‘It seems crazy’: SpaceX’s orbital data centre bet tests physics and finance

SpaceX has sold investors on a trillion-dollar vision of millions of solar-powered satellites in orbit, packed with the chips needed to power AI.

The promise is a new market in orbital data centres — and a bid by Elon Musk’s company to dominate a critical layer of the AI economy.

SpaceX, which made its blockbuster stock market debut in June, expects to begin launching orbital data centres into space as soon as next year. Its 407ft-tall rocket known as Starship is central to this plan.

“This is not some sort of far-future, distant thing,” Musk told investors this month.

But company filings, technical documents and interviews with aerospace engineers suggest the company needs to solve three significant interlocking problems to achieve this at scale: how to launch rockets cheaply at unprecedented cadence, cool power-hungry computers in a vacuum, and regularly replace them with the latest chip technology.

The company’s own IPO filing makes clear how speculative its bet remains. “Many of our initiatives . . . including those to develop orbital AI compute at scale . . . involve significant technical complexity, unproven technologies or technologies that do not exist, and such initiatives may not achieve commercial viability,” it wrote.

The logistical challenges of such an endeavour were “tremendous”, said Christopher Smith, a senior aerospace engineer at UC Berkeley’s Space Sciences Laboratory.

To maintain a constellation of 1mn satellites, dubbed Starmind by SpaceX, Smith calculated that the company would need to launch more than nine of its Starship rockets a day. The calculation assumes that the chips in its orbital data centres would need to be replaced every five years, and that each rocket could carry 60 satellites.

Starship was built to carry far more, and significantly larger, satellites than previous launch vehicles, bringing down the cost of taking cargo to space. But it has yet to demonstrate full reusability and has failed on several test flights.

It has so far only completed eight successful launches in total, while SpaceX’s most prolific rocket, the Falcon 9, launched 165 times last year, a rate of one every two days.

SpaceX wrote in its IPO filing that it would eventually “be able to scale to thousands of [Starship] launches per year”, but did not say when it would reach that figure.

“It seems feasible that an initial constellation of AI satellites will be deployed by 2028,” said George Lordos, a space systems architect and lecturer at the Massachusetts Institute of Technology. “The question is the scale of that constellation.”

Musk on Monday posted on X that the first AI satellites, packed with chips from Nvidia, would launch in the fourth quarter of 2027, with plans for “significant scale” in 2028.

Lordos said SpaceX would need to “really drive down the turnaround time” for the reusable boosters and upper stages of the Starship rocket in order to build out a large network of satellites over the coming years.

A further challenge will be in finding a way to keep the data centres cool while in orbit.

While space is nominally very cold — around -270C — it is also a near-perfect vacuum. This means it is largely devoid of particles on which conduction and convection of heat depend.

SpaceX has published a mock-up on its website of the satellite it intends to use, which suggests the company will rely on deployable radiators with a total area of 160 square metres, roughly the same footprint as two badminton courts.

The company’s previous Starlink satellites have used passive cooling systems to remove heat, but the scale of power generated by each of its AI data centres — around 250kW at peak — means it requires a heftier set-up.

SpaceX plans to use pumps and a liquid-cooling loop to carry heat away from the central computing unit and into the radiators. The system is similar to that employed on the International Space Station, which pumps ammonia through its cooling system.

“What is being proposed is a ridiculously higher-performance version of this, but it’s doable,” said McKenzie Sandberg, a thermal engineer at Advanced Cooling Technologies who previously worked for SpaceX.

Sandberg said the most efficient way to cool orbital data centres would be to distribute chips across the surface of the radiators.

SpaceX has instead concentrated the chips within a central “compute” unit, a decision Sandberg said was likely linked to the challenge of shielding components from cosmic radiation.

Electronics in space are susceptible to damage from high-energy particles, which can hit chips and cause issues such as bit-flipping, whereby the 0s and 1s stored by electrical circuits are scrambled. This can have significant consequences such as data corruption, computer crashes or navigational issues.

Radiation shielding is likely to be the main way of protecting the electronics, according to spacecraft engineers, and confining this circuitry to a single unit would reduce the amount of material needed, saving precious mass.

Another problem SpaceX will need to contend with is the challenge of replacing satellites, both to keep up with technological developments on Earth and to remove satellites that fail in service.

So far on its Starlink programme, SpaceX has replaced satellites after roughly five years of service, de-orbiting the vessels and letting them burn up in the atmosphere.

But there are growing worries that allowing spacecraft to be incinerated upon re-entry is polluting the atmosphere with metals whose effects on the planet are not yet fully understood.

“There are huge questions when you ramp up what you are doing to the atmosphere by a factor of 100” over SpaceX’s existing launch programmes, said Jonathan McDowell, a former Harvard astrophysicist and expert on satellite launches.

SpaceX said in a May filing to US regulators that it planned to dispose of some satellites through a “controlled atmospheric re-entry” but asked for “flexibility” on the disposal method for others.

The company will also have to manage the increasing risk of space debris, which could collide with satellites.

McDowell said this would be particularly challenging because of the size of the orbital data centres’ solar panels, which he described as “a 700-square-metre debris collector”.

Even if these engineering problems can be solved, SpaceX must still make the economics work against Earth-based data centres.

MIT’s Lordos said the success of SpaceX’s orbital data centre plan hinged on its economic viability. The company has historically been able to reduce costs by taking control of its supply chain and has set about doing the same for Starship and its orbital data centres.

It began work in July on an eight-mile natural gas pipeline to its Texas launch facility, according to documents filed by a SpaceX affiliate to local authorities.

The move could help SpaceX expedite the transport of fuel to Starship, which uses around 3.4mn litres of liquid methane per launch, according to an analysis by Quilty Space of documents filed with the Federal Aviation Administration.

The fuel for each launch, equivalent by volume to more than one Olympic swimming pool, is currently delivered by truck to SpaceX’s launch site, according to Reuters.

SpaceX has also begun work on what it has called the “Gigasat Factory”, a new Texas-based facility that it says will be able to churn out “thousands of AI satellites . . . as soon as late 2027”.

The facility follows its announcement of Terafab, a vast $17bn chip fabrication facility in the same state, in partnership with Tesla and Intel. SpaceX noted in its IPO filing that its plans for “orbital AI” were contingent upon “our ability to access a sufficient number of AI chips, significantly more than are currently available to us”.

For now, though, the company said it expected “to continue sourcing a significant portion of our compute hardware from third-party suppliers”.

Some experts have questioned whether putting data centres into orbit was “the best way” to build computational capacity.

“It seems crazy,” said UC Berkeley’s Smith. “The logistical challenge of getting one million or even hundreds of thousands of large spacecraft in orbit at a cost that competes with the same project cost on the ground seems immense.”

But he said it would be folly to dismiss the plan altogether. “We’ve seen Elon break these conventions before.”