Last December, Sundar Pichai announced that Google had decided to put data centers in space. Project Suncatcher was the moonshot — Pichai's word — and the framing was that the sun puts out "100 trillion times more energy than what we produce on all of Earth today," and Google would like access to that. Two pilot satellites with Planet Labs are scheduled for early 2027. "A more normal way to build data centers" is how Pichai described it, with a horizon of about a decade.
Then a former NASA engineer with a PhD in space electronics, who happened to have spent a decade at Google's YouTube and Cloud-AI infrastructure, sat down at a blog called taranis.ie, opened with "For clarity: I am a former NASA engineer/scientist with a PhD in space electronics. I also worked at Google for 10 years," and proceeded to walk through why none of this works. The post — credited to a single byline, Taranis — went to the front page of Hacker News and ran 361 comments deep, then to Lobsters, and game-engineering veteran Christer Ericson endorsed it on X with "data centers in space is a fantasy."
Here is the math I cannot stop thinking about. The largest solar array ever deployed in space is on the International Space Station. It produces about 200 kilowatts at peak, covers about 2,500 square meters — half a football field — and required several Shuttle missions to install. An NVIDIA H200 draws 700 watts at maximum thermal design power, and Taranis's rule of thumb of about one kilowatt per GPU once support hardware is counted is conservative. So an ISS-sized solar array, the largest humans have ever flown, can power approximately two hundred GPUs. NVIDIA's standard 72-GPU rack ships in a DGX configuration that already exists. One ISS in orbit, powered to its peak, runs three of them.
Stargate Norway — OpenAI's first European AI gigafactory, announced in summer 2025 — targets 100,000 GPUs by the end of 2026. To match that in orbit you would need five hundred ISS-sized solar arrays. The ISS itself took thirteen years and forty-some launches to build.
This is the kind of piece where the rest of the article is the unpacking.
Power
The intuition behind a space data center is that the sun is up there and it is enormous, and so power is, in some loose sense, free. The intuition is wrong in a specific way. Solar panels in orbit are essentially the same panels that cover the roof of any rooftop installation, with the same conversion efficiency, slightly more sunlight (the atmosphere absorbs a few percent), and no night. The advantages are real but they are a couple of multipliers, not a step change. The disadvantage is that the panels have to be deployed in vacuum, which is hard, and held there, which is harder. The four primary ISS solar wings were each delivered on a separate Shuttle mission — STS-97, STS-115, STS-117, STS-119 — between December 2000 and March 2009. That's nine years to get four solar wings in orbit. The deployment is the easy part of operating them.
The other power option is nuclear, and in orbit that means radioisotope thermoelectric generators, which are the small plutonium-powered heat engines that drove the Voyager and Cassini probes. They produce 50 to 150 watts apiece. As Taranis observes, that is not enough to power one H200, even before you have asked anyone for a sub-critical lump of plutonium-238 and explained what you plan to do with it. The reactors NASA actually flies in orbit aren't reactors at all. The reactors that would be reactors aren't ready, and the agencies that would have to license them are unenthusiastic.
So the math is solar, and the math is two hundred GPUs per ISS-sized array, and the marketing slide where the sun is "100 trillion times the energy" of all human civilization is operating on a scale where the energy that matters is not the energy in the sun but the energy you can collect on a panel, beam to a chip, and not boil the chip with.
Thermal regulation
This is the section that broke me on first reading and is, in the long run, the structural argument against the whole project.
There is no air in space, which means there is no convection. On Earth, the way a data center stays alive is that hot air rises off a chip, gets entrained in a heatsink fan, gets dumped into a cold aisle, gets pumped through liquid loops or chilled-water exchangers, and ultimately gets convected into the atmosphere. The atmosphere has been doing the heavy lifting for the entire computing industry for sixty years. The atmosphere is the actual cooling system; everything else is a connector to it.
In orbit there is no atmosphere. There is no medium for the heat to leave through except radiation, which is the same mechanism that makes a hot piece of metal glow. The Stefan-Boltzmann law puts a cap on how much heat per square meter a radiator panel can dump, and the cap is unforgiving.
