Google is preparing to send its Tensor Processing Units into low Earth orbit aboard SpaceX’s Transporter-18 mission, a first hardware test for Project Suncatcher that will examine whether the company’s AI accelerators can survive launch, radiation and the thermal realities of space. The experiment is deliberately modest, but its implications are not: if the hardware works, Google will have taken a concrete step toward testing whether future machine-learning infrastructure could be distributed across constellations of solar-powered satellites rather than confined entirely to terrestrial data centres.
A moonshot moves from modelling to orbit
Project Suncatcher was introduced by Google in 2025 as a long-term research effort rather than a commercial product roadmap. The central idea is straightforward to describe and exceptionally difficult to engineer: place specialised AI computing hardware on satellites with large solar arrays, connect those satellites through high-speed optical links, and use formation flying to make many spacecraft behave more like a tightly coupled distributed computer than a conventional communications constellation. Until now, most of the evidence supporting that concept has come from laboratory tests, mathematical modelling and a preprint paper produced by Google researchers. The coming orbital test changes the nature of the project because hardware will have to cope with the physical environment it is meant to inhabit.
Google said on September 24 that a prototype satellite built with Planet is scheduled to fly on SpaceX’s Transporter-18 rideshare mission. SpaceX currently targets October 1 for launch from Vandenberg Space Force Base in California. The mission is not an orbital data centre in any meaningful commercial sense, and Google has been careful to describe it as a learning mission. Its purpose is narrower: evaluate how TPUs behave during ascent and in orbit, collect operational data, and identify the kinds of failures that ground testing may not expose. That distinction matters because the broader vision still depends on multiple technologies that have not yet been demonstrated together at the necessary scale.
The new mission also updates the project’s practical timetable. Google’s original Suncatcher announcement described two prototype satellites as an early milestone. The company’s latest account separates the work into phases: the immediate test will focus on putting TPU hardware in orbit, while a later two-satellite experiment planned for 2027 is intended to explore the demanding optical links that would be needed between neighbouring compute satellites. In other words, Google is breaking the problem into engineering layers rather than pretending that one launch will validate the entire architecture.
Why AI infrastructure is looking beyond the ground
The timing is not accidental. The physical footprint of artificial intelligence has become one of the defining constraints of the sector. Training and serving large models requires dense clusters of accelerators, high-capacity networking, continuous electrical power and substantial cooling infrastructure. Those requirements have transformed data-centre planning from a relatively specialised property and network question into a strategic problem involving electricity grids, water, transmission equipment, permitting and national industrial policy.
The International Energy Agency estimates that global data-centre electricity consumption rose sharply in 2025 and remains on course to roughly double by 2030, reaching about 950 terawatt-hours in its central outlook. AI-focused facilities are expected to grow faster than the sector overall. Those global numbers are significant, but the more immediate problem is geographic concentration. A hyperscale AI campus can demand electricity on the scale of a large industrial facility, yet such projects are often clustered in regions where grid connections, transformers, generation capacity and transmission expansion already face long lead times.
Space does not make energy free, but it changes the geometry of the problem. In a suitable orbit, solar panels can receive sunlight for much longer portions of each day than panels on the ground and avoid the atmospheric and weather losses that reduce terrestrial output. Google says a panel in the right orbit can be up to eight times more productive than an equivalent panel on Earth. That does not mean eight times more usable computing power after all system losses are included, but it explains why the company sees continuous solar exposure as a potentially valuable resource for very large future compute systems.
The premise is therefore less about escaping Earth than about testing whether some of the hardest constraints on terrestrial AI infrastructure could be traded for a different set of constraints in orbit. The grid connection is replaced by launch. Cooling towers are replaced by radiators. Fibre links are replaced, at least inside a cluster, by lasers. Maintenance crews are replaced by redundancy and fault tolerance. Land-use and local permitting challenges give way to orbital traffic management, debris mitigation and launch economics. None of those substitutions is automatically favourable. Suncatcher exists to test whether enough of them could eventually become favourable at the same time.
The first test: can ordinary AI hardware survive space?
The near-term engineering question is whether hardware designed for data centres can function in an environment for which it was not originally built. Space electronics are traditionally selected, shielded and qualified with radiation tolerance and extreme reliability in mind. State-of-the-art AI accelerators, by contrast, are designed around performance, energy efficiency, memory bandwidth and manufacturing scale. Suncatcher is effectively asking whether commercially derived accelerator technology can be adapted for orbital use without losing the economic and performance advantages that made it attractive in the first place.
