on a slice of a shared facility competing for grid capacity against every other tenant. Because the deployment is sized to one organisation's real usage, it never needs the multi-megawatt draw, the industrial cooling, or the speculative launch schedule that space-based and hyperscale approaches both depend on. The result is an AI capability that is available now, not contingent on a rocket program succeeding.

This also changes the risk profile in ways that matter beyond electricity. Centralised infrastructure, whether it sits in a shared data center or in orbit, means a business's prompts and documents travel to hardware it does not control. Aphelion keeps that data inside infrastructure the organisation governs directly, which is why data enrichment and private hosting sit at the centre of the platform rather than as an afterthought. Compliance work becomes a matter of reviewing your own controls, not vetting a third party's facility, whether that facility is in Virginia or in low Earth orbit.

"No one in data center land is losing any sleep," said Raul Martynek, whose company operates 75 terrestrial data centers, when asked whether orbital compute threatens his business. The bigger lesson for enterprise AI buyers is similar: the practical answer to today's infrastructure strain is not a more ambitious centralised facility, it is a smaller, private one you actually control.

What This Means for Your AI Strategy

Whether or not orbital compute ever becomes commercially viable, the underlying lesson for any business planning its AI infrastructure holds regardless:

• Dependence on shared, centralised capacity is a supply risk, not just a cost one. If your AI runs on infrastructure a provider is fighting to power, your access is only as secure as their next energy deal.
• Right-sizing beats scaling up. Most enterprise AI workloads do not need hyperscale power draw, they need a reliably running private agent connected to the systems the business already uses.
• Physical control of infrastructure is also data control. Wherever the hardware sits, routing sensitive prompts and documents through infrastructure you do not own is an exposure, and one that is straightforward to remove.
• Integration should not be sacrificed for independence. A private deployment only works day to day if it connects cleanly to the CRMs, ERPs and document stores already in use, which is why Aphelion treats system integration as a core capability rather than a bespoke add-on.

The space data center story is a striking one, and it may well prove out over a longer horizon than its proponents currently admit. But it is also a clear signal that the industry's default response to AI's growing infrastructure appetite is to go bigger and more centralised, whether that means a larger terrestrial campus or a constellation of satellites. Aphelion's answer is the opposite: build a private AI agent sized to what a business actually needs, running on infrastructure it already controls, so growth in AI usage never depends on someone else's power supply, launch schedule, or grid connection.

Frequently Asked Questions

What are AI data centers in space and why are companies pursuing them?

AI data centers in space are proposed satellite constellations that run AI chips in orbit, powered by continuous solar energy rather than a terrestrial grid connection. SpaceX, Starcloud and Google are each pursuing versions of the idea because ground-based AI infrastructure is running into a global power ceiling, with data center electricity demand expected to roughly double by 2030. The pitch is that space offers effectively free, uninterrupted power, though cooling, launch costs and satellite size remain unresolved engineering problems.

How does Aphelion AI avoid the power and infrastructure constraints driving the space data center race?

Aphelion sidesteps the problem by design rather than by orbit. A private AI agent is sized to a single organisation's real workload and deployed on infrastructure that organisation already owns or controls, rather than requiring a share of a hyperscale facility drawing tens of megawatts. Because the deployment is not competing for grid capacity against every other AI customer on a shared platform, an enterprise is never waiting on a utility upgrade, a new turbine, or a satellite launch schedule to run its own AI reliably. You can read more about how the agent is deployed on the AI Agent page.

Is a private, on-premises AI deployment more secure than relying on centralised or space-based infrastructure?

Yes. Centralised infrastructure, whether a shared hyperscale data center or a speculative orbital cluster, means an organisation's prompts, documents and outputs travel to and are processed on hardware it does not control. Aphelion keeps that data inside infrastructure the business governs directly, which removes the exposure that comes with routing sensitive information through a third-party facility, and simplifies compliance work under frameworks like GDPR, HIPAA and ISO because there is no external infrastructure provider to vet.

Does Aphelion's model still integrate with the business systems a company already runs?

Yes. A private deployment does not mean an isolated one. Aphelion treats system integration and data enrichment as core platform capabilities, connecting a private AI agent to the CRMs, ERPs, databases and document stores an organisation already operates, so the same infrastructure independence that protects against power and supply shocks does not come at the cost of everyday usefulness.

Private, right-sized AI infrastructure vs hyperscale or orbital data centers: which is the safer bet for enterprises today?

For most enterprises, right-sized private infrastructure is the safer and more immediate option. Orbital data centers depend on unproven cooling, satellite mass and launch-cost breakthroughs that even proponents place years out, while hyperscale terrestrial capacity is itself power-constrained and increasingly expensive to secure. A private deployment scoped to one organisation's actual usage needs a fraction of the power and none of the speculative engineering, and it can be running today rather than sometime after 2027. You can learn more about the team behind that approach on our About page.