Florida Man Spots Data Centers In Space

On August 21 Starcloud announced a $250 million extension to its Series A at a $2.3 billion post-money valuation, more than doubling the $1.1 billion it set back in March and bringing total capital raised to $450 million since the company was founded in 2024. Manhattan West led it. Nvidia and Cisco Investments came in new, alongside Cedar Capital, Goanna Capital and Standard Capital, with Benchmark, EQT, Soma, NFX and 776 following on from the earlier round. Philip Johnston’s announcement post says the company moves into a 100,000 square foot facility this week and will ramp to a hundred satellites a week.

I read that post on a Friday afternoon, in the middle of trying to raise twenty million dollars for a company that has paying merchants.

So I ran the arithmetic. Not the vision, the mass budget: radiator area, failure rates, tonnes to orbit, who fixes it when it breaks.

Starcloud’s white paper, published back when they were still called Lumen Orbit, benchmarks a 40 megawatt orbital cluster against a terrestrial one. Ten-year terrestrial operating cost: $167 million, made up of $140 million in electricity at four cents a kilowatt-hour, $7 million in cooling, $20 million in backup power, and 1.7 million tons of water. Ten-year orbital cost: $8.2 million, made up of $2 million in solar arrays, $1.2 million in radiation shielding, and $5 million for one Starship launch.

One launch. That single assumption is carrying the entire company.

The paper puts $5 million against 100 tons of payload and calls the result approximately $30 per kilogram, which is arithmetic that actually produces $50, and I only mention it because a company asking for a $2.3 billion valuation on the strength of a launch-cost figure might want to check the launch-cost figure. The long-term vision is a 5 gigawatt cluster with solar arrays four kilometers on a side, scaling out to 88,000 satellites and 20 gigawatts of orbital compute.

Nvidia’s own blog post about Starcloud calls the vacuum of deep space an infinite heat sink. That gets the physics exactly backwards, and it is not a small thing to get backwards.

A vacuum flask keeps coffee hot for six hours using nothing but a vacuum. That’s what a vacuum is for. It removes conduction and convection, which happen to be the two mechanisms every data center on Earth uses to get heat out of a rack. On the ground you push air across a heatsink or pump water through a cold plate and dump it into a tower. In orbit both channels are gone and you’re left with thermal radiation, which is slow, and which is why spacecraft thermal engineering is a specialty and data center thermal engineering mostly isn’t.

Stefan-Boltzmann gives you the ceiling. A two-sided black plate at 20 degrees Celsius radiates roughly 770 watts per square meter, and once you subtract what it absorbs back from direct sunlight and from Earth’s albedo and infrared you’re netting about 633. Rejecting 40 megawatts of waste heat, and essentially every watt a data center draws turns into waste heat, needs about 63,000 square meters of radiator.

Area turns out to be the cheap part. Space radiators aren’t sheets of foil, they’re pressurized fluid networks that have to carry heat from the source out to every square centimeter of panel, which means ammonia loops, manifolds, internal channels, pumps, mounting hardware. The ISS External Active Thermal Control System radiators mass 1,122 kilograms each across 79.2 square meters, or 14.16 kilograms per square meter, and the photovoltaic radiators are heavier still at 17.48. The radiating surface is light. The plumbing is what kills you.

Run flight-proven ISS mass density across 63,000 square meters and the radiators alone land between 895 and 1,103 tonnes, which is nine to twelve Starship launches for the cooling system before a single GPU goes up.

Angadh Nanjangud, who has published peer-reviewed work on in-space assembly of large telescopes, worked this through in detail last year, and his sourcing checks out where I’ve followed it back: the ISS radiator masses and the iROSA figures are straight from NASA documentation.

Servers come to 408 tonnes. Starcloud’s own benchmark is 300 Nvidia GB200 NVL72 racks at 1.36 tonnes and roughly 120 kilowatts each, and that 1.36 is the compute rack by itself. The full deployment also ships an 800 kilogram NVLink switch rack, a 400 kilogram coolant distribution unit and a 300 kilogram power distribution unit per system, none of which anybody has counted here.

Solar arrays come to 397 tonnes, using the ISS Roll-Out Solar Array at 3.10 kilograms per square meter across the 128,000 square meters that Starcloud’s own power density numbers demand. This is the one line they can fairly contest. The white paper calls for thin-film cells thinner than 25 microns at better than 1,000 watts per kilogram, which is a real technology and would cut the solar mass substantially below the iROSA benchmark. It would not touch the radiators, and the radiators are the problem. Radiators come to 895 tonnes taking the optimistic end.