Launch is the first obstacle. Google says the spacecraft may experience sustained acceleration of around 10 times Earth gravity during ascent, while individual components can see substantially higher local loads. The company has already subjected hardware to vibration tests across multiple axes to reproduce the frequencies and mechanical stress of launch. According to Google’s latest update, the hardware survived those tests. That is useful evidence, but the actual rocket environment will add a complex mix of acoustic loading, vibration, temperature change and acceleration that laboratory rigs can only approximate.
Once in orbit, radiation becomes a more persistent threat. High-energy particles can alter data in memory or logic, causing transient bit flips, while accumulated radiation can gradually degrade electronic components. Google tested Trillium TPUs at the Crocker Nuclear Laboratory at the University of California, Davis, using a proton beam while the chips were running AI workloads. The company says the devices tolerated a total ionising dose greater than the amount expected during a five-year mission and that the tests were used to characterise error behaviour rather than merely determine whether the chips switched on afterward.
That result is encouraging but should not be mistaken for orbital qualification. Radiation in space is variable, includes different particle types and energies, and can interact with electronics in ways that depend on shielding, memory technologies and workload. A processor that survives a controlled proton test may still encounter rare faults in orbit that matter at scale. For AI infrastructure, the problem is especially interesting because a system with thousands of accelerators can be statistically exposed to failures much more often than a small spacecraft carrying a handful of processors. Resilience therefore has to exist at the chip, system and software levels.
Heat is the problem that makes space counterintuitive
Popular descriptions of orbital computing often imply that space is naturally cold and therefore ideal for cooling servers. In engineering terms, that is misleading. A vacuum contains almost no matter to carry heat away through convection. On Earth, data centres can move enormous volumes of air or liquid through equipment and then reject that heat into the surrounding environment. In orbit, heat generated by processors must be conducted away from the chips and ultimately radiated as infrared energy.
Google’s first Suncatcher hardware uses heat pipes and radiators, technologies with long histories in spacecraft, but the power density of modern AI accelerators makes the application demanding. A small satellite carrying several high-performance chips can generate concentrated thermal loads while also cycling between changing illumination and orientation conditions. A thermal design must keep the accelerators within safe operating ranges without using so much mass for radiators, pumps, shielding and structure that launch becomes uneconomic.
The company says it has tested its cooling approach in a thermal-vacuum chamber, reproducing some of the temperature and pressure conditions that the hardware will encounter in space. The orbital mission will provide the more valuable data: how quickly the system heats under real workloads, how effectively heat moves through the structure, how radiator performance changes with attitude and sunlight, and whether thermal management limits how long the TPUs can operate at high utilisation.
This may ultimately prove more consequential than radiation. Modern data-centre economics depend on keeping expensive accelerators busy. If an orbital TPU can only run intensive workloads in short bursts before waiting for heat to dissipate, the theoretical advantage of near-continuous solar power would be weakened. Conversely, if a compact passive or semi-passive thermal system can sustain useful duty cycles without excessive mass, it would remove one of the most frequently cited objections to orbital compute.
From one spacecraft to a distributed machine
Even a successful single-satellite experiment would leave the central architectural problem unresolved: one spacecraft cannot provide anything like the compute capacity of a major terrestrial AI cluster. Google’s research therefore imagines tightly grouped satellites that share work across optical links. The 2025 Suncatcher paper used an illustrative formation of 81 satellites within a radius of roughly one kilometre to explore how such a system might be controlled and connected.
The attraction of a modular constellation is scale. Instead of launching one enormous orbital computer, an operator could theoretically add capacity by launching more standardised compute satellites. Faults could be handled by routing workloads around failed units. Generational upgrades could arrive incrementally. Manufacturing could resemble high-volume satellite production rather than the bespoke engineering of a traditional flagship spacecraft. This philosophy mirrors how terrestrial cloud infrastructure has evolved around replaceable servers and distributed software, but orbital mechanics adds a layer of complexity that does not exist inside a data hall.
Satellites in a compute cluster would need to maintain very precise relative positions while all of them move around Earth at orbital velocity. Small differences in drag, gravitational forces and solar radiation pressure can gradually separate spacecraft. Repeated station-keeping consumes propellant and changes the economics of mission life. Google’s paper models formation control and argues that clustered operation is plausible, but the real system would have to demonstrate years of safe, predictable manoeuvring without creating unacceptable collision risks.
