Why the economics of orbital AI are so brutal

In a way, this entire factor was inevitable. Elon Musk and his coterie have been speaking about AI in area for years—primarily within the context of Iain Banks’ science fiction collection a couple of far-future universe the place sentient spaceships roam and management the galaxy. 

Now, Musk sees a possibility to appreciate a model of this imaginative and prescient. His firm SpaceX has requested regulatory permission to construct solar-powered orbital knowledge facilities, distributed throughout as many as 1,000,000 satellites, that might shift as a lot as 100 GW of compute energy off the planet. He has reportedly recommended a few of his AI satellites will probably be constructed on the Moon.

“By far the most cost effective place to place AI will probably be area in 36 months or much less,” Musk stated final week on a podcast hosted by Stripe cofounder John Collison. 

He’s not alone. xAI’s head of compute has reportedly guess his counterpart at Anthropic that 1% of world compute will probably be in orbit by 2028. Google (which has a big possession stake in SpaceX) has introduced an area AI effort known as Venture Suncatcher, which can launch prototype autos in 2027. Starcloud, a start-up that has raised $34 million backed by Google and Andreessen Horowitz, filed its personal plans for an 80,000 satellite tv for pc constellation final week. Even Jeff Bezos has stated that is the long run.

However behind the hype, what is going to it truly take to get knowledge facilities into area? 

In a primary evaluation, as we speak’s terrestrial knowledge facilities stay cheaper than these in orbit. Andrew McCalip, an area engineer, has constructed a helpful calculator evaluating the 2 fashions. His baseline outcomes present {that a} 1 Gw orbital knowledge heart may cost $42.4B—nearly thrice its ground-bound equal, due to the up-front prices of constructing the satellites and launching them to orbit. 

Altering that equation, consultants say, would require expertise improvement throughout a number of fields, huge capital expenditure, and plenty of work on the provision chain for space-grade elements. It additionally is determined by prices on the bottom rising as sources and provide chains are strained by rising demand.

Designing and launching the satellites

The important thing driver for any area enterprise mannequin is how a lot it prices to get something up there. Musk’s SpaceX is already pushing down on the price of attending to orbit, however analysts what it can take to make orbital knowledge facilities a actuality want even decrease costs to shut their enterprise case. In different phrases, whereas AI knowledge facilities could appear to be a narrative a couple of new enterprise line forward of the SpaceX IPO, the plan is determined by finishing the corporate’s longest-running unfinished mission—Starship. 

Take into account that the reusable Falcon 9 delivers, as we speak, a price to orbit of roughly $3,600/kg. Making area knowledge facilities doable, per Venture Suncatcher’s white paper, would require costs nearer to $200/kg, an 18-fold enchancment which it expects to be out there within the 2030s. At that worth, nevertheless, the power delivered by a Starlink satellite tv for pc as we speak can be price aggressive with a terrestrial datacenter. 

The expectation is that SpaceX’s next-generation Starship rocket will ship these enhancements—no different car in improvement guarantees equal financial savings. Nonetheless, that car has but to turn into operational and even attain orbit; a 3rd iteration of Starship is predicted to make its maiden launch someday within the months forward. 

Even when Starship is totally profitable, nevertheless, assumptions that it’s going to instantly ship decrease costs to prospects could not cross the scent check. Economists on the consultancy Rational Futures make a compelling case that, as with the Falcon 9, SpaceX is not going to need to cost a lot lower than its finest competitor—in any other case the corporate is leaving cash on the desk. If Blue Origin’s New Glenn rocket, for instance, retails at $70 million, SpaceX received’t tackle Starship missions for exterior prospects at a lot lower than that, which would go away it above the numbers publicly assumed by area knowledge heart builders.

“There are usually not sufficient rockets to launch 1,000,000 satellites but, so we’re fairly removed from that,” Matt Gorman, the CEO of Amazon Net Providers, stated at a latest occasion. “If you consider the price of getting a payload in area as we speak, it’s huge. It’s simply not economical.”

Nonetheless, if launch is the bane of all area companies, the second problem is manufacturing price. 

“We at all times take without any consideration, at this level, that Starship’s price goes to be lots of of {dollars} per kilo,” McCalip advised TechCrunch. “Individuals are not making an allowance for the satellites are nearly $1,000 a kilo proper now.”

Satellite tv for pc manufacturing prices are the most important chunk of that price ticket, but when high-powered satellites might be made at about half the price of present Starlink satellites, the numbers begin to make sense. SpaceX has made nice advances in satellite tv for pc economics whereas constructing Starlink, its record-setting communications community, and the corporate hopes to attain extra via scale. A part of the reasoning behind 1,000,000 satellites is undoubtedly the associated fee financial savings that come from mass manufacturing. 

Nonetheless, the satellites that will probably be used for these missions have to be massive sufficient to fulfill the complicated necessities for working highly effective GPUs, together with massive photo voltaic arrays, thermal administration programs, and laser-based communications hyperlinks to obtain and ship knowledge.

A 2025 white paper from Venture Suncatcher presents one strategy to examine terrestrial and area knowledge facilities by the price of energy, the fundamental enter wanted to run chips. On the bottom, knowledge facilities spend roughly $570–3,000 for a Kw of energy over a 12 months, rely upon native energy prices and the effectivity of their programs. SpaceX’s Starlink satellites get their energy from on-board photo voltaic panels as a substitute, however the price of buying, launching, and sustaining these spacecraft delivers power at $14,700 per Kw over a 12 months. Put merely, satellites and their elements must get quite a bit cheaper earlier than they’re cost-competitive with metered energy.

The area atmosphere shouldn’t be playing around

Orbital knowledge heart proponents usually say that thermal administration is “free” in area, however that’s an oversimplification. With out an environment, it’s truly tougher to disperse warmth.

