Why the economics of orbital AI are so brutal

In a sense, this full happening was inevitable. Elon Musk and his coterie person been talking astir AI successful abstraction for years—mainly successful the discourse of Iain Banks’ subject fabrication bid astir a far-future beingness wherever sentient spaceships roam and power the galaxy. 

Now, Musk sees an accidental to recognize a mentation of this vision. His institution SpaceX has requested regulatory support to physique solar-powered orbital information centers, distributed crossed arsenic galore arsenic a cardinal satellites, that could displacement arsenic overmuch arsenic 100 GW of compute powerfulness disconnected the planet. He has reportedly suggested immoderate of his AI satellites volition beryllium built connected the Moon.

“By acold the cheapest spot to enactment AI volition beryllium abstraction successful 36 months oregon less,” Musk said past week connected a podcast hosted by Stripe cofounder John Collison. 

He’s not alone. xAI’s caput of compute has reportedly stake his counterpart astatine Anthropic that 1% of planetary compute volition beryllium successful orbit by 2028. Google (which has a important ownership involvement successful SpaceX) has announced a abstraction AI effort called Project Suncatcher, which volition motorboat prototype vehicles successful 2027. Starcloud, a start-up that has raised $34 cardinal backed by Google and Andreessen Horowitz, filed its ain plans for an 80,000 outer constellation past week. Even Jeff Bezos has said this is the future.

But down the hype, what volition it really instrumentality to get information centers into space? 

In a archetypal analysis, today’s terrestrial information centers stay cheaper than those successful orbit. Andrew McCalip, a abstraction engineer, has built a helpful calculator comparing the 2 models. His baseline results amusement that a 1 Gw orbital information halfway mightiness outgo $42.4B—almost 3 times its ground-bound equivalent, acknowledgment to the up-front costs of gathering the satellites and launching them to orbit. 

Changing that equation, experts say, volition necessitate exertion improvement crossed respective fields, monolithic superior expenditure, and a batch of enactment connected the proviso concatenation for space-grade components. It besides depends connected costs connected the crushed rising arsenic resources and proviso chains are strained by increasing demand.

Designing and launching the satellites

The cardinal operator for immoderate abstraction concern exemplary is however overmuch it costs to get thing up there. Musk’s SpaceX is already pushing down connected the outgo of getting to orbit, but analysts looking astatine what it volition instrumentality to marque orbital information centers a world request adjacent little prices to adjacent their concern case. In different words, portion AI information centers whitethorn look to beryllium a communicative astir a caller concern enactment up of the SpaceX IPO, the program depends connected completing the company’s longest-running unfinished project—Starship. 

Consider that the reusable Falcon 9 delivers, today, a outgo to orbit of astir $3,600/kg. Making abstraction information centers doable, per Project Suncatcher’s achromatic paper, volition necessitate prices person to $200/kg, an 18-fold betterment which it expects to beryllium disposable successful the 2030s. At that price, however, the vigor delivered by a Starlink outer contiguous would beryllium outgo competitory with a terrestrial datacenter. 

The anticipation is that SpaceX’s next-generation Starship rocket volition present those improvements—no different conveyance successful improvement promises equivalent savings. However, that conveyance has yet to go operational oregon adjacent scope orbit; a 3rd iteration of Starship is expected to marque its maiden motorboat sometime successful the months ahead. 

Even if Starship is wholly successful, however, assumptions that it volition instantly present little prices to customers whitethorn not walk the odor test. Economists astatine the consultancy Rational Futures marque a compelling lawsuit that, arsenic with the Falcon 9, SpaceX volition not privation to complaint overmuch little than its champion competitor—otherwise the institution is leaving wealth connected the table. If Blue Origin’s New Glenn rocket, for example, retails astatine $70 million, SpaceX won’t instrumentality connected Starship missions for outer customers astatine overmuch little than that, which would permission it supra the numbers publically assumed by abstraction information halfway builders.