The proposed architecture also assumes that each future satellite could carry dozens of TPUs rather than the limited hardware of the first test. Packing that amount of compute into a spacecraft magnifies every related challenge: power generation, radiation shielding, heat rejection, structural mass and networking. The current mission is therefore best understood as one small experiment at the bottom of a very steep scaling curve.
Laser links are the nervous system of the concept
A distributed AI cluster is only as useful as the network connecting its processors. Large model training and many forms of distributed inference require accelerators to exchange data rapidly and predictably. Terrestrial AI systems use specialised electrical and optical networking inside data centres, often with enormous aggregate bandwidth. A satellite cluster cannot rely on radio links if it is expected to behave like a tightly coupled computing fabric.
Project Suncatcher therefore depends on free-space optical communication. Lasers can provide high bandwidth without physical fibre, but the pointing problem is difficult. Google describes the requirement as analogous to aiming at a very small target from kilometres away while both endpoints are moving. The company’s earlier research reported high-bandwidth bench demonstrations, but an orbital link must also contend with vibration, thermal distortion, relative motion and the need to acquire and maintain lock between spacecraft.
Google says it plans to test this part of the system with two satellites in 2027. That experiment may be more decisive for the long-term vision than the first TPU mission. If the company can establish stable, high-capacity optical links between close-flying spacecraft, it will have evidence that separate satellites can function as elements of a broader machine. If the links are unreliable or require excessive pointing hardware and power, the architecture may need to be redesigned around more independent workloads.
Latency is another reason Google is interested in close formation. Light travels quickly, but distributed AI systems can become sensitive to even small communication delays when thousands of synchronisation events accumulate. Keeping satellites hundreds of metres rather than hundreds of kilometres apart reduces propagation delay and makes high-bandwidth short-range optical networking more plausible. It also makes collision avoidance and formation management more demanding. Suncatcher’s core trade-off is visible again: technical advantages in one subsystem create new burdens in another.
The economics hinge on launch costs that do not exist yet
The largest uncertainty may be financial rather than scientific. Google’s 2025 paper explicitly treats launch cost as a gating factor. Its learning-curve analysis suggests that low-Earth-orbit launch prices could fall to around or below $200 per kilogram by the mid-2030s under optimistic assumptions about reusable launch systems and flight rates. At that level, the researchers argue, the annualised launch cost associated with delivering power-generating spacecraft could begin to approach the range of terrestrial electricity costs for machine-learning data centres on a per-kilowatt basis.
That is a scenario, not a current market price. Launch today is still dramatically more expensive, and the $200-per-kilogram figure depends on high utilisation of future launch vehicles, continued reusability gains and a large commercial market capable of supporting frequent flights. The calculation also cannot by itself capture the cost of spacecraft manufacturing, insurance, replacement launches, ground infrastructure, optical terminals, mission control or the financial value of hardware that becomes obsolete while still in orbit.
Amazon Web Services chief executive Matt Garman expressed scepticism about orbital data centres earlier this year, saying the concept remained far from economic reality because launch costs and available rocket capacity were not yet sufficient. That scepticism is important precisely because it comes from another company deeply exposed to the economics of large-scale cloud infrastructure. Google’s experiment does not resolve the disagreement; it begins collecting the engineering data needed to make the debate less theoretical.
There is also a technology-refresh problem. Terrestrial accelerators can be swapped when a new generation offers better performance per watt. An orbital accelerator is much harder to service. A constellation designed around a five-year spacecraft life could spend much of that period competing with newer ground hardware. The space system would therefore need strong advantages in energy access, deployment scale or specialised workloads to offset the risk of technological ageing.
What kind of AI would actually run in orbit?
Project Suncatcher’s public materials focus on infrastructure rather than specific commercial workloads, leaving a critical question open: what computations make sense to perform in space? Training frontier AI models requires enormous datasets and intense communication among accelerators. Moving training data from Earth to orbit would add an uplink burden, while returning model checkpoints and results would require reliable high-capacity downlinks. Cloud inference, meanwhile, is often latency-sensitive because users expect rapid responses.
Those constraints suggest that the earliest useful applications, if orbital compute becomes viable, may not mirror ordinary terrestrial cloud services. Space-based systems could have advantages for processing data that is already generated in orbit, including Earth-observation imagery, communications traffic or scientific measurements. Instead of transmitting every raw dataset to the ground, satellites could analyse information locally and send back only high-value results. That is already an area of interest in edge computing, although it is much less demanding than Google’s vision of large distributed AI clusters.
