Space Data Centers: The $500 a Kilo Question

A space data center is a satellite carrying ordinary data center GPUs: solar panels for power, a radiator to shed heat, laser links to move the data. Two have flown: Starcloud put an Nvidia H100 into orbit in November 2025, and a Google prototype carrying TPUs reached orbit on 1 October 2026. Nothing is operating at commercial scale. Whether anything ever does comes down to a single number.

That number is missing from most of the argument. Ask whether data centers in space make sense and you get two confident camps, one saying orbit is where AI compute is going because the sun never sets and nobody files a zoning complaint, the other saying launch costs make the whole idea absurd. Both argue in adjectives.

The founder actually building it argues in arithmetic.

Philip Johnston is co-founder and CEO of Starcloud, which is putting GPUs in orbit and selling the compute. On Y Combinator's The Lightcone he walked through how the company was derived, and the method is worth more to most founders than the destination. He did not start with a passion for space hardware. He started with a cost curve that was obviously falling, asked what becomes sensible on the other side of it, and then picked the one business on that list that did not require getting anything back down.

What follows is the founder-facing version: the single number the thesis lives or dies on, why compute goes to orbit before anything else does, what is genuinely unsolved, what the objectors have filed with the regulator, and the two process moves that transfer whether or not you ever touch a rocket.

What a space data center actually is

The hardware is less exotic than the phrase suggests. A space data center is a satellite, or a constellation of them, carrying standard data center GPUs instead of cameras or radios: solar panels for power, a radiator that dumps waste heat into the vacuum as infrared, and optical laser links that carry the data up and the results back down. The work it does is ordinary, model inference mostly. The reason to do it in low Earth orbit is not that computing is better there. It is that the inputs are cheaper there, and getting steadily harder to buy on the ground.

Starcloud launched its first satellite, Starcloud-1, on a Falcon 9 in November 2025. It carried five GPUs, the headline one an Nvidia H100, widely reported as the first data center class GPU to operate in orbit, and the company says it trained nanoGPT on it. The cooling was improvised by Johnston's own account: the assembly was submerged in a phase change material, and the thermal cycling test that normally books a vacuum chamber was done overnight in an ice bath and with hot air guns the morning the hardware shipped. The verdict from the room: "It is a miracle that it works to be honest."

That is the demo. The business is the constellation behind it.

The entire thesis is one number

Johnston's cleanest version of the argument is not about space at all. It is a line-by-line comparison against the cheapest power on Earth.

Build a solar project to power a new data center terrestrially and you pay three main costs: permitted land, which he says is the largest single cost for most new North American solar projects, battery storage and backup power for the roughly 20 hours a day you are not at peak, and the solar cells themselves.

Now price the same project in orbit. "we don't need to pay for permitted land. So your biggest cost is gone." In a dawn-dusk sun-synchronous orbit you are in sunlight continuously, so the batteries go too. And "1 square meter of solar panel in space produces eight times the energy of 1 square meter of solar panel on Earth," so you need roughly an eighth of the cells.

Three costs down. One cost added, and it is the only one that matters: "the only additional cost or the main additional cost we have in space is the launch cost."

That turns a sprawling argument into an arithmetic problem with a break-even. Johnston's team spent a month on it and got "a much closer to reality number of $500 a kilo." Below that launch price, orbit wins on energy economics. Above it, it does not.

The discipline here is what to copy. He had already run the same calculation for space-based solar, the beam-power-down idea he actually started with, and got roughly $50 per kilo, a number far enough from reality to kill it. Same method, different answer, and the answer reassigned the company. He also named the physics that killed the first idea rather than talking around it: "The problem with space-based solar is you actually lose 95% of the energy in transmission from space to Earth."

Where is launch pricing today? SpaceX publishes rideshare at $350,000 for the first 50 kg and about $7,000 per kilogram beyond that, and a dedicated Falcon 9 works out near $3,200 per kilogram at its full rated payload. So $500 is about a 14x cut from the published rideshare rate, or roughly 6x from a dedicated Falcon 9; Johnston rounds the gap to about 10x. Starship is the vehicle expected to close it, and he puts its design target at roughly $10 to $20 per kilogram. Treat that as SpaceX's stated ambition rather than an observed price: Starship has no commercial per-kilogram rate yet, and published estimates of its eventual cost span an order of magnitude. The honest statement of the thesis is that it is a bet on one vehicle's cost curve, and Johnston says so plainly, noting the company is "still not out of the woods" on low-cost launch.