That’s 1,687 tonnes against a white paper budgeting 100, or 167 on its own corrected math. Seventeen to twenty-two Starship launches for what the paper prices as one.

And the 1,687 excludes the structural truss holding a 357-meter-square array rigid, the propellant for station-keeping against drag, micrometeoroid and debris shielding, the assembly robotics, and the optical comms hardware. All additive. Treat it as a floor.

Divided out: 42.2 tonnes per megawatt.

Falcon 9 is the only vehicle flying commercial payloads at scale right now. SpaceX raised the list price to $74 million a launch in February. Against the vehicle’s maximum 22,800 kilograms to low Earth orbit that works out to $3,246 per kilogram, and if you want the booster back the usable payload drops to 17,500 kilograms and the price climbs to $4,229. I’ll use the lower figure throughout, which assumes Starcloud fills every rocket to its expendable limit and never wastes a gram. You will see the $2,720 number quoted around the industry. That one comes from the old $62 million list price and it is stale.

Forty-two point two tonnes per megawatt at $3,246 per kilogram comes to $137 million per megawatt, and that is the shipping line only. No hardware, no assembly, no operations.

A complete AI-optimized data center on the ground, shell and power distribution and liquid cooling fit-out and the GPUs themselves, runs $30 to $40 million per megawatt in 2026. JLL puts standard shell-and-core at $11.3 million per megawatt globally, and Epoch AI models a full gigawatt build at $38 billion in upfront capital. So the freight on the orbital version costs roughly four times what the entire finished terrestrial facility costs, GPUs included.

Scale it. Five gigawatts is 211,000 tonnes, 12,000 Falcon 9 flights, $684 billion. Twenty gigawatts is 843,000 tonnes, 48,000 flights, $2.74 trillion. For their part, the white paper puts 5 gigawatts at fewer than 100 launches for the compute and a similar number again for the solar and radiator modules, so call it 200. We disagree by a factor of ten, and the disagreement is entirely in the mass budget.

Every nation and company on Earth has flown 7,386 orbital launches since Sputnik in 1957. Starcloud’s stated constellation needs six and a half times all of human spaceflight. SpaceX’s record year was 165 launches, so at that cadence you finish in 292 years, and the first ones would be long dead before the last ones flew.

To their credit the white paper does not ignore this. There is a section called Maintenance, and it describes a real architecture: compute containers dock and undock independently, faulty ones get swapped out, the old ones either ride home in the launcher’s payload bay or are built to burn up completely on reentry, and redundancy is designed in at system level so performance degrades gracefully instead of falling over. That is a sensible design and I would sign off on the drawing.

It appears in their cost table as a zero.

When Meta trained Llama 3 405B on 16,384 H100s, the cluster took 419 unexpected hardware interruptions across 54 days. One every three hours. A hundred and forty-eight of those were outright GPU faults, seventy-two were HBM3 memory failures, nineteen were GPU SRAM, seventeen were the GPU system processor, thirty-five were network switches and cables. Two were CPUs. The GPU is the fragile component, which anyone who has run a dense cluster already knew, and single-GPU mean time between failures sits around 50,000 hours.

Annualize Meta’s numbers and you get roughly one GPU in ten dying per year, which is conservative against that MTBF figure. On the ground this is a solved problem and barely even an incident: detection in minutes, a technician pulls the tray and swaps the card in two to four hours, validation another hour or two, back in production the same day. Maintenance runs something like forty percent of data center operating cost, and a 30 megawatt facility spends on the order of $100 million a year all in. It is a line item. You staff it, you stock spares, you move on.

In orbit the line item exists, but every entry on it is a rocket launch.

Starcloud’s 40 megawatt benchmark is 21,600 GPUs. At one in ten a year that’s roughly 2,200 dead GPUs annually that nobody will ever touch again. Scale to the 5 gigawatt vision and you’re looking at three million GPUs and something like 300,000 permanent failures a year. Not offline. Not queued for RMA. Dead, in a sealed box, moving at 28,000 kilometers an hour.