Another possibility is that orbital infrastructure could run workloads tolerant of longer communication paths and delayed results. Large batch jobs, model evaluation, synthetic-data generation or scientific simulations may fit better than interactive consumer services. This remains analysis rather than a stated Google product plan. The important point is that energy availability alone will not determine whether space is competitive; workload placement, data gravity and network architecture will matter just as much.
Google’s advantage is that it controls both the accelerator architecture and a global cloud software stack. If Suncatcher progresses, the company can experiment with hardware, compilers, scheduling and fault tolerance as one system rather than treating the satellite as a generic server rack. That vertical integration could prove essential in an environment where every watt, kilogram and retry matters.
An orbit full of computers would create new external costs
Any serious assessment of orbital data centres must include the space environment itself. Low Earth orbit is already becoming more crowded as communications and Earth-observation constellations expand. The European Space Agency’s 2026 Space Environment Report warns that current trends are increasing long-term collision risk and that end-of-life disposal remains inconsistent. ESA says active debris removal will be required to prevent a self-sustaining growth in fragments caused by collisions.
A compute constellation consisting of tens, hundreds or eventually thousands of satellites would therefore need a credible debris-mitigation and disposal strategy. Close formation introduces additional operational complexity because spacecraft must avoid one another while also responding to conjunction warnings involving unrelated satellites and debris. A failed compute satellite cannot simply become an inert server left in place; it remains a physical object travelling at several kilometres per second.
Astronomers have also raised concerns about the brightness and radio effects of large satellite constellations. Orbital computing may use optical inter-satellite links rather than conventional radio for much of its internal traffic, but the spacecraft would still reflect sunlight and require communications with the ground. Environmental questions would extend to launch emissions, manufacturing, re-entry and the fate of spacecraft materials in the upper atmosphere.
These issues do not make the concept impossible, but they change what counts as success. An engineering demonstration can prove that a TPU works in orbit without proving that a million-accelerator orbital infrastructure system would be environmentally or operationally acceptable. The larger the vision becomes, the more its feasibility depends on regulation and public policy as well as physics.
The terrestrial alternative keeps improving too
Suncatcher is racing against a moving target. Ground-based data centres are not static systems waiting to be displaced. Chipmakers continue to improve performance per watt, liquid cooling is becoming more sophisticated, and cloud operators are signing long-term agreements for renewable energy, nuclear power and emerging geothermal resources. Grid operators and governments are accelerating transmission and generation projects in regions where data-centre demand is strongest.
The IEA’s latest outlook underscores this competition between growth and efficiency. Data-centre electricity use is rising rapidly, but the agency also describes substantial uncertainty because hardware and software efficiency can improve at extraordinary speed. A future model may require far more total compute than today’s systems while using much less energy for each individual task. If efficiency gains outpace workload growth, some of the pressure motivating orbital infrastructure could ease.
Terrestrial facilities also have advantages that are difficult to price in a simple energy comparison. Technicians can repair them. Network connectivity is abundant. Failed components can be replaced. New accelerators can be installed without a rocket launch. Physical security, fire suppression and redundancy are well understood. A space system must offer enough benefit to overcome all of that operational convenience.
Google is not arguing that terrestrial data centres are about to disappear. Its own language presents Suncatcher as a research moonshot designed to explore whether space could eventually become one additional layer of computing infrastructure. That framing is more credible than portraying the October mission as the start of an immediate migration of AI to orbit.
Why Planet and SpaceX matter to the experiment
Project Suncatcher also illustrates how the commercial space industry has changed the threshold for experimentation. Google does not need to develop its own rocket or become a satellite manufacturer from scratch. Planet brings experience building and operating spacecraft at scale, while SpaceX’s rideshare programme allows a relatively small payload to reach orbit as part of a larger mission. That modular ecosystem makes it possible for a technology company to test a specialised computing concept without funding an entire launch stack.
SpaceX’s Transporter missions have become a regular path to orbit for commercial and research payloads, compressing the time between laboratory development and flight testing. For Suncatcher, that means the company can learn from a real orbital environment sooner than would have been practical in an era of infrequent bespoke launches. The same trend also intensifies congestion, because lower barriers to launch encourage more experiments and constellations.
Planet’s involvement is equally significant. Operating large numbers of small Earth-observation satellites has given the company experience in power budgets, thermal management, ground communication and fleet operations. Google’s first objective is not to invent every spacecraft subsystem; it is to learn whether its compute hardware can be integrated into a proven commercial satellite development process. That division of labour may be a model for future orbital computing projects if the field expands.
