The single most persistent misunderstanding about orbital compute is that space is cold, and therefore cooling is easy.
Space is cold and it does not cool anything, because cooling requires a medium to carry heat away and vacuum has none. There is no convection. There is no conduction to anything beyond the spacecraft itself. The only mechanism available is thermal radiation, which is governed by the Stefan-Boltzmann law and which requires surface area in quantities that terrestrial engineers never have to think about.
NVIDIA’s chief executive put it plainly: it is cold in space, and there is no airflow. A space station operator put it more bluntly, calling it counterintuitive that cooling in space is hard precisely because there is no medium to transmit hot to cold.
Which produces the finding that should govern how anyone reads this subject. Space data centers relieve exactly one of the three constraints that have organized everything else, and they make the most fundamental one substantially worse. Every serious space data centers proposal is therefore a trade rather than an escape, and the question is only whether the trade is favourable.
What space data centers actually solve
The case is real and deserves stating at full strength before the objections, because the advantages are genuine and specific.
Power is the big one. In a dawn-dusk sun-synchronous orbit a satellite sits near the terminator and receives nearly continuous solar illumination, with no night, no weather, and no atmospheric attenuation. Solar panels in that configuration produce substantially more energy per unit area than the same panels on the ground, with estimates running as high as eight times terrestrial output depending on the comparison, and the power is carbon-free without a fuel supply chain, a combustion permit, or a grid interconnection queue. It also requires no fuel cycle, no enrichment capacity, and no reactor licensing, which removes an entire category of multi-year dependency.
Water use goes to zero, because there is no evaporation and nothing to evaporate.
Land use goes to zero, and with it the entire apparatus of county boards, zoning hearings, ballot measures, and tax abatement negotiations that has become the binding political constraint on terrestrial siting.
Transformer lead times, turbine backlogs, and electrician shortages become irrelevant, because none of that equipment exists in orbit. A satellite carries its own generation and distribution, which sidesteps the three-to-five-year transformer queues and multi-year turbine backlogs that currently determine which terrestrial announcements become buildings.
That is a serious list. Three of the four are the exact constraints that have made terrestrial buildout difficult, and orbit removes them entirely rather than mitigating them.
Worth naming the political point precisely, because it is the one operators discuss least publicly and value most. A terrestrial project can be stopped by a county commission, a ballot measure, a water permit, a rate case, or an air permit, and increasingly is. A satellite constellation is licensed by a federal regulator and an international spectrum body, and no locality has standing. That is worth weighing against the cost-allocation fights now consuming utility commissions, since a satellite has no rate case and no ratepayers to shift costs onto. The entire apparatus of local objection that has become the binding constraint on siting simply does not apply, which is a structural advantage independent of any physics.
The heat problem, quantified
Then the physics arrives. Radiative heat rejection scales with the fourth power of absolute temperature and linearly with area, and the numbers that produces are not intuitive.
At around one hundred and twenty-seven degrees Celsius, roughly the practical upper limit for electronics, a radiator surface rejects approximately 1,450 watts per square meter. That is the theoretical best case, before accounting for view factors, radiator efficiency, the temperature drop between the chip and the radiator surface, and the fact that a radiator facing the sun or the illuminated Earth absorbs heat rather than rejecting it.
Work the arithmetic. A one-megawatt cluster requires something in the range of three thousand to ten thousand square meters of radiator area, depending on operating temperature and configuration. A gigawatt-scale facility, which is the unit of ambition in this industry, needs radiator area measured in square kilometers.
Every square meter of that has to be manufactured, folded into a fairing, launched, deployed reliably in orbit, and then survive micrometeoroid impacts and thermal cycling for years without a maintenance visit. Radiator mass and area do not scale gracefully, which is the central engineering fact about space data centers, and at megawatt scale they plausibly dominate the entire spacecraft.
The scale distinction matters enormously and gets collapsed. At ten to five hundred watts per node, the thermal problem is largely solved with flight-proven technology carrying decades of heritage in low Earth orbit, and thousands of satellites already reject heat at that scale routinely. At a megawatt, the radiator becomes the spacecraft. At a gigawatt, the structure being described is a megastructure rather than a satellite, and no comparable object has ever been assembled in orbit. Collapsing a fifty-watt edge node and a gigawatt facility into one conversation about space data centers is how the difficulty gets hidden.
