The previous article in this series ended on a deferral. After working through electricity, water, land, noise, generators, taxes, and the long argument over where these buildings belong, it noted that computing infrastructure might eventually move toward "more remote, unconventional, or even off-world locations," and then set the idea aside as a separate problem.
This is that separate problem.
The case for off-world compute is not a science-fiction flourish attached to an ordinary infrastructure argument. It is a direct consequence of the two articles before it. The first described a machine whose defining constraints are power delivery and heat rejection. The second described a political conflict that exists almost entirely because that machine has to sit somewhere, on land, next to people, drawing from a grid and a watershed that other people also use.
Space does not make those constraints disappear. It rearranges them so completely that almost every assumption in terrestrial data-center design has to be re-derived from scratch. Some problems evaporate. Others become far worse. A few become interesting in ways that have no terrestrial analogue at all.
What follows is an argument that the rearrangement is favorable—eventually, at scale, for a specific class of workload—along with an honest accounting of everything that has to go right first.
A dawn-dusk orbital platform: solar arrays permanently facing the sun, radiators permanently facing deep space, and the machine sandwiched between them. 1. The constraint swap
Here is the entire thesis in one table's worth of prose.
On Earth, mass is cheap and energy is expensive . You can pour as much concrete and steel as you like; a rack weighing one and a half tonnes costs nothing meaningful to move across a data hall. What you cannot easily get is a firm 500 megawatts, an interconnection date inside three years, a water allocation that survives a drought, and a county that wants you.1
In space, energy is cheap and mass is expensive . Sunlight is free, unfiltered, and—in the right orbit—continuous. Nobody is queuing for it, nobody rate-cases it, and nobody lives downwind. What you cannot easily get is a kilogram in orbit, a technician's hands on a failed component, or a way to dump waste heat into anything.
Every serious argument about orbital and lunar data centers is a variation on that swap. The question is never "is space better." It is "at what scale, and at what launch price, does trading an energy problem for a mass problem come out ahead?"
That question now has real money behind it. As of this writing, an NVIDIA H100 has been operated in low Earth orbit, Google has published a research program for putting TPUs on satellites, a European feasibility study has modeled a gigawatt-scale orbital constellation for 2050, and at least three companies have flown data-storage or compute payloads toward or onto the Moon. This is no longer purely a thought experiment.
2. What space actually gives you
Power that never sets
Above the atmosphere, the solar constant is about 1,361 watts per square meter . At Earth's surface, the best you see on a clear day at noon is roughly 1,000 W/m², and the annual average delivered by a fixed terrestrial panel—after night, clouds, dust, haze, seasonal angle, and latitude—is a small fraction of that.
The interesting orbit is dawn–dusk sun-synchronous low Earth orbit : a polar orbit whose plane rides the terminator, the moving line between Earth's day and night sides. A satellite there sees the sun essentially all the time, with only brief eclipse seasons. Its arrays can be fixed rather than tracking. Google's published work on this concept estimates that a panel in such an orbit can deliver on the order of eight times the annual energy of the same panel on the ground.2
This is the single largest advantage, and it is not marginal. It means a space platform needs roughly one-eighth the deployed photovoltaic area for the same annual energy. It means no batteries sized for a winter night, no dispatchable backstop, no capacity market, no interconnection study, and no gas turbine.
It also means the thing that makes terrestrial AI construction politically radioactive—the demand for hundreds of megawatts of firm grid power in somebody's utility territory—simply does not apply.
Water: zero
Not "closed-loop." Not "reduced." There is no evaporative cooling in a vacuum because there is nothing to evaporate into. The entire water-consumption fight described in the previous article—withdrawal versus consumption, potable versus reclaimed, drought-stage curtailment, basin-level budgets—has no orbital equivalent.
The indirect water footprint of the electricity goes to zero too, because the electricity comes from a photovoltaic panel in sunlight rather than from a thermal power plant with a cooling tower.
Land, noise, viewshed, neighbors: none
No rezoning. No setbacks. No cooling-tower fans running at 3 a.m. next to a subdivision. No 40-to-59 dBA complaint measurements. No historic viewshed. No forest clearing, no grading, no stormwater basin, no impervious surface, no habitat fragmentation, no transmission corridor cutting across private land.
The externalities that dominate the local politics of terrestrial data centers are almost entirely externalities of location . Remove the location and you remove them.
Somebody else's grid stops being your problem
The stranded-cost argument, the rate-design argument, the "who pays for the substation" argument, the speculative-load argument—all of it exists because a data center is a customer of a shared public utility. An orbital platform generates its own power at the point of use. It is not a large-load customer. It cannot strand a transmission investment or shift costs onto residential ratepayers, because it is not connected to anything.
And a genuine security property
A facility in orbit or on the lunar surface is immune to flooding, wildfire, hurricanes, ice storms, seismic events, regional grid collapse, and the physical-security threat model that shapes terrestrial site design. Its threat model is different and not obviously easier—but it is different, which is the entire point of geographic diversity.
3. The problem nobody skips: you cannot cool anything with vacuum
This is where the popular version of the argument goes wrong, and it is worth getting right, because it determines the shape of every real design.