If you are building anything capital-intensive, this is the artifact to steal. Not the conclusion. The falsifiable threshold: one number, stated in public, that tells you and everyone else when you are wrong.

Why data centers go to space before mining or manufacturing

Johnston looked at several orbital businesses before this one: in-space manufacturing, asteroid mining, even space hotels. The filter that selected data centers is a single sentence, and it is a business-model insight rather than a technical one.

"Data centers makes a lot of sense to be the first thing you do because you don't need to re-enter a product."

Mining, manufacturing and tourism all have to bring something back through the atmosphere, and "they require this very expensive re-entry process." A data center's product is a packet. It leaves by laser. The hardest, most expensive, least mature part of the whole space value chain simply does not appear in the business.

That is a general move: among the businesses a new capability makes possible, the first viable one is usually the one that touches the fewest unsolved problems. Cheap launch makes a dozen things conceivable. Only one of them ships its output over a link you already know how to build.

The scale ambition follows the same logic. Starcloud has filed with the FCC for authority to operate up to 88,000 satellites, each about 200 kilowatts, which the company puts at roughly 20 gigawatts of compute capacity. For reference, Johnston gives the largest data center deployments on Earth today as about a gigawatt. Fifty of the satellites fit per Starship, so roughly 10 megawatts of new capacity per launch. He pegs the capex for the full constellation at "a hundred billion dollars of capex spend," which he argues is lower than the terrestrial equivalent. His framing of the whole programme: "this really is the start of the largest infrastructure project ever."

Note what it is for. Almost entirely inference, because "inference is going to be like 99% of the compute market very soon anyway," and because training needs a single contiguous machine that would have to be docked together in orbit. On that, he is unusually candid about the timeline: "it'll probably be at least 15 years before we get to anything like that."

What is actually still unsolved

Two problems, and Johnston says the engineering team is split roughly 50/50 between them: "how do you get rid of this heat in a vacuum" and "how do we make the chips work in a higher radiation environment."

Heat is counterintuitive. Space is cold but empty, so there is no air to carry warmth away. The only exit is infrared radiation off a surface, which means the radiator area is the design constraint. His napkin math: solar panels yield around 200 watts per square meter, a radiator held near 50 degrees C sheds about 800, so you carry roughly a quarter as much radiator area as solar panel area. A 400 square meter array needs about 100 square meters of radiator on top of it. The lever is temperature, because the Stefan-Boltzmann relation makes dissipation scale with the fourth power of it, so running the chips hotter shrinks the radiator substantially. Starcloud's claim is a deployable radiator "at least 10 times less mass per watt of dissipation than the ISS radiator."

Radiation is brute-force empirical. The company runs chips through a cyclotron in Knoxville for high-velocity protons and Brookhaven National Lab for heavy ions, compressing a five-year dose into 24 hours, then uses the telemetry to pick shielding and write software mitigations. The by-product is a genuine proprietary asset, and his own framing keeps the hedge: "I think we're the only people in the world now that know where both an H100, a B200, H200 will fail" under that bombardment. That data is why Nvidia is a design partner. At GTC in March 2026 Nvidia announced the Space-1 Vera Rubin Module, built for orbital inference and claimed at up to 25 times the AI compute of an H100 for space workloads, with Starcloud among the named partners. The terrestrial version of that same platform question, what inference silicon is actually optimised for and where it mismatches agent workloads, is the subject of why inference chips are built for the wrong job.

Neither problem is physics-breaking. Both are manufacturing and cost problems, which is the shape a startup can actually attack. The ISS already flies a liquid-cooled radiator; Starcloud's job is to make one cheap and light, and the benchmark it cites is a vehicle nobody built for cost.

The regulatory tailwind is real, and smaller than the pitch

This is where the interview needs a fact-check, and where the honest version is more useful to a founder than the pitch version.

A YC partner puts the thesis to him directly: AI data centers "are becoming politically incredibly unpopular," possibly to the point where building them in democratic countries gets hard. Johnston agrees and says the shift moved faster than he expected, pointing at New York, and concludes "building these things in space will definitely be much easier from a regulatory perspective."

The direction is right. The specifics are different from how they sound in the room. On 14 July 2026, Governor Kathy Hochul signed an executive order establishing a temporary statewide moratorium pausing state environmental permits for new hyperscale data centers, those drawing 50 megawatts or more. New York was the first state to do it. It is a permitting pause on the largest projects, not a ban on data centers; projects whose permits were already complete were unaffected; and its length is open-ended rather than fixed, billed as up to a year but formally running until the state finishes a generic environmental impact statement.