There is a genuine mitigating trend and it deserves credit. Data center accelerators have become dramatically more tolerant of warm coolant. A 2014-generation GPU required coolant below fifteen degrees Celsius to avoid throttling. Current-generation platforms run efficiently at forty-five degrees. Because radiative rejection scales with the fourth power of temperature, a higher operating temperature is worth far more than the linear improvement it looks like, and that trend does real work for orbital feasibility.
It does not eliminate the problem. It moves the radiator area required for a given load down by a meaningful factor while leaving the scaling relationship exactly where it was.
Where the demonstrations actually are
The distinction between what has flown and what has been announced is the whole analytical task, and it is unusually clean in this case because both are documented.
What has flown: in November 2025 Starcloud placed a roughly sixty-kilogram satellite carrying an unmodified NVIDIA H100 into low Earth orbit at around three hundred and fifty kilometers, and trained a small language model on it. That is the first state-of-the-art data center GPU to operate in space and it is a genuine milestone. Axiom Space deployed orbital data center nodes in January 2026 with optical links in the low gigabits per second, which is real hardware in a real orbit and is bandwidth roughly four orders of magnitude below what a rack backplane moves internally.
What is scheduled: Starcloud-2 in October 2026, carrying several H100s alongside Blackwell hardware, with plans to deploy AWS Outposts hardware in orbit. Google’s Project Suncatcher prototype, two satellites built with Planet Labs, targeting early 2027 to test Trillium TPUs, optical inter-satellite links, and thermal management, with Google in launch services discussions with SpaceX as of May 2026.
What has been filed: SpaceX applications for up to one million data center satellites. Starcloud’s February 2026 filing for an 88,000-satellite constellation totaling roughly twenty gigawatts, with a stated vision of a five-gigawatt orbital hypercluster powered by a solar array spanning four square kilometers. A five-month-old company filing in June 2026 for up to 100,000 satellites at around ten gigawatts.
The gap between those three categories is four to five orders of magnitude. One satellite with one GPU has flown. Filings describe constellations of a million. An FCC filing is a document with an author who wanted something, it costs comparatively little to submit, and it establishes a regulatory position rather than a capability.
Google’s ground-based radiation testing produced a genuinely useful result, confirming that its TPU v6e can withstand the radiation environment of a five-year low Earth orbit mission, and its laboratory optical link demonstrations reached 1.6 terabits per second on a single transceiver pair. Those are real technical de-riskings of specific subsystems, and they are not the same as an operating cluster.
Launch economics, and the number everything depends on
The entire orbital case rests on launch cost falling, and the current numbers are further from the projections than the coverage suggests.
As of 2026, a reused Falcon 9 delivers payload to low Earth orbit at roughly $2,700 to $3,100 per kilogram, which is a ninety to ninety-five percent reduction from the Space Shuttle era and a genuine achievement. Small payloads on rideshare missions run considerably higher, around six to seven thousand dollars per kilogram. The industrial supply chain underneath launch is also not infinitely elastic, and launch cadence has its own regulatory ceiling that analysts expect to bind through 2028 regardless of vehicle capability.
Starship is the vehicle every orbital data center business case assumes. As of mid-2026 it had completed a series of test flights, with a record of roughly seven successes across twelve to thirteen attempts, and had deployed functional satellites on a suborbital trajectory rather than into a stable orbit. It has not reached stable orbit on any flight and has not begun selling launches to outside customers.
Analyst estimates put current Starship flight costs in the range of eighty to one hundred million dollars, which against a hundred-tonne payload implies roughly eight hundred to a thousand dollars per kilogram, an order of magnitude above the target figure.
The famous sub-hundred-dollar number traces to a 2019 projection of eventual operating cost, made years before the current vehicle existed. SpaceX’s own 2026 prospectus is more restrained, stating an aim to reduce the cost of reaching orbit by ninety-nine percent or more against a historical benchmark of $18,500 per kilogram, which computes to $185 per kilogram with room below.
