Google put its first TPUs in orbit, but the data center is still on Earth

THE SIGNAL IN ONE SENTENCE

The newest piece of Google AI infrastructure is not answering prompts. It is circling Earth and trying not to cook itself. Google confirmed on October 1 that its Project Suncatcher prototype satellite launched aboard SpaceX's Transporter-18 rideshare mission. Planet built the spacecraft with Google. The company says it has established contact and the satellite is operating as expected. That is the verified event. The next part needs a much larger asterisk. Project Suncatcher is Google's research plan for putting machine-learning hardware on solar-powered satellites and eventually linking many satellites with optical connections. The long-range concept is an orbital computing system that can draw near-continuous solar energy in a suitable low-Earth orbit. The launch does not demonstrate that system. It begins a smaller experiment: expose Google's Tensor Processing Units, or TPUs, to the physical stress, radiation and thermal conditions of space, then measure what happens. The plain signal is that Google has moved one important question out of the laboratory and into orbit. The company has not moved a data center there. The distinction matters because the attractive part of the proposal is easy to explain. In its original research description, Google says a solar panel in the right orbit can be up to eight times more productive than one on Earth and can generate power nearly continuously. AI computing needs enormous amounts of electricity. Put the computers closer to the sunlight, the argument goes, and perhaps some future infrastructure can scale without placing every new megawatt on a terrestrial grid. Space then sends the bill. A rocket launch shakes hardware violently. Google says the spacecraft can experience sustained acceleration up to ten times gravity, while individual components can see 50 to 100 times gravity. Outside Earth's protective atmosphere, solar events and cosmic radiation can alter bits or damage electronics. Heat is another problem. A TPU concentrates power in a small area, but a vacuum has no air to carry that heat away. The satellite must move heat through pipes and radiate it into space. Google tested parts of this problem on Earth before launch. The company says it shook the satellite along three axes, ran Trillium TPUs in a proton beam at the University of California, Davis, and used a thermal-vacuum chamber to test its cooling system. Its research paper reports that the tested TPUs survived a total ionizing dose equivalent to a five-year mission without permanent failures and were characterized for bit-flip errors. Those tests are promising evidence about selected hardware under selected conditions. They are not a production service-level agreement. Orbit adds combinations, durations and failure modes that a laboratory cannot perfectly reproduce. That is why the current satellite exists. Over the coming weeks, Google says it will collect data on how the TPUs handle spaceflight stress, radiation and thermal extremes. The useful result may be a list of problems. A hot component, a radiation-induced error pattern or a cooling limit would tell engineers where the next design needs margin. A quiet flight would still need enough time and telemetry to show that quiet means robust rather than merely lucky. Even a perfect hardware-survival result would leave the harder system questions on the ground. Large AI workloads do not become useful just because several chips are nearby. Training and serving models require accelerators to exchange data at high bandwidth and low latency. Google's proposed satellites would fly in close formation and communicate through free-space optical links. Its research says data-center-like operation could require tens of terabits per second between satellites. A laboratory demonstration reached 800 gigabits per second in each direction, according to Google's earlier work, but keeping narrow laser links aligned between moving spacecraft is a different job. Google plans a two-satellite test in 2027. That mission is the first public milestone aimed at the networking problem. Until satellites exchange real workload data in orbit, the cluster remains a design, not a cluster. Then there is the workload path. Data must reach orbit, be processed, survive interruptions and return. Sensitive information needs protection in transit and at rest. Operators need a way to schedule jobs, replace failed capacity, update software and prove that a result was not corrupted by radiation. A satellite network also needs ground stations, launch providers, tracking, collision avoidance, licensing and debris plans. The data center may be in the sky, but its supply chain, control room and consequences stay stubbornly terrestrial. Economics are equally unfinished. Google's research argues that falling launch prices could make the concept competitive in the mid-2030s and models launch costs at or below $200 per kilogram. That is a projection built from a learning curve, not a price available today. The comparison must include launch, spacecraft manufacturing, failed missions, replacement cadence, ground links, insurance, operations and the energy used across the whole system. Terrestrial data centers are expensive and resource-intensive, but they are also reachable by trucks and technicians. A failed power supply can be replaced without booking a rocket. An orbital system trades some land, grid and water constraints for launch emissions, material throughput, space-traffic risk and hardware that is difficult or impossible to repair. The environmental case therefore needs a ledger, not a mood board. Count the energy collected in orbit, the energy and material required to manufacture and launch the system, the cooling and ground infrastructure, the expected hardware life, the replacement rate and the disposal path. Compare that with a realistic Earth-based alternative using the same workload and reliability target. A solar panel seeing more sunlight is only the first line of the spreadsheet. There are practical applications if the engineering works. An orbital cluster could run workloads whose power demand is difficult to place on a crowded grid. It could support space-based sensing or scientific instruments without sending every raw observation to Earth first. A modular satellite design could add capacity in increments rather than through one immense facility. Optical networking developed for the project might improve other spacecraft communications even if the orbital data-center vision never arrives. None of those applications has been demonstrated by this launch. That sentence is not pessimism. It is the job description of a prototype. The current mission should be judged by whether Google publishes enough evidence to separate survival from useful computing. How much TPU work runs in orbit? What kinds of errors appear under radiation? How does performance change across thermal cycles? How much power reaches the chips, and how much heat can the radiator reject? What fails first? How does the hardware recover? What telemetry will be public, and what will remain a company summary? The 2027 mission needs another set of questions. What throughput and latency do the two satellites sustain while moving? How often does the optical link drop? How much pointing power and mass does it require? Can a distributed workload make progress across the link, or does the network spend its life recovering alignment? Those measurements would turn a spectacular premise into an engineering record. Independent researchers would also need enough detail to reproduce the assumptions behind cost, energy, orbital formation and reliability claims. Google deserves credit for naming the current launch as an experiment. Its October 1 announcement says the satellite will gather data, refine designs and inform later missions. The company's September preview calls this first launch a hardware-survival test and says future satellites would carry dozens of TPUs. That is more useful than pretending a small prototype is an orbital cloud region. The tempting headline is that AI has gone to space. The more accurate one is that a few of AI's physical constraints finally have nowhere to hide. Radiation, vibration, heat, bandwidth, launch cost and orbital control do not care about a demo reel. They produce measurements. Project Suncatcher now has the first chance to collect them. The data center is still on Earth. The experiment is finally above it.