Space is cold. Space is also empty , and that is the property that matters. On Earth, a chip sheds heat by conduction into a cold plate, then convection into a fluid, then convection into outdoor air or evaporation into the atmosphere. In vacuum there is no conduction path out of the vehicle and no convection at all. There is exactly one mechanism: thermal radiation .
Vacuum is not a coolant. Vacuum is the best insulator known. A thermos works by being a vacuum. The cold background of space is real—the cosmic microwave background sits at about 2.7 K—but you cannot touch it. You can only glow at it.
On Earth, heat has three exits. In vacuum it has one: radiation from a surface, and nothing else. The radiator math, done out loud
Radiated power follows the Stefan–Boltzmann law:
P = ε σ A T⁴
where σ is 5.67 × 10⁻⁸ W/m²K⁴, ε is the emissivity of the surface (call it 0.9 for a good radiator coating), A is area, and T is the absolute temperature of the radiating surface.3
Run the numbers for a one-gigawatt platform—roughly the scale of the largest AI campuses now being proposed on Earth:
Radiating at 300 K (27 °C, a comfortable coolant temperature): about 413 W/m² . You need roughly 2.4 million square meters —2.4 km²—of radiating surface. A flat panel radiating from both faces cuts the physical panel area to about 1.2 km² .
Radiating at 350 K (77 °C): about 766 W/m² . Panel area drops to roughly 0.65 km² .
Radiating at 400 K (127 °C): about 1,306 W/m² . Panel area drops to roughly 0.38 km² .
Three things fall out of that.
First, the numbers are large but not absurd. A square kilometer of thin deployable film is an enormous structure by current spaceflight standards and a completely ordinary structure by terrestrial industrial standards. A large solar farm covers more ground than that. The concept art showing kilometer-scale platforms is not artistic license; it is what the physics requires. It is also, at 5 to 15 kilograms per square meter for deployable designs, very likely the largest single mass item on the vehicle—a number that returns with force in the economics section below.3
Second, temperature is the whole ballgame. Because radiated power scales with T⁴, raising the radiator from 300 K to 400 K—a 33% increase in absolute temperature—more than triples the heat rejected per square meter. Every degree you can push the coolant loop upward buys enormous radiator mass back.
This inverts a terrestrial instinct. On Earth, cooler is better and warm-water cooling is a modest efficiency play. In space, running hot is a structural design goal. A chip and a coolant loop that tolerate 90 °C rather than 40 °C do not just save energy; they cut the largest single mass item on the vehicle roughly in half.
That single fact quietly reshapes the hardware. It argues for direct-to-chip and two-phase cold plates over air, for silicon rated at high junction temperatures, for packaging that tolerates thermal cycling, and eventually for the kind of microfluidic in-package cooling that terrestrial vendors are already exploring for their own reasons. The AI-era trend toward liquid cooling described in the first article is, conveniently, the exact trend orbital compute needs.
Third, deep space is not always your sink. A radiator only sees the 2.7 K background if it is pointed at empty sky. In low Earth orbit, a panel facing the planet absorbs roughly 240 W/m² of outgoing longwave infrared from Earth, plus reflected sunlight on the day side. The effective sink temperature for an Earth-facing surface is closer to 250–260 K than to 3 K. Orbital radiators therefore have to be oriented, articulated, and held edge-on to both the Sun and the Earth—which constrains the platform's attitude, and through that its array pointing, its antenna pointing, and the formation geometry of every satellite flying near it. Thermal design does not merely size the spacecraft. It dictates how the spacecraft is permitted to fly.
Radiated power scales with the fourth power of temperature. Running the loop hotter is the cheapest mass savings available to an orbital design. And the array is just as big
The other side of the platform has a similar geometry. At 1,361 W/m² and, say, 30% end-to-end conversion efficiency, a photovoltaic array delivers about 408 W/m². One gigawatt of electrical power therefore wants roughly 2.5 km² of array .
So a gigawatt-class orbital data center is, physically, a pair of kilometer-scale sheets —one facing the sun, one facing deep space—with a comparatively tiny dense core of computing hardware sandwiched between them. The compute is the smallest part of the machine. Almost all of the structure exists to gather photons on one side and emit them on the other.
That is a genuinely different building than a windowless warehouse in Loudoun County, and it is worth sitting with the image for a moment. The data center becomes mostly surface .
4. The new problem list
Space removes the terrestrial constraint set and installs a fresh one. Honesty requires enumerating it.
Radiation
Outside the atmosphere and, above low orbit, outside much of the magnetosphere, hardware is exposed to trapped protons and electrons, galactic cosmic rays, and solar particle events. The effects fall into a few families:
Single-event upsets: a particle flips a bit in memory or a register. Common, survivable, correctable.Single-event functional interrupts and latch-up: a particle puts a device into a bad state, requiring reset—or, in the latch-up case, a parasitic conduction path that can destroy the part unless power is cycled fast enough.Total ionizing dose: cumulative degradation of transistor behavior over the mission life, eventually causing permanent failure.Displacement damage: lattice damage, particularly relevant to optical and photovoltaic components.4
The traditional answer is radiation-hardened silicon, which is reliable, expensive, and generations behind commercial parts. That trade is fatal for AI infrastructure: a rad-hard processor from a decade-old process node is not going to train anything.
