Part of Big Tech's War on Users So someone's actually going to try it. On Thursday, October 1, Google launches the first hardware test for Project Suncatcher, its orbital data center moonshot. Back in May, in Evil vs Evil, I spent a good chunk of a trial post explaining why the orbital data center pitch doesn't survive contact with physics. Launch mass. Radiation-hardened everything. And cooling, which is the one that doesn't care how much money you have. Now someone's actually going to find out. And the numbers Google has already shared about its own test satellite are something. I read them twice. Meet MVP The satellite is called MVP. It's about the size of a refrigerator, built with Planet Labs, and it's riding up on a SpaceX Falcon 9 as part of the Transporter-18 rideshare out of Vandenberg. On board: four of Google's TPUs. Four TPUs is, per the reporting, roughly the compute of one data center server. One. And that one server can run for about fifteen minutes before it has to shut down and let the radiators catch up. Google's own write-up is pretty candid that cooling in a vacuum is one of the big unsolved problems. There's no air in space. No fans blowing heat anywhere. The only way to get rid of it is to radiate it off a panel, and that panel has to be big enough to keep up with whatever you're running. So the first orbital data center test is a single server that needs a break every quarter hour. That's my cooling paragraph from May, showing up in the spec sheet before the thing has even left the ground. So What's the Point? That was my first reaction too. What good is a server that runs fifteen minutes, stops, cools off, and starts again? As a data center? None. As an experiment? Actually quite a bit. MVP isn't supposed to do useful work. It runs Gemini so there's a realistic workload on the chips, but what Google actually wants is data it can't get on the ground. Do the TPUs survive launch vibration? How often does radiation flip bits in memory, and does error correction keep up? Do the chips degrade over a year of exposure? Does the thermal system behave the way the models said it would? The fifteen-minute limit is one of the measurements. A year of real numbers on how fast heat builds up and how fast it sheds, across every point in the orbit. The catch is that this constraint doesn't go away when you scale up. Radiated heat depends on surface area. Very roughly, a practical radiator sheds somewhere between a few hundred watts and a kilowatt per square meter, depending on temperature and design. More compute means proportionally more radiator. More radiator means more mass, more launch cost, and a bigger target for debris. A small satellite hitting a thermal wall after fifteen minutes isn't a first-flight glitch that gets patched. It's the physics showing up at small scale. Google's bet is that clever engineering can bend that ratio enough to matter. Maybe it can. This is where they find out how far off they are. Wasn't This Elon's Bright Idea? Here's where I have to correct my own framing a bit, because I'd filed orbital data centers under "Elon's latest" in my head. It wasn't his. Google announced Project Suncatcher last November, complete with a preprint paper. Altman was musing about space data centers on Theo Von's podcast back in July 2025. SpaceX didn't file its FCC application for up to a million orbital data center satellites until this January. Musk didn't invent it. He just made it the loudest. And then he used it as the official reason SpaceX swallowed xAI in February, which, as I noted back in May, had a lot more to do with xAI's burn rate than with anything happening in orbit. The difference between the two approaches is the whole story. Google's plan: one satellite now. Two more in 2027 to test laser links between them. Designs for clusters of around 80 if that pans out. Cost parity with ground data centers maybe by the mid-2030s. Their own head of research says it's early and will take years. "Moonshot" is right there in the name. SpaceX's plan, per the filing: a million satellites, described as "a first step towards becoming a Kardashev II-level civilization." That's a real sentence, in a real FCC document. Musk has also reportedly told xAI employees they'd need a factory on the Moon to build the satellites, plus a mass driver to launch them. Which, if you remember the lunar base paragraph in Evil vs Evil, is exactly the "one expensive station to service another expensive station" problem. Just with a railgun. For those of you unfamiliar with the concept: a railgun takes a solid piece of metal and accelerates it, rapidly and massively, using electromagnetic force instead of gunpowder. Technically, the mass driver Musk described is the railgun's cousin: a track of electromagnets flinging a payload bucket instead of a slug between two rails. Same basic idea, same basic physics. And that physics has a second act, which is what happens when that piece of metal hits something. For a Hollywood rendition, for better or worse, look at Devastator's big scene in Revenge of the Fallen. The USS Kidd, a very real Navy destroyer that the scene was actually filmed on, fires an experimental railgun and takes him apart. Yes, it's a disgrace to such a legendary combiner. But it does illustrate the concept. The real-world version is also instructive, just less explode-y. The Navy spent about $500 million and 15 years on an actual railgun before shelving it in 2021 without ever putting one on a ship. Among the problems: enormous power demands and poor reliability. Every shot meant pulling a huge chunk of power from the ship, and the rails themselves take a beating each time they fire. That was on a destroyer, with a crew, on Earth. Now picture that on the Moon, launching satellites, with a maintenance plan of "we'll figure it out." Last month, around the IPO, SpaceX COO Gwynne Shotwell told CNBC the first AI1 data center satellites would launch in late 2027, with compute going onto Starlink satellites before then. SpaceX claims each AI1 will handle about 120 kilowatts. 120 kilowatts. Per satellite. The ISS, the largest radiator array humans have ever put in orbit, handles about 70. Google's four TPUs need a nap after fifteen minutes. The Eggheads Already Know Now, to be fair to the people actually building things at SpaceX: they're very good at their jobs. Nobody over there thinks you go from a drawing to a million satellites. You can't launch a brand-new rocket design straight off the drawing board either, and Starship has blown up plenty of test articles on its way to working. The engineers will mostly do the same boring incremental tests Google is doing. Bolt some compute onto a Starlink. Watch it overheat. Iterate. Okay, realistically, there'll probably be a few not-exactly-ready showpieces along the way too. Something flashy that flies early so Musk has a livestream to point at and a "we're already working on it" to post about, whether or not it survives the week. Dramatic cannon fodder. But underneath the theater, it'll largely be the same incremental testing. That part doesn't change no matter whose name is on the rocket. The gap was never between the engineers and physics. Physics always wins. The gap is between the pitch and what the engineers deliver. We've seen this one before. Full Self Driving has been one to two years away since 2016. The engineers built a genuinely good driver-assist system and presumably knew exactly what it could and couldn't do. Musk promised a Tesla would drive itself from LA to New York by the end of 2017. A few years later, Tesla's lawyers were in court calling "self-driving" aspirational. The eggheads were right the whole time. They just weren't holding the mic. So here's how I expect this to play out. SpaceX ends up doing something that looks a lot like what Google's doing, just years behind the pitch. The goalposts move. The next vision gets announced. And the million-satellite number sits in an IPO story that investors already paid for. The difference in stakes is worth sitting with. If MVP's numbers come back bad, Google quietly shelves a research project. If the same numbers come back bad for SpaceX, that's a problem in a public company's core growth story. The Irony Is Flying on a Falcon 9 My favorite part: Google's test is launching on a SpaceX rocket. SpaceX gets paid to carry the satellite that may end up producing the most honest public data on how hard orbital compute really is. Data that lands, eventually, right next to a publicly traded company's pitch about doing it at a million times the scale. I'll give Google this. They're doing it the honest way. One server. Clearly labeled as an experiment. Promising to share what they learn. It's the most transparent orbital compute story out there right now. Well, second most. Starman's still up there playing Bowie with zero thermal throttling. He doesn't have a GPU, but he's never needed a cool-down either. I'll check back once the telemetry starts coming in. Find me on Mastodon at @ppb1701@ppb.social.