The ISS has the largest active thermal-control system humans have ever flown, the Active Thermal Control System (ATCS). It uses an ammonia coolant loop and a series of large radiator panels that face away from the sun. The system's dissipation cap is 16 kilowatts. Sixteen kilowatts is the peak heat budget of the ISS thermal system. That is the equivalent of approximately sixteen H200 GPUs, or roughly one-quarter of an NVIDIA DGX rack. The radiator panels themselves are 13.6 m × 3.12 m, about 42.5 m². To dissipate the full 200 kW from a notional ISS-sized solar array, the same scaling on radiator area lands around 531 square metres of additional radiator panel, on top of the 2,500 m² solar array that's already there. The satellite is now substantially larger than the ISS, and it is dissipating the heat of three server racks.
There is no engineering shortcut here. Stefan-Boltzmann is a temperature-to-the-fourth-power relationship and the radiator surface temperature is bounded by the temperature at which the chips you're cooling stay alive. You can play tricks at the margin — heat pipes, two-phase loops, anti-sun-side radiators — but the margins are tens of percent. The marketing pitch "space is cold, so cooling is easy" is the inverse of the truth: space is insulating, and the only way to lose heat is to radiate it, and your radiator is what bounds the size of the satellite.
Radiation
The third reason this doesn't work is that the chips you would put in orbit don't survive in orbit.
GPUs and TPUs and the high-bandwidth memory they depend on are the worst-case silicon for radiation tolerance. The transistors are small, which means a single charged particle passing through one is a larger fraction of the gate's relevant area, which means a single hit is more likely to flip a bit (a single-event upset) or, worse, to cause a single-event latch-up where a transistor turns itself on, draws current it shouldn't, and burns the chip out. The die area is also enormous, which means more hits per second per chip. And the cumulative dose effect over months — transistors switching slower, drawing more power, eventually crossing into nonfunctional — is exactly the failure mode you don't want at scale, because you can't service it.
Chips actually designed for space use a different gate topology and much larger geometry — typically the BAE RAD750, based on a PowerPC architecture from the late 1990s, or its successors. Per Taranis's framing, the typical processor on an actually-flying spacecraft has compute roughly equivalent to a 2005-era PowerPC. The relationship is not negotiable: radiation hardness comes from larger transistor geometry; performance comes from smaller. Pick one.
You could ship the H200s anyway — Taranis calls this the "YOLO approach" — and that's how cubesats often work, which is also why cubesats often fail within weeks. Shielding helps a little, except past a thin layer it makes the problem worse: a cosmic ray hitting a sufficient mass of shielding produces a shower of secondary particles, and now you have many hits where you used to have one. Mass is always at a premium on a satellite. The result is that for a long-duration orbital data center — which it has to be, because at $5,000 per kilogram of launch mass, it isn't economic for anything short — you can't ship the GPUs you actually want, and the chips you can ship aren't the ones the customer is paying for.
Google's own Project Suncatcher paper acknowledges this and reports that Google ran TPUs in a particle accelerator and they survived for a simulated five years. This is a real result and it is not the same thing as the chips being the chips you would ship to a customer. The accelerator simulates dose; it doesn't simulate every failure mode at sufficient fidelity, and the chip you put in space still has to be one that exists, which means it's a TPU one or two generations old by the time it flies, and it's competing in the AI-compute market against a TPU running on the ground that is being upgraded twice a year.
Communication
The smaller of the four problems, but worth saying. A typical orbital satellite communicates with the ground at single-digit gigabits per second on radio. Optical inter-satellite links are improving but they require atmospheric clarity to reach the ground, and a single rack of GPUs in a terrestrial data center can saturate hundred-gigabit interconnect to its neighbors as a floor. The space data center, until somebody builds the orbital optical mesh that doesn't exist, has the I/O of a rack from 2010 with the compute of three.