And that’s before you count the things that fail around the GPUs. On July 31, 2010, a single 350 kilogram ammonia pump module failed on the ISS and took out half the station’s cooling. Fixing it took three spacewalks, plus a fourth later just to stow the dead unit. Fifteen and a half hours of EVA across the first two alone. An M3 quick-disconnect refused to unlatch and Doug Wheelock had to pry it with a lever, at which point it started venting ammonia, and somewhere in the middle of all this the CO2 sensor in his suit failed. The station ran degraded for two weeks. All of that happened with a crew of six on board, an airlock, a robotic arm to ride, a spare pump already parked outside, and two of the most highly trained human beings alive on the end of the tether.

Starcloud’s satellites have no crew, no airlock, no arm, no spare and no astronaut. When a pump seizes, the satellite is finished.

The only hardware ever designed to be serviced in orbit was Hubble, and NASA flew five shuttle missions between 1993 and 2009 to do it. The 1993 repair was budgeted at $251 million for hardware and prep, but shuttle flights ran around $1.5 billion apiece in 2011 dollars, and NASA priced the final 2008 servicing mission at roughly $900 million cradle to grave. Hubble’s lifecycle cost including servicing came to something like $16 billion in then-year dollars. That is the actual historical price of changing parts in space, for one telescope, with a national space program behind it.

Commercial servicing does exist now, barely. Northrop Grumman’s SpaceLogistics flew MEV-1 to Intelsat 901 in February 2020 and gave it five years of life extension before undocking in April 2025. Except MEV doesn’t repair anything. It clamps onto the client’s apogee engine and takes over propulsion and attitude control, which is a tow truck, not a mechanic. Two MEVs have serviced three satellites and delivered about ten combined years of extension. The Mission Robotic Vehicle, the first one with actual arms, launched this month. Six years of commercial in-orbit servicing history, three customers, zero component repairs.

Starcloud is proposing 88,000 satellites.

So price their own architecture. A dead container gets replaced, which means every maintenance event is a launch event. At one GPU in ten per year, holding 40 megawatts steady means relaunching about a tenth of the compute mass annually, call it 41 tonnes a year against the 408 tonnes of servers. Across the ten-year window their own table uses, that is another 408 tonnes of replacement containers, or $1.3 billion in freight on top of the $5.5 billion it cost to put the thing up in the first place.

Call it $6.8 billion over ten years, against the $8.2 million in Table 1. Off by a factor of eight hundred.

The alternative is to skip resupply and overprovision instead, and that is worse. Ten percent compounding against the fifteen-year design life they claim leaves twenty-one percent of the GPUs alive at the end and averages under half of nameplate across the life. Sizing the launch so that 40 megawatts is still breathing in year fifteen means putting up 194 megawatts of nameplate, which is 8,200 tonnes and $27 billion of freight, to operate a cluster the size of one mid-sized terrestrial hall.

Either way the maintenance budget is denominated in kilograms and settled at $3,246 a kilo. There is no version of this where you pay it in technicians.

The obvious rebuttal is that you send people up to build and fix it. Or robots. Or some mix of the two. The industry has actually tried all three, so the receipts exist and we can price them.

People first. A seat to low Earth orbit runs about $55 million on Crew Dragon, which is what Axiom’s private astronauts paid. On top of the seat, NASA’s published price list charges $11,250 per person per day for life support and the toilet, another $22,500 per person per day for food, air and consumables, $42 a kilowatt-hour for power, and up to $164,000 per person per day just to launch the supplies that person will consume. There’s $5.2 million per mission for station crew time and $4.8 million for mission planning and space-to-ground comms. Those figures come from two different NASA pricing frameworks that overlap on consumables, so don’t just add them up. Take the seats alone and four people is $220 million. Add the daily charges at either framework’s rate and a ninety-day rotation lands somewhere between $250 and $310 million before anybody opens a toolbox. And that pricing assumes the ISS is sitting there as your habitat. Starcloud has no habitat. No airlock, no galley, no lifeboat, no way home. All of that is additional mass on top of the 42.2 tonnes per megawatt, and every kilogram of it prices at $3,246.

Now look at how much work a human actually gets done out there. A spacewalk runs six to seven and a half hours, and between prebreathe protocols, decompression risk, suit consumables and plain hand fatigue inside a pressurized glove, a crew tops out around two EVAs a week. Assembling the ISS took more than 160 spacewalks across 42 launches and thirteen years to put 420 tonnes together. Round it to one spacewalk per two and a half tonnes.