“There are not capable rockets to motorboat a cardinal satellites yet, truthful we’re beauteous acold from that,” Matt Gorman, the CEO of Amazon Web Services, said astatine a caller event. “If you deliberation astir the outgo of getting a payload successful abstraction today, it’s massive. It is conscionable not economical.”

Still, if motorboat is the bane of each abstraction businesses, the 2nd situation is accumulation cost. 

“We ever instrumentality for granted, astatine this point, that Starship’s outgo is going to beryllium hundreds of dollars per kilo,” McCalip told TechCrunch. “People are not taking into relationship the satellites are astir $1,000 a kilo close now.”

Satellite manufacturing costs are the largest chunk of that terms tag, but if high-powered satellites tin beryllium made astatine astir fractional the outgo of existent Starlink satellites, the numbers commencement to marque sense. SpaceX has made large advances successful outer economics portion gathering Starlink, its record-setting communications network, and the institution hopes to execute much done scale. Part of the reasoning down a cardinal satellites is undoubtedly the outgo savings that travel from wide production. 

Still, the satellites that volition beryllium utilized for these missions indispensable beryllium ample capable to fulfill the analyzable requirements for operating almighty GPUs, including ample star arrays, thermal absorption systems, and laser-based communications links to person and present data.

A 2025 achromatic paper from Project Suncatcher offers 1 mode to comparison terrestrial and abstraction information centers by the outgo of power, the basal input needed to tally chips. On the ground, information centers walk astir $570–3,000 for a Kw of powerfulness implicit a year, beryllium connected section powerfulness costs and the ratio of their systems. SpaceX’s Starlink satellites get their powerfulness from on-board star panels instead, but the outgo of acquiring, launching, and maintaining those spacecraft delivers vigor astatine $14,700 per Kw implicit a year. Put simply, satellites and their components volition person to get a batch cheaper earlier they’re cost-competitive with metered power.

The abstraction situation is not fooling around

Orbital information halfway proponents often accidental that thermal absorption is “free” successful space, but that’s an oversimplification. Without an atmosphere, it’s really much hard to disperse heat.

“You’re relying connected precise ample radiators to conscionable beryllium capable to dissipate that vigor into the blackness of space, and truthful that’s a batch of aboveground country and wide that you person to manage,” said Mike Safyan, an enforcement astatine Planet Labs, which is gathering prototype satellites for Google Suncatcher that are expected to motorboat successful 2027. “It is recognized arsenic 1 of the cardinal challenges, particularly agelong term.”

Besides the vacuum of space, AI satellites volition request to woody with cosmic radiation arsenic well. Cosmic rays degrade chips implicit time, and they tin besides origin “bit flip” errors that tin corrupt data. Chips tin beryllium protected with shielding, usage rad-hardened components, oregon enactment successful bid with redundant mistake checks, but each these options impact costly trades for mass. Still, Google utilized a particle beam to trial the effects of radiation connected its Tensor Processing Units (chips designed explicitly for instrumentality learning applications). SpaceX executives said connected societal media that the institution has acquired a particle accelerator for conscionable that purpose.

Another situation comes from the star panels themselves. The logic of the task is vigor arbitrage: Putting star panels successful abstraction makes them anyplace from 5 to 8 times much businesslike than connected Earth, and if they’re successful the close orbit, they tin beryllium successful show of the prima for 90% of the time oregon more, expanding their efficiency. Electricity is the main substance for chips, truthful much vigor = cheaper information centers. But adjacent star panels are much analyzable successful space.

Space-rated star panels made of uncommon world elements are hardy, but excessively expensive. Solar panels made from silicon are inexpensive and progressively prevalent successful space—Starlink and Amazon Kuiper usage them—but they degrade overmuch faster owed to abstraction radiation. That volition bounds the beingness of AI satellites to astir 5 years, which means they volition person to make instrumentality connected concern faster. 