His "seven other states" aside has since half come true, in a form weaker than a ban and stronger than nothing. New York is still the only state with a formal statewide moratorium. But in September 2026 Governor Greg Abbott directed the Texas Commission on Environmental Quality to stop issuing air and water permits for data center projects statewide, with no megawatt threshold, while the grid operator and the state water board audit electricity and water use. Maine's governor, meanwhile, vetoed an outright moratorium bill in April 2026. So the regulatory risk he is pricing is real and spreading, and it is arriving as conditional pauses tied to audits rather than as the bans the pitch implies. His characterization of the reasoning as not grounded in science is his opinion, not a finding.

So the tailwind exists and is worth underwriting. It is also two pauses, both tied to studies that will report, against a company whose first revenue at scale is years out. If your deck leans on a regulatory trend, state it at the resolution the record supports. The version that survives contact with a diligent investor is the smaller, true one. The broader case for treating local disposition as an input to a physical build, rather than a footnote, is the subject of deep tech startups and the clock that is their real moat, which works the terrestrial side of this problem: what to do when the bottleneck is a permit and a supply chain rather than a cost curve.

What the objectors filed against orbital data centers

An honest read of this thesis has to include the people formally arguing against it, because they are not anonymous commenters. They filed with the FCC.

The Secure World Foundation, the American Astronomical Society through its COMPASSE committee, the Center for Space Environmentalism, DarkSky International, the satellite operator Viasat and NASA itself have all filed on Starcloud's application, opposing it or asking for far more scrutiny. The objections, in their terms: the request is precedent-setting and short on technical detail about cumulative collision risk and aggregate radio interference; a constellation this size implies a continuous re-entry cycle whose atmospheric effects are not characterized; and sun-synchronous terminator orbits are exactly the orbits that ruin twilight observing for sensitive telescopes. The asks differ, which is worth noticing. Secure World wants the waivers denied or deferred and authorization staged against demonstrated performance. The astronomers want the application denied or stayed pending an independent study of aggregate impact. DarkSky International has pressed for environmental review under NEPA. Viasat argues the Ka-band coexistence case has not been made. NASA, the objection a founder should read first, asked for coordination to protect crewed spacecraft and a clearer account of automated collision avoidance. The economics have been challenged directly too: in a technical critique of the single-Starship cost case, Angadh Nanjangud argues that the mass a workable thermal design implies pushes the launch count, and therefore the cost, well past the company's own figures.

Johnston's answer on debris is specific rather than dismissive. Starcloud-1 flies low, around 400 km on his account and lower still at deployment by independent tracking, and he argues that an orbit that low decays on its own, so a collision there does not seed a lasting debris field. He points at SpaceX operating roughly 10,000 satellites without a collision as the existence proof for collision avoidance at scale, and says the apparent congestion in satellite maps is an artifact of plotting each spacecraft as a dot the width of California. He does not have an answer on the light-pollution objection in this interview, and it is the objection least likely to be solved by flying lower.

He is also not alone, which cuts both ways. SpaceX filed with the FCC on 30 January 2026 for an Orbital Data Center System of up to a million satellites between 500 and 2,000 km, the largest application in the field by an order of magnitude, and the entrant Johnston says surfaced mid-round. Blue Origin filed in March 2026 for Project Sunrise, up to 51,600 orbital computing satellites. Google's Project Suncatcher, built with Planet Labs, put its first prototype satellite carrying TPUs into orbit on 1 October 2026 to test how the chips survive launch, radiation and vacuum. When the incumbents file, the idea stops being a contrarian bet and becomes a race, and a race changes which risks matter.

Book the launch before you know what you are launching

Here is the part that transfers cleanly, to any startup, in any industry.

Starcloud was founded on 1 January 2024. On 2 January 2024, before there was a design, a prototype, or a decided payload, they booked a SpaceX rideshare slot. Johnston's framing: "The first thing every space company should do is book the first available launch they can," because "Booking a launch is such a good forcing function for a space company."

Two details make this method rather than bravado.

It was cheap. Johnston on the price of the slot: "it was cheap. It's like 300 grand", which is roughly SpaceX's published rideshare base price for the first 50 kg. Set against the $2 million first round, which was still months from closing when they booked, the forcing function cost about 15 percent of the money they were raising, which is a lot and still less than one bad quarter of indecision.

It was far enough out to be real. The slot was 18 months away and slipped to 21, so the deadline was aggressive but not absurd. The result, in his words: "something is going to be on that rocket. I'm not sure 100% sure what it's going to be at this point."