The modern answer is radiation-tolerant commercial hardware : fly ordinary parts, shield them with mass you were carrying anyway, add aggressive error correction, checkpoint constantly, watchdog everything, power-cycle on upset, and accept a failure rate that software absorbs. This is precisely the hyperscale reliability philosophy described in the first article—reliability moved upward into distributed software—applied to a harsher environment.
The early evidence is encouraging. HPE's Spaceborne Computer-2 has operated commercial off-the-shelf servers aboard the ISS since 2021, running hundreds of experiments and demonstrating that ordinary hardware plus software mitigation survives real orbital conditions.5 2
Nobody can touch it
The first article described commissioning: teams deliberately failing utility power, tripping breakers, killing pumps, and confirming the building recovers. It described technicians replacing dead servers on a schedule while software routes around them.
None of that exists in orbit. There is no hot-swap. There is no cable-management retrofit. There is no walking the row with a thermal camera.
The closest terrestrial analogue is instructive: Microsoft's Project Natick sealed a rack assembly in a nitrogen-filled cylinder and sank it off Orkney for two years. The reported result was roughly an eight-fold improvement in server reliability versus an identical land control group—attributed to the absence of oxygen, humidity, temperature swings, and human hands. Lights-out sealed operation is not only survivable; under the right conditions it is better .6
Even so, the operating model has to change:
Overprovision and expect to run with dead units permanently in place.
Design failure domains so a dead module isolates cleanly rather than taking a string with it.
Build robotic servicing where the economics justify it—in-orbit servicing vehicles are an active commercial field for other reasons.
Treat the platform as consumable: for a hardware generation that is economically obsolete in four years anyway, "replace the satellite" may beat "repair the satellite."
That last point deserves emphasis. The first article noted the awkward mismatch between decades-long building lifetimes and years-long hardware lifetimes. In orbit that mismatch partially resolves itself, because the "building" and the hardware can be launched, deorbited, and replaced as a single unit on the hardware's schedule.
Debris and micrometeoroids
Low Earth orbit is increasingly crowded, and a kilometer-scale structure is a very large collision cross-section. Any serious orbital compute program has to address conjunction avoidance, shielding of critical elements, graceful degradation of punctured radiator panels, and end-of-life disposal. A platform that fragments in a heavily used orbital shell is a liability to everyone, and the Kessler-syndrome concern is a real constraint on how casually anyone should scale this.7
Radiators are, fortunately, one of the more forgiving structures to perforate—a hole in a panel is a local loss, not a catastrophic one—but coolant loops running through them are not.
Getting the data home
An orbital data center that cannot move its results is a very expensive space heater.
Radio-frequency downlinks are spectrum-limited and licensed. The enabling technology is free-space optical communication —lasers. The demonstrated capability is already substantial:
NASA's TBIRD payload demonstrated multi-hundred-gigabit-per-second downlink rates from a CubeSat in low Earth orbit.8
NASA's Laser Communications Relay Demonstration has operated an optical relay in geostationary orbit.9
NASA's Deep Space Optical Communications experiment, flying with the Psyche spacecraft, returned data at hundreds of megabits per second from tens of millions of kilometers away—far beyond lunar distance.10
Satellite-to-satellite optical links are also now routine at commercial scale; large broadband constellations already fly them. A compute constellation would use those links twice: once to move data down, and once—more interestingly—to bind many satellites into a single cluster.
Optical links do have a hard limitation: clouds . Ground stations must be geographically diverse and sited in clear-sky regions, with enough site diversity that the constellation always has a viable path down.
Latency, honestly
This is where careful distinctions matter more than enthusiasm.
Low Earth orbit is not far away. At 550 km, the one-way light time to the ground is under 2 milliseconds; a round trip including a ground station hop is on the order of a few milliseconds. That is comparable to a metro fiber hop between neighborhoods of the same city. LEO compute is not latency-prohibitive. For a great many workloads it is indistinguishable from a regional cloud zone.
The Moon is far away. One-way light time averages about 1.28 seconds ; a round trip is about 2.6 seconds . No amount of engineering fixes this—it is the speed of light. Lunar compute is permanently unsuitable for interactive request-response workloads and permanently fine for anything batch-shaped.
That distinction produces a natural architecture, which the speculation section returns to below: orbit for anything that talks to users, the Moon for anything that talks mostly to itself.
The lunar night
The Moon rotates once per orbit, giving any equatorial site roughly 14 Earth days of continuous sunlight followed by 14 days of darkness . Surviving a two-week night with a multi-megawatt load is not a battery problem in any conventional sense; it is a stored-energy problem on the scale of a small power plant. Regenerative fuel cells—electrolyze water in daylight, recombine hydrogen and oxygen at night—are the usual proposal, along with radioisotope or fission surface power.
Or you avoid the problem geographically, which is the far better answer. See below.
