What did Google's Suncatcher satellite just do?
Google's first orbital AI test is now producing output. After roughly three weeks of commissioning, the prototype satellite in Project Suncatcher received its first prompt this week and answered it, TechCrunch reported on 30 October 2026.
The prompt asked the model loaded on the spacecraft, Google DeepMind's Gemma 4, to write a haiku about space. The reply opened with “Blue marble below” and closed with “Starlight feeds my core”. A Google spokesperson told TechCrunch that the satellite was operating nominally.
The spacecraft launched on 1 October 2026 on a SpaceX Falcon 9, flying as part of the Transporter-18 rideshare mission with more than 100 other payloads aboard. Planet Labs built the satellite on one of its standard platforms, with Google supplying the compute payload. Reporting by The New York Times, cited by Data Center Dynamics, named the craft MVP and put its hardware at four Trillium TPUs powered by about 1 kW of solar electricity.
That is a tiny machine by data centre standards. SpaceX has described its proposed AI1 satellite as delivering 150 kW at peak and 120 kW of sustained compute. Google's prototype is therefore a physics and reliability experiment, not a commercial service. Its value lies in the data it sends home about how AI silicon behaves outside the atmosphere.
For Google, the milestone is less about the poem than about the telemetry behind it. The company has spent a year arguing that its TPUs can be made radiation-tolerant enough for orbit. A chip that boots, runs a model and returns coherent output after three weeks in space is the first public evidence that the argument holds outside a test chamber.
Why is running AI in orbit so difficult?
Power is the first constraint. A solar panel in the right orbit can be up to eight times more productive than one on Earth, according to Google Research, because a dawn-to-dusk sun-synchronous orbit keeps a satellite in near-constant sunlight. That reduces the need for heavy batteries. It does not remove the ceiling on how much power a small spacecraft can collect and dissipate.
Cooling is harder than power. In a vacuum there is no air to carry heat away by convection, so heat must be radiated into space. Google says its team is developing a combination of heat pipes and radiators, and has tested the design in a thermal vacuum chamber that simulates both the cold and the vacuum of orbit. Whether that works in sunlight, on a satellite whose chip is the hottest object on board, is one of the questions the mission is meant to answer.
Radiation is the third obstacle. Cosmic rays and solar events can flip bits in memory, corrupting data without setting off the usual monitoring tools. Google tested its Trillium chips in a 67 MeV proton beam at the Crocker Nuclear Laboratory at UC Davis. High-bandwidth memory, the most sensitive subsystem, began showing irregularities only after a cumulative dose of 2 krad(Si), nearly three times the 750 rad(Si) expected behind shielding over a five-year mission. No hard failures were recorded up to the maximum tested dose of 15 krad(Si).
Launch itself is a trial. The ride to low Earth orbit lasts about ten minutes, with sustained acceleration up to 10 g and individual components seeing 50 to 100 g, according to Google. The team shook the satellite on all three axes to mimic launch vibrations. “Tests like this rarely go as planned, so we were pleasantly surprised that the hardware held up to the force,” wrote Travis Beals, who leads the project.
Even with all of that solved, the chip cannot run continuously. Power generation and heat rejection together cap how long the TPU can operate, so the team fires it in short bursts rather than around the clock. That complicates the already difficult task of scheduling work and getting results back to the ground.
Getting results back is a constraint in its own right. Google's research lists high-bandwidth ground communications as an unresolved engineering challenge. A satellite is only in range of a given ground station for part of each orbit, so a service would need either many stations or a relay architecture that does not exist yet. Pre-loading the model, as Google did with Gemma 4, avoids streaming weights down a limited link.
How big is Project Suncatcher meant to become?
The long-term plan is not one satellite but a tightly packed cluster of them. Google's research paper models an 81-satellite constellation at a mean altitude of 650 km, with the whole cluster inside a radius of 1 km. Next-nearest neighbours would oscillate between roughly 100 and 200 metres apart under the influence of Earth's gravity.
That closeness is the point. Google argues that running large machine-learning workloads across many chips needs links in the tens of terabits per second, far beyond conventional long-range satellite communications. Flying satellites within a kilometre of each other shortens the distance the light has to travel and closes the link budget. The team has demonstrated 800 Gbps each way, 1.6 Tbps in total, across a single pair of transceivers on a bench.
Keeping those links alive requires extraordinary pointing precision. Google compares it to hitting a coin-sized target from miles away while both points are moving. The company plans to test the idea in 2027, when it intends to fly two satellites equipped with optical inter-satellite links and more power than the current prototype.
Future satellites in the design would each carry dozens of TPU chips, not four. Google also describes eventual gigawatt-scale constellations that might need a more radical mechanical design, with solar collection, compute and heat rejection integrated rather than bolted together. None of that exists today. The current mission is the first data point on whether the pieces can survive.
The formation is designed to stay put. Google's models suggest that with satellites a few hundred metres apart, station-keeping will need only modest manoeuvres to hold the cluster inside its sun-synchronous orbit. That matters because fuel is finite, and a constellation that drifts apart cannot share work over laser links.
Can orbital data centers compete on cost?
Google's economic case rests on launch prices falling far below today's levels. Its research paper projects that with a sustained learning rate, launch prices could drop below $200 per kg by the mid-2030s. At that point, Google estimates, launching and operating a space-based data centre could become roughly comparable to the reported energy costs of an equivalent terrestrial facility on a per-kilowatt-year basis.