Every optimistic figure rests on the same conditions: both stages returning and reflying with minimal refurbishment, and a flight cadence high enough to spread pad, factory, workforce, and development costs across many launches. Neither has been demonstrated.
Which means the honest framing is that orbital data centers are a bet on a launch vehicle achieving an operating profile it has not yet achieved, priced against a cost per kilogram that is currently a projection.
Radiation, and the hardware nobody designed for this
Commercial accelerators are not built for orbit and the environment does two distinct things to them.
Single-event effects occur when an energetic particle strikes the silicon and flips a bit or induces a transient fault. These are survivable with error correction, redundancy, and checkpointing, all of which cost performance and memory overhead, and which interact badly with memory-bandwidth-bound inference since error-correcting overhead consumes exactly the resource that is already scarce.
Total ionizing dose is the cumulative one and it is what limits mission life, and it is the constraint that makes commercial silicon awkward, since chips optimized for terrestrial density carry no radiation margin by design. Radiation gradually degrades semiconductor performance, and the degradation is not repairable in place. Google’s testing establishing TPU v6e tolerance across a five-year mission is the most useful public data point available, and it is a five-year number.
That interacts badly with the refresh cadence. NVIDIA moved to an annual product cycle. A five-year radiation-limited service life against a one-year hardware generation means an orbital facility is running increasingly obsolete silicon for most of its operating life, with no possibility of replacing individual components. Terrestrial operators treat component failure as routine and budget for it; the replacement supply chain for accelerators and cooling hardware is a standing operational cost rather than an emergency. In orbit there is no such supply chain, and proposals for in-situ manufacturing and orbital fabrication are considerably earlier in development than the compute platforms they would service. Terrestrial facilities swap failed accelerators, drives, and power supplies continuously. In orbit, a failed unit is dead capacity until the entire satellite is deorbited and replaced.
The distributed architecture is the answer to this, and it is why Google’s approach uses clusters of many small satellites replaceable incrementally rather than large monolithic platforms. Distributed failure characteristics are better and the upgrade path exists. It also multiplies the number of objects requiring launch, tracking, and eventual disposal, and it caps the size of any single coherent compute domain, since a model that would occupy a seventy-two-GPU NVLink domain on the ground has to be split across satellites connected by optical links with vastly lower bandwidth than a copper backplane.
Latency, bandwidth, and what the workload has to look like
Low Earth orbit adds roughly twenty to forty milliseconds round trip, plus Doppler compensation and handover between satellites as they move relative to a ground station.
That rules out interactive inference, which is the highest-value and fastest-growing workload in the industry. It is acceptable for batch training and for processing data that originates in orbit, which is the genuinely defensible use case: Earth observation constellations already generate more imagery than they can downlink, and processing in place rather than transmitting raw data is a real argument with real economics behind it, and it is the version of orbital compute that would exist whether or not anybody had ever proposed a gigawatt constellation.
The bandwidth constraint runs the same direction. Optical inter-satellite links have improved dramatically and space-to-ground optical links remain weather-dependent, since clouds block them. Radio frequency downlink has spectrum limitations, and spectrum is allocated internationally through a coordination process with its own multi-year timeline, which puts it in the same category as every other permitting constraint the terrestrial buildout runs into. Getting a training corpus up and a trained model down is a substantial data movement problem, and the scale-across communication penalty that already constrains multi-site terrestrial training is considerably worse across a link that is intermittent and weather-limited.
So the workload profile that fits orbit is narrow: batch, latency-tolerant, and ideally operating on data already up there. That is a real market and it is not the market the gigawatt constellation filings describe.
Debris, slots, and the governance problem
A million-satellite constellation is a regulatory and orbital-mechanics proposition before it is an engineering one.
Orbital debris accumulation is the obvious concern, and the mechanism that worries people is cascading collision, where fragments from one collision raise the probability of the next. Deployment on the scale being filed for would change the population of tracked objects in low Earth orbit by orders of magnitude.