Dust
Lunar regolith is not sand. It is unweathered, mechanically shattered rock: sharp, abrasive, electrostatically charged, and clingy. Apollo hardware degraded visibly from it over days. It fouls seals, abrades surfaces, scratches optics, and—critically for this application—coats radiators and solar panels , degrading exactly the two surfaces whose performance the entire design depends on. Dust mitigation is a serious, unsolved, actively researched engineering problem.11
Law, liability, and jurisdiction
The Outer Space Treaty forbids national appropriation of celestial bodies, makes launching states internationally liable for damage caused by their space objects, and requires that activities be authorized and continuously supervised by a state party.12
Which jurisdiction's data-protection law governs a database physically located on the lunar far side? Who is liable when an orbital platform fragments? Can a company claim exclusive use of a specific ridge on the lunar south pole? These are unanswered and will become contentious well before they become urgent.
The Artemis Accords attempt to fill part of the gap with norms including deconfliction “safety zones” around operations, but they are a political agreement among signatories rather than binding international law, and they leave the property question exactly where the treaty left it.13
Scarcity will force the issue sooner than the law expects. There are only so many near-permanently illuminated ridges at the lunar south pole, and they are simultaneously the best sites for power generation, ice prospecting, communications relay, and human outposts. The same is true of the most useful orbital shells. “No neighbors” is a property of being early, not a property of space.
And there is a governance problem running the other way. A gigawatt-class facility owned by one company, in a jurisdiction-free environment, processing sovereign data, subject to no local regulator and no practical inspection regime, is a concentration of power the terrestrial data-center debate has not begun to model. The previous article argued that communities should demand disclosure of full buildout, resource use, and decommissioning plans. Those demands do not become less reasonable in orbit. They become much harder to enforce.
The quietest place in the solar system
One lunar property deserves separate treatment, because this site is named after it and because the obvious use of it is self-defeating.
The far side of the Moon is permanently shielded from terrestrial radio noise by 3,475 kilometers of rock. It is the most radio-quiet location human beings can reach, and it is the only place in the inner solar system where certain low-frequency radio astronomy is possible at all. The ITU Radio Regulations recognize this by designating a Shielded Zone of the Moon and recommending it be protected from transmissions.14
A data center is an emitter. Putting a large one in the quietest place in the solar system, specifically in order to exploit how quiet it is, would destroy the property that made the location worth having. This is a real conflict rather than a hypothetical one, and the honest version of the lunar argument states it plainly instead of leaving it for radio astronomers to discover later.
Constructive resolutions exist, and none of them are exotic. Compute installations do not intrinsically need to radiate: a facility linked by fiber or by tightly collimated optical links, inside shielded enclosures with disciplined emission control, can be quieter than a comparable terrestrial site. Siting near the limb rather than the deep far side preserves most of the shielding while reducing the conflict. Coordinated exclusion zones around radio-astronomy installations are a well-established instrument on Earth and would transfer.
All of that has to be designed in from the beginning. The version of this argument that treats far-side radio silence as a free amenity to be consumed is the version that gets it wrong.
5. Where the Moon beats orbit—and where it doesn't
Orbit and the lunar surface are usually discussed together and are actually quite different propositions.
Orbit's advantages: cheap to reach relative to a lunar landing, low latency to Earth, continuous sun in the right orbit, no dust, no gravity well to climb back out of, easy replacement cadence.
Orbit's disadvantages: every gram is launched, radiators must be deployed structures, debris risk, no local materials, no shielding except what you carry.
The Moon's advantages are more interesting than they first appear:
Free shielding. Regolith is mass you did not launch. A few meters of it over a vault provides radiation protection, micrometeoroid protection, and enormous thermal inertia. This is the single largest argument for the surface over orbit, and it flips radiation from a design-driving problem to a solved one.A stable platform. Radiators can be built as ground structures rather than deployable spacecraft appendages. There is a floor to stand robots on and a gravity vector to work with.Peaks of eternal light. Certain ridges near the lunar poles—the rim of Shackleton crater is the canonical example—receive sunlight for the overwhelming majority of the lunar year, because the Moon's spin axis is nearly perpendicular to the ecliptic. Sited there, the 14-day night problem largely dissolves. Nearby permanently shadowed craters are simultaneously among the coldest surfaces in the solar system and hold water ice, which is both a coolant reservoir and a propellant feedstock.15 The far side is radio-quiet. The lunar body itself blocks terrestrial radio noise. This is why radio astronomers want the far side, and it is a genuine, physically unique property—a natural electromagnetic sanctuary that exists nowhere else in the inner solar system. It is also the one lunar advantage a data center cannot use without damaging, which is the subject of its own section above.In-situ resources, eventually. Lunar regolith is roughly 40% oxygen and about 20% silicon by mass, with substantial aluminum, iron, calcium, magnesium, and titanium. Everything a solar panel, a radiator, a structural frame, and—far more speculatively—a semiconductor is made of, is present in the dirt.11
The Moon's disadvantages: the 2.6-second round trip, the dust, the cost and difficulty of landing mass, thermal extremes at non-polar sites, and a far less forgiving repair-and-replace cycle.
