--- title: "Heat is the hard part." date: 2026-10-05 url: https://remiam.co.uk/notes/heat-is-the-hard-part tags: [Space, AI, Hardware] read_time_minutes: 5 description: "Remiam's read on Google's Project Suncatcher TPU satellite: why cooling, not radiation, is the real constraint, and what that means for the kiosks, installations and edge hardware we build." --- # Heat is the hard part. *Published 2026-10-05 · 5 min read · by Liam (Remiam)* On 1 October Google put four of its Trillium TPUs into orbit on a Planet-built satellite, the first step of Project Suncatcher. The chips had already passed their radiation tests on the ground. The open question is the one we meet on every enclosure we build: where does the heat go. On 1 October a Falcon 9 lifted off from Vandenberg on SpaceX's Transporter-18 rideshare, and one of the payloads it deployed was a fridge-sized satellite built by Planet carrying four Google Trillium TPUs. It is the first hardware step of Project Suncatcher, Google's long-term research into whether machine learning compute can run in orbit, powered by sunlight instead of a grid connection. Google has since confirmed the satellite is in contact and working. Four TPUs is roughly one server's worth of accelerators, on a power budget of about a kilowatt. Nobody is training a frontier model up there this year. The point of the mission is to find out what breaks. ## Radiation was the problem everyone expected The obvious worry about putting data-centre chips in space is radiation. Google tested that before launch, firing a proton beam at Trillium chips at UC Davis while they ran real workloads. They took a total dose of around 15 krad, many times the roughly 750 rad a shielded five-year mission in that orbit would see, without the core TPU logic failing. That is a strong result for silicon that was never designed for space. So the question that is left is less glamorous. In a vacuum there is no air to blow over a heatsink. The only way to lose heat is to radiate it from a surface as infrared, which is slow, and the satellite swings between full sun and shadow every orbit. Reporting since the launch has been blunt about it: the radiators, not the radiation, decide how long the chips can run flat out before they have to back off. > Compute is easy to put in a box. Getting the heat back out of the box is the design. ## We have this problem on the ground We do not build satellites. We do build sealed boxes with computers in them: kiosks, booths, wall displays, installations that run all day in a shopping centre or at an event, and increasingly a small model running locally because the venue Wi-Fi cannot be trusted. Almost every one of those projects has had the same quiet moment where the software is finished and the enclosure is not, because the machine throttles after forty minutes in a closed cabinet under hot lights. It is never in the brief. Nobody asks for thermal design. They ask for a screen that responds instantly at four in the afternoon on the third day of a show, which turns out to be the same thing. ## What Suncatcher gets right - It tests the boring constraint in the real environment. The radiation work was done on the ground because it could be. The thermal behaviour over hundreds of sun and shadow cycles can only be learned in orbit, so that is what the satellite is for. - It treats duty cycle as a design input. If the chips have to rest between bursts, the software has to know that, and schedule around it, rather than pretending the hardware can sustain its peak. - It is honest about the economics. Google's own 2025 paper puts rough cost parity with ground data centres in the mid-2030s, and only if launch prices fall to around $200 per kilogram. A prototype is a way to find out, not a promise. ## What we would take from it - Put the hardware in its real enclosure, in its real conditions, before the software is signed off. A bench test on an open desk tells you very little. - Design the software for sustained load, not peak load. If a local model or a render loop will run all day, measure it after an hour, not after a minute. - Budget for the dull part. Ventilation, power and heat sit at the bottom of the spec and at the top of the list of things that go wrong on site. Should clients care about AI in orbit? Not yet, and maybe not for a decade. The useful lesson is closer to home. Every piece of hardware we ship has a heat budget whether anyone wrote it down or not, and the projects that go smoothly are the ones where somebody did. ## References 1. [Converge Digest, Google accelerates Project Suncatcher with first orbital TPU test (2026)](https://convergedigest.com/google-project-suncatcher-orbital-ai-data-center-tpu/) 2. [Tech Times, Google Project Suncatcher reaches orbit; cooling, not radiation, now defines mission (2026)](https://www.techtimes.com/articles/328482/20261002/google-project-suncatcher-reaches-orbit-cooling-not-radiation-now-defines-mission.htm) 3. [Mixed, Google is sending TPUs to orbit, after a five-year radiation dose at UC Davis (2026)](https://mixed-news.com/en/google-project-suncatcher-tpus-orbit-transporter-18-radiation-test/) 4. [Google, Towards a future space-based, highly scalable AI infrastructure system design, arXiv (2025)](https://arxiv.org/pdf/2511.19468)