The regulatory apparatus was not designed for this. SpaceX has requested waiver of FCC milestone requirements that normally mandate half a constellation deployed within six years and full deployment within nine, which is an acknowledgment that the schedule implied by the filing is not achievable under existing rules.
Spectrum coordination, international frequency allocation through the ITU, and end-of-life disposal obligations all scale with constellation size, and none of them has been tested at the numbers being proposed.
There is also a jurisdictional question that mirrors the terrestrial one exactly. A facility in orbit is subject to the licensing state’s law, which makes orbital siting a jurisdiction and export control decision in the same way terrestrial siting has become one. China’s Three-Body Computing Constellation began launching in May 2025 as part of a planned multi-thousand-satellite program, which means the geopolitical vector arrived in orbit before the commercial one did. Orbital slots and spectrum are allocated on a first-come basis in practice, which turns a filing into a claim on a finite resource and explains why the applications describe constellation sizes nobody expects to build. The same behaviour appears wherever a scarce permitted position has option value.
What the serious people are actually claiming
Separating the operators’ claims from the analysts’ assessments clarifies where the disagreement sits.
Elon Musk has projected cost parity between orbital and terrestrial compute within two to three years. Jeff Bezos has suggested gigawatt-scale orbital data centers within ten to twenty years. Google frames Suncatcher explicitly as early research toward eventual in-space scaling rather than as a near-term product, which is the most careful public framing any operator has offered and is worth noting as a contrast to the confident timelines elsewhere in the sector.
Deutsche Bank puts cost parity well into the 2030s, which is a bank taking a position and should be read as one. Analysts covering the sector characterize orbital data centers as speculative near-term revenue, citing unproven economics, hardware aging, latency limits, and narrow use cases.
Notice that the spread is not about physics. Everybody agrees on the Stefan-Boltzmann law, the radiation environment, and the latency. The disagreement is entirely about the launch cost curve and the timeline, which means the argument is a financial one wearing a technical costume. That is worth registering because technical-sounding disputes with financial content resolve on financial evidence, and the evidence here is a flight-rate curve rather than a thermal calculation.
The capital has arrived regardless. Starcloud raised a $170 million Series A at a $1.1 billion valuation in March 2026 against roughly $200 million total raised. Another entrant reached a reported $2 billion valuation in late March 2026. Venture money is now underwriting the demonstrations, which means investors can get exposure to the outcome without funding the experiment themselves. That is the same structure as any speculative infrastructure buildout financed against a projected cost curve, and it resolves the same way: the demonstrations either hit their gates or the valuations reprice.
The mass budget, and why it decides everything
Everything above resolves into one number, and working it explicitly is the most useful thing anybody can do with this subject.
A satellite carrying compute has to launch its processors, its solar array, its radiators, its structure, its attitude control, its communications hardware, and its propulsion for station-keeping and disposal. Of those, the radiators and the solar array scale with power while the rest scale more slowly, which means at high power the thermal and generation hardware dominate the mass.
Take a rough case. A megawatt of orbital compute needs radiator area in the thousands of square meters. Deployable radiator panels for spacecraft run in the range of several kilograms per square meter depending on technology and durability requirements. That alone puts radiator mass in the tens of tonnes per megawatt before anything else is counted, and the solar array to generate the megawatt adds its own.
Now apply launch cost. At the current demonstrated Falcon 9 figure near three thousand dollars per kilogram, tens of tonnes per megawatt implies launch costs in the range of a hundred million dollars per megawatt of compute, against a terrestrial data hall that runs roughly ten million dollars per megawatt all-in including the building. At the aspirational hundred dollars per kilogram, the same mass costs a few million per megawatt, and the comparison inverts.
That single sensitivity is the entire orbital thesis. Everything else in the engineering is a detail relative to whether launch cost falls by a factor of thirty from a demonstrated figure to a projected one. Anyone modelling this should build the case as a function of dollars per kilogram and observe how little else matters, which is the same structure as the depreciation-schedule sensitivity that governs terrestrial GPU economics: one unobservable input determines whether the business exists.