Starcloud’s 40 megawatt benchmark is 1,687 tonnes, which is four times the mass of the International Space Station. At ISS assembly intensity that’s on the order of 640 spacewalks, or roughly six years of a permanent crew doing nothing but suiting up twice a week, to stand up the smallest system in the deck. The 5 gigawatt version is 211,000 tonnes. Five hundred space stations. I’m not going to insult anyone by writing down that EVA count.

And the ISS is the only completed example of large-scale orbital assembly humans have ever produced, which makes it the only real cost benchmark we have: $150 billion for 420 tonnes, or about $357 million per tonne assembled. Nobody thinks Starcloud would pay ISS prices, and I’m not claiming they would. But nobody has demonstrated any other number either, and a company valued at $2.3 billion on an assembly-dependent architecture ought to be able to name the number it is claiming instead.

So robots, then. This is where the honest answer gets uncomfortable, because NASA already ran that experiment and shut it down. OSAM-1 was a single robotic servicer, built to do one refueling of Landsat 7 and assemble one Ka-band antenna in orbit. It started life around 2016 as Restore-L with a projected cost between $626 and $753 million. By the October 2023 Inspector General report it had passed $2 billion. NASA cancelled it in March 2024. Congress didn’t like that and appropriated $227 million to force a descoped 2026 launch, NASA reviewed the descoped plan, and in September 2024 reconfirmed the cancellation anyway, citing significant risk to mission success and low return on investment for the servicing community. Roughly $2 billion and a decade for one robot that never flew.

Canadarm2 is the counterexample, and it’s a real one. It works, it’s brilliant, it pulled most of the ISS modules out of the shuttle bay and bolted them together, and it cost about $2 billion for a single arm. It has also needed maintenance itself: a wrist joint installed in 2001 is being replaced this year, by astronauts, which tells you something about the mixed-team model. The robot needs the crew. Northrop Grumman’s Mission Robotic Vehicle, the first commercial servicer with actual manipulator arms rather than a docking clamp, launched this month with zero flight heritage behind it.

Mixed teams are not a third option that dodges the first two. Mixed teams are what the ISS was: Canadarm2 plus Dextre plus 160-odd spacewalks plus 42 launches plus thirteen years plus $150 billion, for 420 tonnes in a 109-meter span. That is the state of the art in orbital assembly, it is twenty-five years old, and Starcloud needs to beat it by three orders of magnitude on cost while also beating it on schedule while also doing it uncrewed.

There is a way out of this, and it’s worth naming because it’s probably what Starcloud actually intends. Don’t assemble a structure. Fly a constellation. Build 88,000 small self-deploying spacecraft that each unfurl their own arrays and radiators the way a Starlink does, network them with optical links, and never send a human or a robot anywhere. That architecture genuinely sidesteps the assembly problem.

It also throws away most of the economics. Every one of those 88,000 spacecraft needs its own bus, its own propulsion, its own station-keeping propellant, its own thermal loop and its own comms, and small thermal systems are dramatically less mass-efficient than large ones because the plumbing overhead doesn’t scale down. You need them holding formation tightly enough to keep optical links locked while atmospheric drag pulls on each one differently. You need seventeen years of uninterrupted production at a hundred a week. And you need to convince the FCC, the ITU and every other operator in low Earth orbit that 88,000 additional spacecraft is a responsible thing to do in the same shell that already holds the largest constellation ever flown.

Meanwhile Starcloud’s own concept art shows a single coherent installation with a solar wing four kilometers on a side. That is not a constellation. That is a structure, and it is forty times longer than anything humanity has ever assembled in orbit. The public materials describe both architectures, which suggests to me that the question is still open internally, and it is not a small question. It’s the whole question.

Everything above assumes Falcon 9, and Starcloud would fairly say that isn’t the plan. Fine, take Starship at its word.

At 100 tonnes to LEO and the paper’s assumed $5 million a flight, the 40 megawatt cluster costs $85 million to launch instead of $5 million. Five gigawatts costs $10.5 billion across 2,108 flights, twenty gigawatts costs $42 billion across 8,433 flights. Price a Starship flight at what serious people in the industry actually model, call it $100 million or $1,000 per kilogram, and 5 gigawatts costs $211 billion in freight alone.