“You’re counting on very massive radiators to simply have the ability to dissipate that warmth into the blackness of area, and in order that’s plenty of floor space and mass that it’s a must to handle,” stated Mike Safyan, an govt at Planet Labs, which is constructing prototype satellites for Google Suncatcher which might be anticipated to launch in 2027. “It’s acknowledged as one of many key challenges, particularly long run.”

In addition to the vacuum of area, AI satellites might want to take care of cosmic radiation as nicely. Cosmic rays degrade chips over time, they usually may trigger “bit flip” errors that may corrupt knowledge. Chips might be protected with shielding, use rad-hardened elements, or work in collection with redundant error checks, however all these choices contain costly trades for mass. Nonetheless, Google used a particle beam to check the results of radiation on its Tensor Processing Items (chips designed explicitly for machine studying purposes). SpaceX executives stated on social media that the corporate has acquired a particle accelerator for simply that goal.

One other problem comes from the photo voltaic panels themselves. The logic of the mission is power arbitrage: Placing photo voltaic panels in area makes them anyplace from 5 to eight occasions extra environment friendly than on Earth, and in the event that they’re in the suitable orbit, they are often in sight of the solar for 90% of the day or extra, growing their effectivity. Electrical energy is the principle gasoline for chips, so extra power = cheaper knowledge facilities. However even photo voltaic panels are extra difficult in area.

House-rated photo voltaic panels made from uncommon earth parts are hardy, however too costly. Photo voltaic panels constructed from silicon are low-cost and more and more prevalent in area—Starlink and Amazon Kuiper use them—however they degrade a lot quicker as a consequence of area radiation. That may restrict the lifetime of AI satellites to round 5 years, which suggests they must generate return on funding quicker. 

Nonetheless, some analysts suppose that’s not such an enormous deal, primarily based on how rapidly new generations of chips arrive on the scene. “After 5 – 6 years, the {dollars} per kilowatt hour doesn’t produce a return, and that’s as a result of they’re not state-of-the-art,” Philip Johnston, the CEO of Starcloud, advised TechCrunch.

Danny Discipline, an govt at Solestial, a start-up constructing space-rated silicon photo voltaic panels, says the business sees orbital knowledge facilities as a key driver of development. He’s talking with a number of corporations about potential knowledge heart initiatives, and says “any participant who’s large enough to dream is a minimum of fascinated by it.” As a long-time spacecraft design engineer, nevertheless, he doesn’t low cost the challenges in these fashions.

“You may at all times extrapolate physics out to a much bigger measurement,” Discipline stated. “I’m excited to see how a few of these corporations get to some extent the place the economics make sense and the enterprise case closes.”

How do area knowledge facilities slot in?

One excellent query about these knowledge facilities: What’s going to we do with them? Are they basic goal, or for inference, or for coaching? Based mostly on current use circumstances, they might not be fully interchangeable with knowledge facilities on the bottom. 

A key problem for coaching new fashions is working 1000’s of GPUs collectively en masse. Most mannequin coaching shouldn’t be distributed, however performed in particular person knowledge facilities. The hyperscalers are working to alter this as a way to improve the ability of their fashions, however it nonetheless hasn’t been achieved. Equally, coaching in area would require coherence between GPUs on a number of satellites.

The staff at Google’s Venture Suncatcher notes that the corporate’s terrestrial knowledge facilities join their TPU networks with throughput within the lots of of gigabpits per second. The quickest off-the-shelf inter-satellite comms hyperlinks as we speak, which use lasers, can solely rise up to about 100 Gbps. 

That led to an intriguing structure for Suncatcher: It entails flying 81 satellites in formation so they’re shut sufficient to make use of the form of transceivers relied on by terrestrial knowledge facilities. That, after all, presents its personal challenges: The autonomy required to make sure every spacecraft stays in its appropriate station, even when maneuvers are required to keep away from orbital particles or one other spacecraft.

Nonetheless, the Google examine presents a caveat: The work of inference can tolerate the orbital radiation atmosphere, however extra analysis is required to grasp the potential affect of bit-flips and different errors on coaching workloads. 

Inference duties don’t have the identical want for 1000’s of GPUs working in unison. The job might be performed with dozens of GPUs, maybe on a single satellite tv for pc, an structure that represents a form of minimal viable product and the probably start line for the orbital knowledge heart enterprise. 

“Coaching shouldn’t be the best factor to do in area,” Johnston stated. “I believe nearly all inference workloads will probably be performed in area,” imagining the whole lot from customer support voice brokers to ChatGPT queries being computed in orbit. He says his firm’s first AI satellite tv for pc is already incomes income performing inference in orbit.

Whereas particulars are scarce even within the firm’s FCC submitting, SpaceX’s orbital knowledge heart constellation appears to anticipate about 100 kw of compute energy per ton28, roughly twice the ability of present Starlink satellites. The spacecraft will function in reference to one another and utilizing the Starlink community to share data; the submitting claims that Starlink’s laser hyperlinks can obtain petabit-level throughput. 

For SpaceX, the corporate’s latest acquisition of xAI (which is constructing its personal terrestrial knowledge facilities) will let the corporate stake out positions in each terrestrial and orbital knowledge facilities, seeing which provide chain adapts quicker.

That’s the good thing about having fungible Floating Level Operations Per Second – if you may make it work. “A FLOP is a FLOP, it doesn’t matter the place it lives,” McCalip stated. “[SpaceX] can simply scale till [it] hits allowing or capex bottlenecks on the bottom, after which fall again to [their] area deployments.”

Obtained a delicate tip or confidential paperwork about SpaceX? Attain out to Tim Fernholz at tim.fernholz@techcrunch.com. For safe communication, you’ll be able to contact him by way of Sign at tim_fernholz.21.