That projection depends heavily on SpaceX. TechCrunch reported that a peer-reviewed version of Google's white paper, due in the journal Joule, estimates SpaceX would need to fly Starship about 1,800 times to achieve the same learning-driven cost reductions it saw with its earlier vehicles. Alphabet, Google's parent company, is also a major investor in SpaceX. Starship is still in development and is not yet fully and rapidly reusable.
The researchers are careful about what the paper claims. TechCrunch noted that they stress the analysis is not an economic feasibility study. It is a cost model built on historical launch pricing and projected learning rates, not a demonstration that customers will pay for compute in orbit.
The comparison that matters is with ground-based facilities, which benefit from cheap land, existing power contracts, fibre links and trained staff. Google's own pitch is that space offers near-continuous solar power and avoids competing for terrestrial electricity and water. Whether that advantage survives the cost of launch, insurance, ground stations and replacement satellites is unresolved.
What would change the picture is cadence. Google's cost case assumes launch becomes routine, cheap and reliable enough to replace satellites on a schedule. If Starship's flight rate stalls, or if terrestrial operators keep cutting the cost of a kilowatt-hour of compute, the crossover point moves further into the future. Peer review tests the method and the assumptions, not the projections themselves.
Who else is racing Google to put AI in orbit?
Google is not first. Starcloud, a Washington-based startup backed by Nvidia, launched its Starcloud-1 satellite in November 2025 carrying an Nvidia H100 GPU. CNBC reported in December 2025 that the company had trained NanoGPT, a small language model created by Andrej Karpathy, on the complete works of Shakespeare while in orbit, and was also running Google's Gemma model on the same chip.
CNBC described the H100 as 100 times more powerful than any GPU compute previously sent to space. Starcloud's chief executive, Philip Johnston, framed the goal in terms of terrestrial limits. “Anything you can do in a terrestrial data center, I'm expecting to be able to be done in space,” he told CNBC, citing energy constraints on the ground.
SpaceX is pursuing a larger design. Its AI1 satellite, which the company calls an orbital data centre, is described by Data Center Dynamics as targeting 150 kW at peak and 120 kW sustained. Google's approach differs in architecture: it wants many satellites flying close together and sharing work over laser links, while companies such as Starcloud and Cowboy Space expect most inference to happen on individual satellites. Cowboy Space and Satlyt both had payloads on the same Transporter-18 launch as Google's prototype.
China is also moving. SpaceNews reported in April 2026 that a Beijing-based space startup had secured early-stage funding and $8.4 billion in credit lines as part of a national push toward space-based computing. The competition is therefore not only about chips, but about launch cadence, capital and industrial policy.
Blue Origin has also told US regulators it wants to operate orbital compute clusters, according to Data Center Dynamics. The result is a crowded field in which Google's differentiator is not speed but architecture: a cluster designed for distributed workloads rather than inference on a single spacecraft.
What does this demo not prove?
It does not prove that orbital data centres work. One satellite running inference on a pre-loaded model is a long way from a cluster training or serving models at scale. There are no optical links in orbit yet, no multi-satellite scheduling and no customer workload.
The engineering evidence is also early. Ground tests suggest radiation effects will be manageable for inference and more of a problem for training, but the team now needs real orbital data on bit flips, thermal stress and degradation over time. A single mission provides one sample, not a reliability curve.
The public record has gaps. Reports differ on the chip count: launch-week coverage citing The New York Times said four TPUs, while TechCrunch described the spacecraft as carrying a single TPU. Google's public posts on the mission do not list a chip count, so anyone quoting an exact configuration should check which report they are relying on.
There is also the question of useful work. Running a chatbot prompt proves the chip and the model function together. It does not show throughput, latency or cost per token under a realistic load, which are the numbers that would decide whether anyone buys compute in orbit.
The economics are unproven in both directions. Google's cost case assumes launch prices that do not exist yet, on a rocket that is still being tested. It also assumes demand for compute in orbit that has not materialised. Companies on the ground are adding capacity quickly, and terrestrial data centres remain far cheaper per unit of compute today.
Finally, the demonstration says nothing about the operational questions that would shape a real constellation: how satellites are deorbited at end of life, how much ground-station bandwidth is available, and how debris risk scales with thousands of closely flying spacecraft. Those issues sit outside the scope of a one-satellite test.
What happens next for Suncatcher?
The immediate next step is the 2027 mission, when Google plans to fly two satellites with higher power generation and optical inter-satellite links. That demonstration will test whether the cluster architecture works at all, not just whether a chip survives.
Before then, watch for technical results from the current mission. The most useful signals would be published data on bit-flip rates in orbit, how the radiator design performs across sunlit and shadowed passes, and whether the TPU can sustain useful throughput within its power budget.
Watch Starship as well. Every cost projection in Google's paper depends on launch prices falling by an order of magnitude, and that depends on a vehicle that has not yet settled into routine reuse. Until then, Suncatcher remains what Google calls it: a long-term research project aimed at infrastructure that might be economical in the 2030s.
Google's timeline is deliberately unambitious by industry standards. The company frames the work as infrastructure for the 2030s rather than a product for this decade. That gives it room to publish failures as well as successes, and it means the current mission should be judged on the quality of the data it returns rather than on whether it looks impressive.