For shipping to be a normal fraction of project cost instead of the dominant one, Starcloud needs launch around $170 per kilogram. SpaceX’s 2026 prospectus targets a 99 percent reduction against a historical benchmark of $18,500 per kilogram, which works out to $185. So the business case doesn’t just need Starship to work, it needs SpaceX to hit the most aggressive number in its own investor materials and then hand the entire saving to a customer at cost.

That isn’t how SpaceX prices anything. Commercial customers don’t pay SpaceX’s marginal cost, they pay what SpaceX charges, and what SpaceX charges is set by the next-cheapest alternative, which today is a Falcon 9 at $3,246 per kilogram. Starcloud will get quoted a price, not a cost.

Meanwhile Starship has flown thirteen integrated test flights as of mid-2026 and has never sold a commercial payload delivery. Flight 13 in July put up 20 Starlink V3 satellites, the first time it carried real hardware instead of mass simulators. The schedule has slipped three times this year. Johnston told TechCrunch he’s raising specifically because the launch market is tightening and he can see how much he’ll need to book, which is a completely correct read of his own situation: he is buying a place in line for a vehicle whose price has to fall nineteenfold from anything anyone has ever paid, operated by a company that Nvidia is separately funding to pursue orbital compute of its own.

The power system and the network want opposite things, and this is the conflict that never appears on the slide.

A data center is only worth anything if you can reach it. In orbit that means having a spacecraft above your horizon with a clear line of sight to a ground station. At 550 kilometers a satellite clears a fixed ground station for eight to twelve minutes above a five-degree elevation mask, and a given satellite-and-station pair typically gets four to six of those windows a day. That works out to roughly three percent of the day in contact. The other ninety-seven percent it’s over the Pacific, or over Kazakhstan, or over the night side of the planet with nobody to talk to.

Now layer the power requirement on top. This isn’t my inference, it’s their stated reasoning: the white paper says the most important factor in choosing an orbit is continuous solar generation, and that dawn-dusk sun-synchronous orbit was selected on that basis. It is a near-polar track that rides the terminator in permanent sunlight, and it is the only low-Earth orbit with that property. That orbit is locked to the sun. It is not locked to your customer. From a ground station at mid-latitude the spacecraft turns up at about the same two moments every day, near local dawn and near local dusk, and it is somewhere else the rest of the time. The orbit that makes the electricity free is the orbit that puts you over Virginia twice a day.

Now, the fair rebuttal, because there is one. A constellation fixes coverage. That’s exactly what Starlink does: any given satellite is only overhead briefly, but there is always some satellite overhead, and traffic gets handed off. Fine. Except that solves the wrong half of the problem. Coverage means there is always a spacecraft over your ground station. It does not mean the spacecraft holding your model is the one over your ground station. Almost by definition it isn’t. So every byte in and out of your compute has to cross the constellation first, satellite to satellite, until it reaches whichever unit happens to have a clear shot at a dish.

Which makes the relay fabric the entire product, and the relay fabric is where the announced numbers fall apart. Starcloud’s May contract with Starlink covers more than 50 Starlink Mini Laser terminals across at least 25 satellites, giving the constellation up to 25 gigabits per second of intersatellite optical capacity.

A single GB200 GPU has an 800 gigabit per second scale-out port. One GPU’s network connection carries thirty-two times the bandwidth of the entire announced constellation link budget. Inside the rack it’s worse by a margin that stops being comparable: the NVL72’s NVLink fabric moves 130 terabytes per second across its 72 GPUs, which is 1,040,000 gigabits per second, or about forty-one thousand times the whole constellation.

That gap is not an engineering detail to be closed later, it is the reason the rack exists. NVIDIA built the NVL72 as a single 72-GPU NVLink domain with 13.8 terabytes of unified coherent memory precisely so that a trillion-parameter model doesn’t have to cross a network at all. NVIDIA’s own deployment guidance tells operators to keep jobs inside one rack wherever possible and, when they must scale out, to use rail-optimized topologies to blunt the bandwidth cliff from NVLink down to InfiniBand. That cliff is about eighteen to one. The cliff from InfiniBand down to a Starlink laser link is another two thousand to one.

So you cannot split a training job across spacecraft. You cannot migrate a workload to whichever satellite happens to be over Oregon, because the weights are terabytes and the pipe is 25 gigabits. What you have, if you fly the constellation architecture, is not a 5 gigawatt data center. It is roughly 41,000 isolated single-rack islands, each capable of running whatever fits in 13.8 terabytes and nothing larger, loosely federated over a link slower than a decent office building’s uplink. That is an inference edge network, which is a real product and a decent business, and it is not the product being valued at $2.3 billion.