A lunar facility inverts the orbital design: shielding is free and structural, while the radiators become permanent ground installations. Near-continuous sunlight and permanently shadowed cold traps sit within kilometers of each other at the lunar poles—the two resources an off-world data center needs most. 6. The people actually building this
The concept has moved from conference papers to funded programs and flown hardware faster than most observers expected. What follows is a snapshot as of publication; this field is changing quickly, and specific figures come from company and agency statements rather than independent verification.
Starcloud
A venture-backed startup (formerly Lumen Orbit, and an NVIDIA Inception company) that launched Starcloud-1 in late 2025 carrying an NVIDIA H100 GPU—by the company's account the most capable GPU ever operated in space, by a very large margin over prior flown hardware. Its stated roadmap runs through a Blackwell-generation follow-on toward a 5-gigawatt orbital facility built around roughly 4 km × 4 km solar and radiator structures. The 5 GW concept is a decade-plus aspiration, not a manifest. The H100 is real and it flew.16
Google — Project Suncatcher
Announced in late 2025 as a research moonshot: constellations of solar-powered satellites carrying TPUs , flying in tight formation in dawn–dusk sun-synchronous orbit, connected to each other by free-space optical links at data-center-interconnect bandwidths. Google published radiation testing of its Trillium TPUs in a proton beam, modeled the formation-flying and optical-link physics, and stated a plan to fly two prototype satellites with Planet Labs by early 2027 .2
The most useful thing Google published was not a design but a threshold : the analysis argues orbital compute reaches rough cost parity with terrestrial data centers if launch prices fall to around $200 per kilogram by the mid-2030s. That is a falsifiable claim, tied to a number anyone can watch.
Axiom Space
Flew AxDCU-1 , an orbital data center unit, to the ISS in 2025, working with partners on containerized workloads and optical relay connectivity. Axiom's stated plan is to operate data center nodes on its commercial station, positioning orbital compute as a service offered from crewed infrastructure rather than a dedicated free-flyer.17
Lonestar Data Holdings
Focused on the Moon rather than orbit, and specifically on data resilience —the pitch is off-world backup and disaster recovery, a use case where a 2.6-second latency is entirely irrelevant. Lonestar flew a software-based data payload on Intuitive Machines' IM-1 mission in 2024 and a physical solid-state unit, "Freedom," toward the lunar surface on IM-2 in 2025. That lander's mission was cut short after it came to rest in an unintended orientation, which is itself an honest data point about the difficulty of the surface option.18
Thales Alenia Space — the ASCEND study
The most rigorous public analysis to date. ASCEND (Advanced Space Cloud for European Net zero emission and Data sovereignty), a European Commission–funded feasibility study led by Thales Alenia Space, examined whether orbital data centers could be environmentally preferable to terrestrial ones—not merely technically possible.
Its conclusion was conditional and worth quoting in spirit: yes, but only if launch vehicles become roughly an order of magnitude cleaner in emissions terms. The study sketched a deployment of around 13 satellite building blocks totaling about 10 MW by 2036 , scaling toward 1 GW by 2050 with on the order of 1,300 blocks.19
That conditionality is the most credible thing anyone has said about this field. Orbital compute is not automatically green. It is green if and only if the rocket that puts it there stops being the dominant term in its lifecycle footprint.
SpaceX
Public statements from SpaceX leadership have described using the high-bandwidth optical links on next-generation Starlink satellites as the backbone for orbital compute, with aspirational figures of scaling to enormous annual orbital power deployment. Treat the specific numbers as directional rather than committed. The relevant point is structural: the company that owns both the cheapest launch capability and the largest optical mesh in orbit is the one best positioned to make the mass problem go away, and it knows it.20
Blue Origin
Jeff Bezos has publicly predicted gigawatt-scale data centers in space outcompeting terrestrial ones on a 10-to-20-year horizon, framing space-based industry generally as the destination for energy-intensive activity. Blue Origin's heavy-lift and in-space logistics work is the enabling half of that argument.
The supporting cast
HPE Spaceborne Computer-2 — commercial servers running continuously on the ISS since 2021; the longest-running practical evidence that COTS hardware survives orbit with software mitigation.Ramon.Space — commercial radiation-tolerant space computing, supplying the middle ground between rad-hard and consumer silicon.NASA LunaNet and ESA Moonlight — the communications and navigation infrastructure programs that would make any lunar facility reachable. A lunar data center without a lunar relay network is a stranded asset.21 NASA CLPS providers (Intuitive Machines, Firefly, Astrobotic and others) — the commercial delivery pipeline that makes lunar surface payloads routine enough to iterate on.22 Microsoft Project Natick — terrestrial, underwater, and finished, but the best available proof that sealed, unattended, lights-out operation improves rather than degrades reliability.
The near-term reality: single racks and single accelerators flown as demonstrators, establishing that commercial hardware survives before anyone commits to kilometer-scale structures. 7. Now the speculation
Everything to this point is defensible. What follows is deliberately not.
The economics invert, and then keep inverting
Start with the launch-cost curve, because it drives everything.