What the deadline bought is the interesting part. It converted every design argument into a scheduling decision, and the payload got more ambitious under that pressure rather than less: the original plan was a Jetson module, which had flown before and would have proved little, and a new co-founder pushed for the H100 nobody believed could run in orbit. A date on a calendar is what made that a decision instead of a debate.

This is a different claim from compressing a schedule, which is the moat argument in deep tech startups: buying a deadline is what you do before you have anything to compress. The generalizable bit is the economics. Find a commitment that is cheap in money, expensive in reputation, and dated: a booked slot, a ship date promised to a design partner, an accepted conference talk, a pilot contract with a start date. Software founders have no rockets, so they substitute pressure that costs nothing and therefore buys nothing. Johnston traded about 15 percent of a $2 million round for a hard date. Priced against what the deadline reorganized, that was cheap.

He recruited the engineers before he had the idea

The second transferable move inverts the usual sequence, and it is the one that explains why the company survived its own unpopularity.

Johnston's background does not read as space: math and physics, then software engineering, then McKinsey, where he worked with national space agencies on satellite missions and noticed launch costs falling. So he ran the YC advice literally. "I was kind of blindly following the YC advice. So I was like, Okay, the first thing we need to do before coming up with an idea is get the most incredible space engineers in the world."

He had no idea yet. The pitch was the question itself: "if you have any ideas that would make sense if the launch cost was 10x less than it is today." He messaged a large number of space engineers: "Maybe 10 of them took a call and two of them said yes."

The two who actually became co-founders did not come off that cold list, which is the part worth copying. Ezra Feilden, now CTO, had grown up in the same place in the UK as Johnston; he took a materials engineering PhD at Imperial College London and worked on spacecraft mechanisms at Airbus Defence and Space and then at Oxford Space Systems. Adi Oltean, now chief engineer, arrived through a friend of a friend and had already been thinking about the idea, to the point of registering a domain for it; he spent more than two decades as a principal software engineer at Microsoft before two years in the same role at SpaceX. The cold outreach built the surrounding network. The warm paths produced the founding team. The first idea the three generated together was space-based solar, which the arithmetic then killed, and the team outlasted the idea, which is the entire point.

The hiring discipline afterwards reads as slow to the point of discomfort. After an $11 million round at a $40 million valuation, "it took us 6 months to make our first hire", and that first hire was a space engineer. At the time Starcloud-1 launched, "There was only 12 in the team." After raising far more, "We're still only 20 engineers," plus one commercial hire out of the Space Force. His account of why: "we are so picky about who we hire. Like it is to a point of being very frustrating to be honest, but it's the whole game."

It is worth being clear-eyed that this is survivorship-flavored advice. Being picky is cheap to claim after a $2.3 billion valuation. What makes it credible here is that it is the same claim he made about why his investor said yes. Benchmark's diligence, as he tells it, landed on the engineering team rather than on the feasibility study, and he thinks a report declaring orbital cooling impossible would not have changed the outcome, because the bet was on a team good enough to figure it out.

Which is the same bet he made on himself. His closing advice is a single ordering: "technical talent on the founding team is the thing you need to solve for first."

What the rejections are worth

The fundraising history is useful mostly as calibration on how little signal a no carries.

The first round, $2 million at a $10 million post-money cap, took three months and, by his count, "we got rejected from at least 100 VCs." After YC's demo day they "got rejected from at least 20 VCs" before the first check. For a long stretch the reaction was "this is the dumbest thing I've ever heard." His read: "if you had listened to what all the investors were telling you, you would have quit and not ever done this idea."

What changed was not the pitch. Launch capacity became legible and building terrestrially got visibly harder. He also credits a broader swing toward hard tech, on the theory that "people think that software doesn't have a moat anymore which I think is probably accurate," a view a software founder should weigh rather than adopt. Even the round that worked was messier than it looks: SpaceX disclosed mid-process that it was pursuing the same business, so "anybody with like a conflict policy then was like okay we're conflicted out."

The outcome, as of now: a $170 million Series A at a $1.1 billion valuation led by Benchmark and EQT Ventures in March 2026, reported as the fastest a Y Combinator company has reached unicorn status, 17 months after demo day; then a $250 million extension in August 2026 at $2.3 billion with Nvidia among the investors, taking total funding to roughly $450 million. In September 2026 the company signed Firefly Aerospace to carry an AI computing payload to lunar orbit, targeted for no earlier than 2028.

At least 120 rejections preceded all of it. They were not wrong about the risk. They were wrong about the trend line, which is the only thing the founder had been underwriting the whole time.