Downlink to the ground has the same shape of problem. The state of the art is NASA’s TBIRD, which pushed 200 gigabits per second from a 6U CubeSat at 530 kilometers and moved more than a terabyte in a single sub-five-minute pass, over two years and 110 passes. Genuinely superb work. It also needed two dedicated optical ground stations, Table Mountain in California and Haleakala in Hawaii, both sited specifically for minimal cloud cover, both running adaptive optics on meter-class telescopes, because an optical downlink is stopped dead by a cloud. Your gigawatt data center’s availability is now a function of the weather over a mountain in Maui.

There’s a line in the white paper, in the passage reassuring astronomers that orbital data centers won’t ruin the night sky, noting that the satellites will only be visible at dawn and dusk. It is meant as a comfort. Read it as an operations statement instead and it tells you exactly when your data center is over your head.

And this is the point where the orbit stops being an abstraction. The thing does come back. Everything in low Earth orbit does, on a schedule set by atmospheric drag rather than by anyone’s business plan.

I want to be careful here, because the easy version of this argument is wrong and I nearly made it. Starlink designs to a five-year life because it flies at 550 kilometers where the atmosphere still bites, and the FCC’s five-year rule governs post-mission disposal, not service life. Starcloud has chosen dawn-dusk sun-synchronous orbit, which sits higher, where drag is a much smaller tax. They claim a fifteen-year design life and benchmark it against the ISS cooling and power subsystems. Fine. Take them at their word.

Fifteen years is worse for them than five would have been.

Attrition compounds the whole time. Ten percent a year leaves fifty-nine percent of the GPUs alive at year five and twenty-one percent at year fifteen, so the longer the design life, the emptier the box gets before anyone retires it. A shorter life would at least let you replace the fleet with current silicon.

Obsolescence is the sharper edge. Take the industry rule of thumb that GPU performance per watt roughly doubles every couple of years, and an H100 launched in 2025 is facing something like a fourfold ground-side advantage on the same power envelope by 2029, and there is no aisle to walk down and no tray to pull. You deorbit the satellite. Terrestrial operators upgrade in place, in a building that doesn’t fall out of the sky.

On which: the reentry problem doesn’t go away with a longer life, it just spreads out. Eighty-eight thousand satellites on a fifteen-year cycle is 5,900 reentries a year, sixteen a day, every day, forever, and roughly 56,000 tonnes of hardware vaporizing in the upper atmosphere annually. The FCC requires 95 percent post-mission disposal reliability and SpaceX reports better than 99 percent for Starlink. At Starlink’s best-in-class number that is still about sixty large uncontrolled reentries a year from spacecraft nobody can reach to fix. At the FCC’s floor it is closer to three hundred. NASA is paying up to $843 million for a single-use tug whose only job is making sure one 420-tonne structure comes down in the South Pacific rather than somewhere with people, and Starcloud’s smallest benchmark is four times that mass.

And the silicon was never built for any of this. The H100 is a 4-nanometer part with 80 billion transistors and no radiation hardening at all, no protection against total ionizing dose, none against single-event latch-up. No rad-hard 4-nanometer GPU exists, because hardening means larger geometries and larger geometries destroy the performance you launched it for. Starcloud’s answer is software-level error correction and redundancy. Starcloud-1 did run past thirty days without a radiation-induced crash, which is a real result and I’m not going to wave it off, but thirty days is not fifteen years and one GPU is not 88,000 satellites.

On which: a hundred satellites a week means seventeen years of continuous uninterrupted production to build 88,000 of them. The first ones would be burning up on reentry before the last ones came off the line.

Starcloud took $3 million of pre-seed money and put a working H100 in orbit in twenty-one months. It ran inference on Gemma and trained a small model from scratch. That is genuinely excellent execution, and I’ve watched plenty of companies with fifty times the capital ship nothing at all.

The near-term product is a real business too. Processing Earth observation data in orbit instead of downlinking terabytes of raw imagery through a starved pipe solves an actual customer problem, and Starcloud-2 is aimed straight at it. The constraints they’re pointing at are also real: interconnection queues running past 2030, communities fighting gigawatt campuses, power rather than silicon as the binding constraint on terrestrial buildout.