Price out the one-gigawatt platform from section 3, using current hardware as the unit. A rack in the class of NVIDIA's GB200 NVL72 masses roughly 1.4 tonnes and draws roughly 120 kilowatts.23 11,000 tonnes of compute hardware. Add the radiators: at 10 kg/m², the 1.2 km² of panel a 300 K loop requires is another 12,000 tonnes . Add the array: 2.5 km² of thin-film photovoltaics at 1 to 2.5 kg/m² is 2,500 to 6,000 tonnes . Put structure, avionics, propulsion, and margin on top.
Call it 30,000 tonnes for a gigawatt, give or take a factor that depends almost entirely on how hot you are willing to run the coolant.
Now price the launch:
At $1,500/kg , roughly today's commercial heavy-lift rate: about $45 billion .
At $500/kg : about $15 billion .
At $200/kg , the threshold Google's analysis identifies: about $6 billion .2
Compare that to the hardware. At roughly $3 million per rack, 8,300 racks cost about $25 billion before anything leaves the ground.
That comparison is the entire argument in one line. At today's prices, getting the platform to orbit costs nearly twice what the accelerators inside it cost—which is why this is currently a demonstration business rather than an industry. At $200/kg, launch drops to about a quarter of the silicon bill , and the free power, the absent water, the absent interconnection queue, and the absent county zoning board stop being interesting side benefits and start being the entire business case.
The interesting part is what happens after parity. On Earth, adding a gigawatt means finding a gigawatt, which means transmission, generation, politics, and years. In orbit, adding a gigawatt means unrolling more square kilometers of film. The marginal cost of the next increment of capacity has an entirely different shape: it is a manufacturing-and-launch problem, subject to learning curves and mass production, rather than a permitting-and-utility problem, subject to negotiation and consent.
Manufacturing problems get cheaper at a predictable rate. Permitting problems do not.
Compute becomes the only sane export from space
Space industrialization has always had a shipping problem. Asteroid metals, orbital manufacturing, lunar helium-3—every proposal founders on the cost of moving physical goods down a gravity well.
Computation has no shipping problem. You launch the mass once, and thereafter you export photons . A result weighs nothing. The bandwidth is a solved engineering problem and getting rapidly better.
This may be the actual answer to a question the space industry has been asking for fifty years: what can you make up there that is worth making up there? Not alloys. Not crystals. Not fuel. Answers.
Thermal zoning: an architecture of distance
Push the latency distinction to its conclusion and a natural three-tier architecture appears:
Ground: everything genuinely latency-critical and everything legally required to stay in a jurisdiction. This never goes away, and the previous article's argument about siting it responsibly remains the near-term priority.Low Earth orbit: the interactive tier. Milliseconds from users, continuous power, no water, no land. Inference serving, real-time processing, and—elegantly—Earth-observation processing done in situ , so satellites downlink conclusions instead of raw imagery.The Moon and beyond: the batch tier. Training runs, simulation, rendering, archival processing, and cold storage that nobody needs to reach in under three seconds. Shielded, thermally privileged, and buried under regolith that costs nothing.
The 2.6-second lunar round trip stops looking like a defect and starts looking like a filter that sorts workloads into the right place.
Silicon from dirt
The long game is not launching data centers. It is launching the factory that builds them.
Lunar regolith contains, in abundance, silicon, oxygen, aluminum, iron, and titanium. Solar arrays, structural members, radiator panels, and reaction mass are all in principle manufacturable from local material. Vacuum is a free ultra-high-vacuum processing environment, which is exactly what semiconductor fabrication spends enormous money to create artificially. Sunlight is a free, intense, unfiltered process-heat source.
Nobody is close to fabricating an advanced logic node on the Moon, and the gap between "extract silicon from regolith" and "produce a modern processor" is a chasm, not a step. But the direction is clear: the moment any meaningful fraction of the platform can be built from local material, the mass constraint—the entire other half of the constraint swap—begins to relax. And when the mass constraint relaxes against an energy budget that is effectively unbounded, the growth curve stops looking like infrastructure and starts looking like something else.
The thermodynamic ceiling
Here is the argument that makes all of this feel less like enthusiasm and more like arithmetic.
Every watt of computing becomes heat. The first article said so in its cooling section, and it is true everywhere, not just in a data hall. Earth as a whole is a closed thermal system that sheds heat by radiating to space at a fixed rate. There is a finite quantity of waste heat the planet can absorb before the waste heat itself—entirely separate from carbon emissions—becomes the problem. That ceiling is comfortably distant at current consumption and not distant at all on the growth curves people are drawing for the next century.
The way out of a closed thermal system is to not be inside it.
But there is a second and stranger thermodynamic argument underneath the first, and it points somewhere more specific than "off the planet."
Landauer's principle sets a floor on the energy cost of irreversible computation: erasing one bit costs at least kT ln 2 joules, where k is Boltzmann's constant and T is the temperature at which the computation happens.24
Note the T . The floor is linear in temperature . Computing at 40 K—the measured temperature of a permanently shadowed crater floor near the lunar south pole—has a thermodynamic floor roughly seven and a half times lower than computing at room temperature. Computing against the 2.7 K cosmic background has a floor more than a hundred times lower.
For anything operating near the physical limits of computation, cold is not a comfort. Cold is compute. The same joule buys an order of magnitude more thinking at the bottom of a lunar cold trap than it buys in Virginia, and no amount of cleverness at the warm end closes the gap, because it is not an engineering gap.