The claims that do not hold up
An audit, because this subject generates more confident assertion per unit of demonstrated capability than anything else in the buildout.
Space is cold so cooling is free is the foundational error and it inverts the actual difficulty.
Orbital data centers are coming in two to three years describes filings and prototypes rather than operating capacity, and the operators making the claim have an interest in it being believed.
Starship will cost a hundred dollars per kilogram is a projection resting on operating conditions not yet demonstrated, and current analyst estimates put the figure roughly an order of magnitude higher.
A GPU has been operated in space so the technology is proven conflates a sixty-kilogram demonstration with a gigawatt facility, and the scaling problems are exactly the ones the demonstration did not test.
Orbital compute eliminates environmental impact ignores launch emissions, atmospheric effects of large constellation reentry, orbital debris, and the manufacturing footprint of the satellites themselves, including the critical minerals in solar cells and spacecraft structures that carry their own extraction and processing burden.
Latency does not matter because it is all batch workloads is true for the defensible use case and inconsistent with the revenue projections attached to the large constellation filings, which require serving general demand.
The FCC filings show the industry is committed shows that filings are cheap and regulatory position is valuable.
Space data centers will replace terrestrial ones is the strongest version of the claim and is not supported by anything demonstrated, since the workload profile that suits orbit is narrow and the workload profile driving the buildout is not.
Nothing in orbit will ever be economic is the mirror error, since orbital edge compute for space-originated data has a genuine near-term case and the demonstrations are real.
What space data centers are actually telling us
The reason this belongs at the end of the sequence is that it functions as a test of everything established earlier.
The terrestrial constraints are power, thermodynamics, industrial supply, and politics. Orbit removes politics entirely, removes the industrial supply constraint for electrical equipment, and improves power dramatically. It makes thermodynamics categorically harder, adds radiation, removes maintenance, adds latency, constrains bandwidth, and substitutes a launch cost curve that has not been demonstrated for a set of equipment lead times that have.
That is not a solution to the problem. It is a different allocation of the same problem, which is the pattern this entire subject keeps producing. Move cooling water off the site and it reappears at the power plant. Move generation behind the meter and the cost allocation reappears in a rate case. Move the load offshore and the domestic political fight resolves at the cost of the strategic argument. Move the whole facility to orbit and the heat rejection problem, which was the original reason the buildings got expensive, becomes the dominant engineering constraint of the entire enterprise.
That pattern has a name in every other capital-intensive industry, which is that constraints are conserved rather than eliminated. The materials sector spent a century discovering it: substituting one input for another moves the bottleneck rather than removing it, and the substitution is worth making only when the new bottleneck is genuinely cheaper to relieve than the old one. Whether orbit clears that bar depends entirely on the launch cost curve, which is why the mass budget is the whole argument.
The genuinely useful version of orbital compute is the narrow one nobody is filing hundred-thousand-satellite applications for: processing data that originates in space, at kilowatt to hundreds-of-kilowatt scale, alongside defence and Earth-observation applications where the strategic value of processing in place exceeds the cost premium, where the thermal problem is solved with flight-proven hardware that has decades of heritage and the latency does not matter because nothing is waiting.
That business is real, it is being built, and it is roughly six orders of magnitude smaller than the announcements. That is not an argument against it. It is an argument for reading the unit attached to any figure in this space, since a hundred-kilowatt orbital node and a five-gigawatt orbital hypercluster differ by a factor of fifty thousand and appear in the same articles.
Which is the note the ten-lecture briefing on how AI data centers work ends on, having run the physics, the money, and the politics in the order they bind. Every figure carries a unit, a stage, and a date, and space data centers are where that discipline gets its hardest test, because the gap between what has flown and what has been filed is larger here than anywhere else in the subject. A gigawatt constellation with a 2028 date attached and a prototype flying in 2027 is three of those and only one of them is checkable.
The heat has to go somewhere. That was true of the first mainframe in an air-conditioned room, it is true of a rack drawing a hundred and forty kilowatts in Arizona, and it is true of a satellite in a dawn-dusk orbit with nowhere to put the heat except empty sky and a fourth-power law that does not negotiate.
Leave a Reply