What this is worth if you will never go to space

Strip out the orbit and four moves remain, and they apply to a seed-stage software company as well as a satellite company.

Pick an input cost that is falling for reasons outside your control, and build the business that becomes sensible at the far end of the curve. Johnston's was launch cost. For most founders in 2026 it is inference cost, and the question is identical: what product makes no sense at today's cost per token and obvious sense at a tenth of it. What actually moves AI inference cost is where to start if you want current numbers rather than a guess.

Reduce the thesis to a threshold you can be wrong about in public. He said $500 per kilo, and anyone can now check him.

Among the businesses the trend makes possible, choose the one touching the fewest unsolved problems. No re-entry was worth more than any technical advantage.

Buy a deadline early, while it is still cheap. The booked launch cost about 15 percent of the first round and reorganized everything that came after it.

What to do this week

1. Name the input cost your company is betting falls, and write the number it is at today. If you cannot name one, you are betting on execution alone, which is a worse bet than it sounds.
2. Write the break-even: the single number at which your business works. State it in a sentence a stranger could check. If you cannot, you do not yet have a thesis, you have a direction.
3. Run the calculation that would kill it. Johnston ran it on his own first idea and it came back $50 per kilo against a reality of thousands, so the idea went in the bin and the team stayed. Budget half a day.
4. Book one dated, external, cheap commitment in the next five working days. A design partner ship date, a published launch date, an accepted talk. Cheap in money, expensive in reputation, dated.
5. List the unsolved problems each option on your roadmap depends on, and pick the shortest list. Not the biggest market, the shortest list.
6. Write down the strongest filed objection to your plan, in its advocates' own words, not your paraphrase. If you cannot state it fairly, you cannot answer it in a diligence call.
7. If your plan requires a kind of engineer you are not, go recruit that person before you finalize the plan. Expect cold outreach to build the network and a warm introduction to produce the actual co-founder.

The course is the structured version of this kind of reasoning: AI Operating System for Startups covers how to pick the bet, instrument it, and tell early whether the trend line is actually moving your way. For the wider map of where AI changes a startup's product, engineering, pricing and team, start with AI for startups.

Sources

• The Case For Data Centers In Space, Y Combinator's The Lightcone with Starcloud co-founder and CEO Philip Johnston, published 2026-08-05. The interview this article distills.
• Starcloud's Philip Johnston: Why the Cheapest Compute Will Be in Space (Sequoia Capital, May 2026), for the cost comparison against a terrestrial solar project, the radiator math, the inference-only rationale and the debris answer.
• Starcloud's own Starcloud-1 mission page and team page, and How Starcloud Is Bringing Data Centers to Outer Space (Nvidia), for the first-satellite payload and the founding team.
• Funding: Reuters on the $170 million Series A at a $1.1 billion valuation (March 2026) and Reuters on the $250 million extension at $2.3 billion (August 2026), with SpaceNews on Nvidia joining that round.
• Starcloud files plans for 88,000-satellite constellation (SpaceNews), for the FCC application.
• The objections: Secure World Foundation's FCC filing, the American Astronomical Society's COMPASSE comment and the Center for Space Environmentalism's April 2026 comment.
• New York's permitting pause: Executive Order No. 62 and the Governor's announcement, with Reuters and AP on its scope.
• Texas: Governor Abbott's directive to the TCEQ halting data center environmental permits, with the Texas Tribune and CNBC on its scope. Maine's vetoed bill: TechCrunch.
• SpaceX files plans for a million-satellite orbital data center constellation (SpaceNews), for the January 2026 FCC application.
• 47 CFR 25.157, for how the FCC processes competing non-geostationary applications in rounds rather than in filing order.
• Starcloud's plan sees objections and criticisms (Communications Daily), for the NASA, Viasat and DarkSky International filings alongside the astronomers'.
• Nvidia Launches Space Computing, for the Space-1 Vera Rubin Module announced at GTC 2026 and its named partners.
• The competition: Blue Origin joins the orbital data center race (SpaceNews) on Project Sunrise, and Google's Project Suncatcher prototype launch with Reuters on the TPU test.
• Firefly Aerospace signs Starcloud as a commercial customer (September 2026), for the lunar-orbit mission.
• SpaceX Rideshare pricing for the published per-kilogram rates used as the reference point on launch cost.
• Pushback on the economics: Angadh Nanjangud's technical critique of the single-Starship cost case, and "Orbital Data Centers: Spacecraft Constraints and Economic Viability", for the thermal-mass and launch-count argument.
• Profiles: Philip Johnston, Ezra Feilden and Adi Oltean.