Which reframes those permanently shadowed craters entirely. They are not a convenient heat sink that happens to sit near the peaks of eternal light. In the very long run they are among the most thermodynamically valuable real estate accessible to us—and they sit a few kilometers from the best power sites on the Moon, which is either a remarkable coincidence or simply what a good site looks like.
The far end
Extend the argument and you arrive at a familiar endpoint by an unfamiliar road.
A civilization limited by energy eventually stops collecting the fraction of its star's output that happens to strike a planet and starts collecting an appreciable fraction of the whole. Freeman Dyson made that argument in 1960, and what falls out of it is not a rigid shell but a swarm : an enormous population of independently orbiting collectors.25
The satisfying part is that the governing constraint never changes. You are limited by radiating area, and the useful work you can extract is set by the temperature gradient between your collectors and the sky. P = ε σ A T⁴ governs a hypothetical Matrioshka brain exactly as it governs a hundred-megawatt satellite and exactly as it governs the cooling tower behind a warehouse in Loudoun County. The physics acquires no new terms as the ambition grows. It simply gets applied to larger numbers.
Not a Dyson sphere built out of ambition, then, but a compute swarm built out of the entirely mundane observation that radiators want sunlight on one side and vacuum on the other, that there is a great deal of both available, and that nobody living anywhere has to look at it.
A closing absurdity, offered seriously
There is one more consequence of taking Landauer literally, and it is either the best or the worst idea in this article.
Sandberg, Armstrong, and Ćirković proposed what they called the aestivation hypothesis : that an advanced civilization seeking to maximize its total lifetime computation might rationally do almost nothing now, and instead wait—for billions of years—while the universe expands and cools, because the same stored energy buys vastly more computation against a colder background later.26
That is very probably wrong. It is also a rigorous consequence of an argument that began with the question of where to put a radiator.
The far end of the argument: computation as an off-world industry, sized by available sunlight rather than by available grid capacity. 8. What would have to be true
Speculation is only worth reading if it can be checked. Here is what would have to happen, roughly in order, for any of the above to stop being a story:
Launch cost falls to a few hundred dollars per kilogram. This is the load-bearing assumption. Everything else is downstream. Watch published prices per kilogram to LEO, and watch fully reusable heavy lift actually flying at cadence.Launch emissions fall by roughly an order of magnitude. ASCEND's condition.19 Radiator technology matures at scale. Deployable, kilometer-class, high-emissivity, puncture-tolerant, and capable of running hot. This is a mass-manufacturing problem more than a physics problem, which is encouraging.Commercial silicon proves out over multi-year missions. Not weeks on a demonstrator—years, with measured failure rates, under real solar particle events.Optical downlink and inter-satellite mesh reach data-center-interconnect bandwidths reliably. Including cloud-diverse ground segments.Somebody demonstrates a genuinely unattended multi-year platform. No servicing, no intervention, graceful degradation, useful output at end of life.The legal framework catches up. Data jurisdiction, orbital debris liability, lunar site use, and—eventually—whether "this database is on the Moon" is a compliance answer or a compliance problem.A lunar surface facility survives a full year of dust, thermal cycling, and night. Even a small one.
Two conditions would work the other way, and they deserve equal billing. If advanced nuclear, geothermal, or serious grid reform substantially unblocks terrestrial power, the pressure driving this entire search relaxes—space competes against Earth's worst case, not its best. And if compute-per-watt improves faster than demand grows, the capacity crisis softens and nobody needs a radical answer at all.
If the eight land and the two do not, the case makes itself. If launch costs stall, none of the rest matters and terrestrial data centers remain the only game—which makes the previous article's questions about siting, water budgets, and rate design the urgent ones for a long time yet.
Conclusion: the cloud gets a new address
The previous article ended with the observation that the cloud has an address—that every stored photograph and every generated token depends on land, metal, concrete, water, electricity, and somebody's neighborhood.
That will remain true for decades. The overwhelming majority of computing will stay on Earth, connected to terrestrial grids, subject to terrestrial politics, and owed the same honest scrutiny the second article argued for. Nothing here is an excuse to stop asking hard questions about the campus proposed down the road, and anyone using "we'll move it to space eventually" to deflect a zoning hearing is arguing in bad faith.
But the direction of travel is not really in doubt, because the constraints are not really negotiable. Compute demand wants energy. Energy production on Earth is limited by land, water, grids, emissions, and consent. Energy in space is limited by the size of a sheet of film. Every watt of computing becomes heat, and Earth has one radiator that everybody shares while space has as many as you care to unroll.
Data centers went from a room, to a warehouse, to a campus, to a factory. The interesting question was never whether they would keep going. It was where they would go when they ran out of Earth.
The Moon has been sitting there the whole time: no weather, no neighbors, no water table, no grid, fourteen days of uninterrupted sunlight on the polar ridges, cold traps a few kilometers away that are colder than anything else within reach, and a sky that is nothing but heat sink. It is the best data-center site in the solar system, and the single most honest argument against building on its far side is that the silence there may be worth more than the real estate.