None of that makes a 20 gigawatt constellation a business. It makes it a compelling slide about a real problem. The satellite edge-compute company underneath is fundable and probably good, and it isn’t what $2.3 billion buys. The valuation is being paid for the gigawatt story, and the gigawatt story runs on launch economics and maintenance assumptions that don’t exist.

Nvidia joined a round whose stated purpose, in Johnston’s own words, is building the infrastructure to launch many more of Nvidia’s most advanced GPUs into space. Nvidia is separately developing the Vera Rubin Space-1, its first purpose-built GPU for orbit, and separately backing SpaceX. None of that is a return-seeking financial position. It’s demand manufacturing on the balance sheet, the same circular vendor-financing loop we’ve been watching run through the neoclouds for two years now. Cisco Investments is buying a seat at the optical and networking table. Manhattan West is buying narrative exposure to the hottest constraint story in technology.

Nobody in that syndicate is underwriting free cash flow, and nobody in it needs to be right. Venture capital doesn’t underwrite expected value, it underwrites variance, and a fund needs one position that returns the whole fund. The only candidates are the ones with an unbounded ceiling if the physics happens to cooperate. A 96 percent chance of zero against a 4 percent chance of owning orbital compute is a perfectly rational bet for a fund and a completely irrational one for anybody actually measuring whether the thing works.

I’ve spent nearly thirty years being the guy who has to answer what something actually costs in production. At Volusion I ran infrastructure under forty thousand live storefronts, where capacity planning that was optimistic by twenty percent meant merchants watching checkout fail on a Saturday afternoon and me on the phone about it. In payments, years of PCI-scoped systems and tokenized vaults, you don’t get to hand-wave any of the arithmetic, because the arithmetic is money and the money belongs to someone else. I’ve been handed a lot of decks. The ones that scared me were never the ones with hard numbers I disagreed with, they were the ones where the hardest number in the deck was a price the company didn’t control, sitting on slide nineteen, load-bearing for everything in front of it.

ShopSnap charges merchants zero transaction fees. It has a working agentic shopping assistant with semantic routing between cheap and premium models so the unit economics survive contact with real traffic. It runs on infrastructure I built, in production, for customers who pay. Twenty million dollars takes it national.

And I know exactly why the money isn’t there, which somehow makes it worse rather than better. AI companies took roughly 80 percent of all global venture funding in the first quarter of 2026, and by July that was near 86 percent of US venture dollars. OpenAI and Anthropic between them accounted for 43 percent of everything raised in the first half. Non-AI startups got around $58 billion in Q1, a number that would have led every quarter before 2018 and which, inflation-adjusted, sits below where it was in Q1 2020. Fifteen percent of seed-funded companies from the 2022 and 2023 cohorts made it to a Series A within two years, down from better than thirty percent for 2018 through 2020.

So the market isn’t mad. It’s allocating exactly the way a power-law business allocates once a single narrative has swallowed the ceiling. A company selling better unit economics to merchants is a good business with a bounded story, and bounded stories don’t return funds. A company selling orbital compute is a bad business with an unbounded story, and unbounded stories do. That is the whole reason a founder with revenue takes twenty meetings and a founder with a thermodynamics problem takes $250 million.

I’m not going to pretend that’s fine. I’ll keep shipping and I’ll keep doing the arithmetic, because sooner or later somebody has to hold the invoice for 843,000 tonnes and name the person going up there to change the pump.

About the Author

Gal Ratner is the founder and CTO of Inverted Software and WhiteStar Labs, and Chief Architect at Prana Entertainment in Las Vegas. He has spent nearly thirty years building production systems on the Microsoft and .NET stack for clients including Microsoft, Sony, Rockstar Games, 2K Games, Best Buy and Allegiant Air. He was employee number six at Break.com, ran infrastructure under forty thousand storefronts at Volusion, and was named a Los Angeles Business Journal CTO of the Year finalist. He currently operates ShopSnap, a zero-transaction-fee commerce platform where he is building agent-native merchant infrastructure, and works on production agentic AI: MCP servers, RAG pipelines, SQL Server 2025 vector search, and the PLogger observability framework. He trains Brazilian jiu-jitsu under Sergio Penha, rides motorcycles, co-hosts Edge Grip Podcast, and writes at galratner.substack.com.