The marketing-physics gap
Project Suncatcher's own paper gets to the punchline by halfway through. "Launch costs could drop below $200 per kilogram by the mid-2030s." SpaceX's Falcon 9 is currently around $2,720 per kilogram to LEO; Starship's projected mature cost — if Starship works at full reusable cadence, which it currently does not — is in the $200–500 range. Google's paper is asking the reader to assume that Starship works, at scale, with the cost curve fully realized, in eight to ten years, and then the math becomes plausible.
The math also doesn't actually become plausible. At $200/kg, the launch costs of a 200 kW solar array (a large structure of conservative-but-substantial mass) plus the corresponding 531 m² radiator plus the rad-hard or non-rad-hard compute payload plus the maneuvering and station-keeping plus the redundancy still produces three racks of orbital compute for tens of millions of dollars in launch costs alone, before the spacecraft bus or the deployment operations. You can build the equivalent terrestrial data center, including the grid hookup and the power-purchase agreement and the cooling tower, for less than that.
The honest version of the case for space data centers is that the grid is the binding constraint, not the silicon, and the AI industry has lost so much social license over the past two years on water consumption, farmland conversion, and electricity-rate impacts that putting the next 100,000 GPUs into someone else's atmosphere is starting to look like the path of least resistance. That's not engineering. That's the marketing department's response to the political problem the engineers gave them.
Why the marketing keeps coming back
The reason a fundamentally non-functional idea keeps showing up at scale is that the AI industry, in 2026, has a supply story it cannot tell honestly. Demand for AI compute is growing faster than grid hookups can be brought online. Hyperscaler PR, investor decks, and ESG reports all need a future answer to the question of how this gets built; "we'll figure out the grid eventually" doesn't fit on a slide. "Datacenters in space" does. The fact that the slide is in the deck is itself doing work — it absorbs the question for the duration of the meeting and lets the actual capacity expansion plan, which involves a lot of natural-gas peaker plants and disputed grid interconnects, proceed unexamined.
The other thing the slide does is launder. The single hardest political problem the AI industry has right now is that its capacity expansion is locally visible: the gas plants are in someone's town, the data center is drinking someone's aquifer, the rate hike is on someone's bill. A satellite is in nobody's town. A satellite, in the marketing imagination, runs on solar and harms no one. The fact that the satellite would have to be the size of the ISS, and would deliver three racks of compute, and would do so at a unit cost incompatible with the AI industry's actual capex curve, is information that doesn't survive the trip from the engineering team to the keynote.
There is one argument I find honestly worth taking seriously: the orbital infrastructure has to start somewhere, Project Suncatcher's two pilot satellites are an order of magnitude cheaper than nothing, and the long-run learning curve on rad-hard compute and orbital deployment might justify the spend even if the short-run economics don't close. This is a real argument and it is the one Google's paper actually makes when you read past the marketing layer. It is also a different argument from "this is how we'll meet AI compute demand," and the gap between the two is the gap that Pichai's keynote invites the audience not to notice.
What stays on the ground
Engineering arguments dressed in physics don't shift with political winds. Stefan-Boltzmann doesn't care that 2026 is a hard year for hyperscaler PR. The radiator-size problem will be the same in 2030 as it is now. The radiation-tolerance problem will be the same. The launch-cost problem might come down to within an order of magnitude of viable, and the bet on Starship reaching its target cost by the mid-2030s is the same kind of bet all the other 2010s reusable-rocket projections were — except this one lives in the appendix of a Google research paper instead of the keynote.
Three racks of orbital compute, on a structure the size of the ISS, riding on a launch cost that is almost certainly twice what the marketing assumes, in a chip generation that is one or two cycles behind the ground, with a thermal envelope that constrains everything else — that is the thing the slide is about. The slide is about the slide. The slide is in the deck because the deck needed something. The slide will be in the next deck for the same reason.
Meanwhile the next 100,000 GPUs that OpenAI is bringing online for the next training cycle are getting installed in northern Norway by its joint venture with Nscale and Aker, drawing 230 megawatts of renewable power, and cooled by a closed-loop direct-to-chip liquid loop. The data center stays on the ground because it has to. The space data center stays in the slide because it can.