Somebody is going to build anyway. And the question this series has been circling—where does the machine go, and who pays for where it goes—does not dissolve when the address leaves the planet. It just gets asked, for once, in a place where nobody has to live next door to the answer.
Sources and further reading
Numbered sources are cited inline where the claim they support appears. Figures attributed to companies are drawn from their own public statements and are commercial projections rather than independently verified commitments.
1 U.S. Department of Energy Secretary of Energy Advisory Board, Recommendations on Powering Artificial Intelligence and Data Center Infrastructure . The terrestrial power constraints — hyperscale requests in the 300-to-1,000-megawatt range against multi-year grid lead times — that motivate the off-world argument.
2 Google Research, Exploring a space-based, scalable AI infrastructure system design . Project Suncatcher: the dawn–dusk sun-synchronous orbit analysis and eight-times solar productivity estimate, proton-beam radiation testing of Trillium TPUs, formation-flying optical link modeling, the Planet Labs prototype plan, and the launch-cost parity threshold.
3 NASA, State of the Art of Small Spacecraft Technology: Thermal Control . Radiator coatings and emissivity, deployable radiators, loop heat pipes, and the practical constraints on radiative heat rejection in vacuum.
4 NASA Goddard Space Flight Center, Radiation Effects and Analysis . Single-event upsets, functional interrupts and latch-up, total ionizing dose, and displacement damage in spaceflight electronics.
5 Hewlett Packard Enterprise, Spaceborne Computer-2 . Commercial off-the-shelf servers operating aboard the ISS since 2021 with software-based error mitigation.
6 Microsoft Research, Project Natick . Sealed underwater data center; reported roughly eight-fold server reliability improvement over an identical land-based control group.
7 European Space Agency, Space Debris . The orbital debris environment, collision risk, and the case for disposal planning on large-area structures.
8 NASA Goddard Space Flight Center, TeraByte InfraRed Delivery (TBIRD) . Multi-hundred-gigabit-per-second optical downlink from a CubeSat in low Earth orbit.
9 NASA, Laser Communications Relay Demonstration . Geostationary optical relay operations.
10 NASA Jet Propulsion Laboratory, Deep Space Optical Communications . High-rate laser communications demonstrated at interplanetary distances aboard the Psyche spacecraft.
11 Lunar and Planetary Institute, Lunar science and regolith resources . Regolith composition, its abrasive and electrostatic properties, and in-situ resource potential.
12 United Nations Office for Outer Space Affairs, Treaty on Principles Governing the Activities of States in the Exploration and Use of Outer Space . Non-appropriation of celestial bodies, launching-state liability, and continuing state supervision of non-governmental activity.
13 NASA, The Artemis Accords . Principles for cooperative lunar activity, including deconfliction “safety zones”; a political agreement among signatories rather than binding international law.
14 International Telecommunication Union, Recommendation ITU-R RA.479, “Protection of frequencies for radioastronomical measurements in the shielded zone of the Moon.” The basis for treating the lunar far side as a protected radio-quiet reserve.
15 NASA Science, The Lunar South Pole . Polar illumination conditions, near-permanently lit crater rims, and permanently shadowed regions containing water ice.
16 Starcloud, company site . Starcloud-1, the first H100-class GPU operated in orbit, and the stated roadmap toward gigawatt-scale orbital platforms.
17 Axiom Space, company site . The AxDCU-1 orbital data center unit and planned data center nodes on Axiom Station.
18 Lonestar Data Holdings, company site . Lunar data-resilience payloads flown on Intuitive Machines IM-1 and IM-2.
19 Thales Alenia Space, ASCEND feasibility study on space data centers . European Commission–funded study concluding that orbital data centers can be environmentally favorable only conditional on roughly an order-of-magnitude reduction in launcher emissions.
20 SpaceX, Falcon 9 and Starship . Published payload capacities and pricing, the basis for the cost-per-kilogram figures used above.
21 NASA, LunaNet and European Space Agency, Moonlight . Planned lunar communications and navigation infrastructure.
22 NASA, Commercial Lunar Payload Services . The commercial delivery pipeline for lunar surface payloads.
23 NVIDIA, GB200 NVL72 . Rack mass and power figures used in the launch-cost arithmetic.
24 Rolf Landauer, Irreversibility and Heat Generation in the Computing Process , IBM Journal of Research and Development, 1961. The thermodynamic floor on irreversible computation, and the reason temperature appears in the cost of a bit.
25 Freeman J. Dyson, Search for Artificial Stellar Sources of Infrared Radiation , Science, 1960. The original statement of the stellar-scale energy-collection argument.
26 Anders Sandberg, Stuart Armstrong, and Milan Ćirković, That is not dead which can eternal lie: the aestivation hypothesis for resolving Fermi's paradox , 2017.
A note on figures
Radiator and array sizing in this article is computed directly from the Stefan–Boltzmann law and the solar constant using stated assumptions (ε = 0.9, 30% array efficiency) and is intended to convey scale, not to specify a design. Company roadmaps, capacities, and dates come from public statements by those companies and agencies as of publication; they are commercial projections rather than independently verified commitments, and this field is changing quickly enough that several of them will have moved by the time you read this.