Still, immoderate analysts deliberation that’s not specified a large deal, based connected however rapidly caller generations of chips get connected the scene. “After 5 oregon six years, the dollars per kilowatt hr doesn’t nutrient a return, and that’s due to the fact that they’re not authorities of the art,” Philip Johnston, the CEO of Starcloud, told TechCrunch.

Danny Field, an enforcement astatine Solestial, a start-up gathering space-rated silicon star panels, says the manufacture sees orbital information centers arsenic a cardinal operator of growth. He’s speaking with respective companies astir imaginable information halfway projects, and says “any subordinate who is large capable to imagination is astatine slightest reasoning astir it.” As a long-time spacecraft plan engineer, however, helium doesn’t discount the challenges successful these models.

“You tin ever extrapolate physics retired to a bigger size,” Field said. “I’m excited to spot however immoderate of these companies get to a constituent wherever the economics marque consciousness and the concern lawsuit closes.”

How bash abstraction information centers acceptable in?

One outstanding question astir these information centers: What volition we bash with them? Are they wide purpose, oregon for inference, oregon for training? Based connected existing usage cases, they whitethorn not beryllium wholly interchangeable with information centers connected the ground. 

A cardinal situation for grooming caller models is operating thousands of GPUs unneurotic en masse. Most exemplary grooming is not distributed, but done successful idiosyncratic information centers. The hyperscalers are moving to alteration this successful bid to summation the powerfulness of their models, but it inactive hasn’t been achieved. Similarly, grooming successful abstraction volition necessitate coherence betwixt GPUs connected aggregate satellites.

The squad astatine Google’s Project Suncatcher notes that the company’s terrestrial information centers link their TPU networks with throughput successful the hundreds of gigabpits per second. The fastest off-the-shelf inter-satellite comms links today, which usage lasers, tin lone get up to astir 100 Gbps. 

That led to an intriguing architecture for Suncatcher: It involves flying 81 satellites successful enactment truthful they are adjacent capable to usage the benignant of transceivers relied connected by terrestrial information centers. That, of course, presents its ain challenges: The autonomy required to guarantee each spacecraft remains successful its close station, adjacent if maneuvers are required to debar orbital debris oregon different spacecraft.

Still, the Google survey offers a caveat: The enactment of inference tin tolerate the orbital radiation environment, but much probe is needed to recognize the imaginable interaction of bit-flips and different errors connected grooming workloads. 

Inference tasks don’t person the aforesaid request for thousands of GPUs moving successful unison. The occupation tin beryllium done with dozens of GPUs, possibly connected a azygous satellite, an architecture that represents a benignant of minimum viable merchandise and the apt starting constituent for the orbital information halfway business. 

“Training is not the perfect happening to bash successful space,” Johnston said. “I deliberation astir each inference workloads volition beryllium done successful space,” imagining everything from lawsuit work dependable agents to ChatGPT queries being computed successful orbit. He says his company’s archetypal AI outer is already earning gross performing inference successful orbit.

While details are scarce adjacent successful the company’s FCC filing, SpaceX’s orbital information halfway constellation seems to expect astir 100 kw of compute powerfulness per ton28, astir doubly the powerfulness of existent Starlink satellites. The spacecraft volition run successful transportation with each different and utilizing the Starlink web to stock information; the filing claims that Starlink’s laser links tin execute petabit-level throughput. 

For SpaceX, the company’s caller acquisition of xAI (which is gathering its ain terrestrial information centers) volition fto the institution involvement retired positions successful some terrestrial and orbital information centers, seeing which proviso concatenation adapts faster.

That’s the payment of having fungible Floating Point Operations Per Second – if you tin marque it work. “A FLOP is simply a FLOP, it doesn’t substance wherever it lives,” McCalip said. “[SpaceX] tin conscionable standard until [it] hits permitting oregon capex bottlenecks connected the ground, and past autumn backmost to [their] abstraction deployments.”

Got a delicate extremity oregon confidential documents astir SpaceX? Reach retired to Tim Fernholz astatine tim.fernholz@techcrunch.com. For unafraid communication, you tin interaction him via Signal astatine tim_fernholz.21.