A strategic experiment in the AI infrastructure race
The project also has a competitive dimension. Google’s TPUs are one of the few large-scale alternatives to Nvidia’s GPUs for advanced AI workloads. They are central to Google’s internal systems and available to customers through Google Cloud. Demonstrating that TPUs can operate beyond conventional data centres would not create a near-term revenue stream, but it would reinforce the argument that Google controls a vertically integrated computing platform spanning chip design, software, networking and infrastructure research.
Other companies are examining orbital computing as well, and the interest extends beyond Google. Space companies, semiconductor suppliers and cloud providers increasingly discuss the possibility of moving specialised workloads to orbit. STMicroelectronics has cited orbital data centres as a potential future market for space-grade chips, while AWS has publicly questioned whether the economics are remotely competitive today. The disagreement is itself evidence that the concept has moved from science-fiction language into strategic technology planning.
For governments, the prospect raises additional questions. Large-scale AI infrastructure is increasingly treated as a strategic asset because it affects national security, scientific capability and industrial competitiveness. If meaningful compute capacity eventually sits in orbit, jurisdiction, export controls and physical protection of satellites could become major policy issues. A system built to escape terrestrial power constraints would not escape geopolitics.
What October’s launch can actually prove
The October 1 mission has a narrow but useful test agenda. First, it can confirm whether the TPU hardware and supporting electronics survive the launch environment. Second, it can measure how the processors respond to real radiation exposure rather than accelerated laboratory tests. Third, it can validate thermal behaviour under vacuum and changing orbital conditions. Fourth, it can expose operational problems involving power, firmware, telemetry and remote workload control that may not appear in a ground laboratory.
A successful result would not show that orbital AI data centres are economical, environmentally sustainable or competitive with terrestrial systems. It would show something more basic: that modern high-performance AI accelerators can function as useful spacecraft payloads without requiring such extensive modification that their performance advantage disappears. That is the prerequisite on which every larger Suncatcher claim rests.
A failure would also be informative. If radiation produces unacceptable errors, Google can investigate shielding, error-correcting memory or software redundancy. If thermal limits are severe, the company can redesign heat paths or reduce power density. If launch vibration causes damage, packaging can change. Research missions are valuable because they identify where a concept breaks before capital is committed at commercial scale.
The harder test comes in 2027
The planned two-satellite optical experiment in 2027 will move Suncatcher closer to its defining challenge: making separate spacecraft operate as one computing system. Stable laser links, precise formation control and distributed workload scheduling would transform the project from a demonstration of space-capable AI chips into an experiment in orbital cluster computing.
That phase will need to show more than a connection. Engineers will want to understand link availability, error rates, acquisition time, bandwidth consistency and the propellant cost of maintaining the geometry required for optical communication. They will also need to test how computing workloads respond when a link drops or a spacecraft moves out of position. Terrestrial distributed systems assume networks will fail occasionally; an orbital system must make those failures routine and survivable.
If the 2027 work succeeds, the next questions become much larger: how many satellites can share one formation, how clusters communicate with Earth, how hardware generations are replaced, how debris rules are satisfied, and whether launch economics can support a business case. Each step adds another discipline to the project. That is why Suncatcher is better understood as a systems-engineering programme than as a satellite product.
A real experiment, not yet a revolution
Google’s decision to put TPUs into orbit is significant because it crosses the boundary between conceptual advocacy and empirical testing. The company is no longer only publishing models of future space-based compute. It is preparing to expose the hardware to the environment that will decide whether those models have practical value.
At the same time, almost every transformative claim remains conditional. Solar energy in orbit is abundant, but collecting it requires large, lightweight spacecraft. Space is free of clouds, but it is also a vacuum that makes heat rejection difficult. Laser links can deliver enormous bandwidth, but they demand extraordinary pointing precision. Modular satellites could scale capacity, but they increase traffic and debris concerns. Falling launch prices could transform the economics, but the prices assumed in long-term scenarios have not yet been achieved.
That combination of promise and uncertainty is what makes the coming flight worth watching. The most important result will not be whether a TPU turns on above Earth. It will be the quality of the engineering data Google brings back about how accelerator hardware behaves when launch forces, radiation, heat, power and remote operations all arrive at once.
For an AI industry accustomed to measuring progress in larger models and faster chips, Project Suncatcher introduces a different metric: whether computing can scale by changing where infrastructure lives. The October mission will not answer that question, but it will provide the first orbital evidence from Google’s own hardware. In a sector where demand for compute is colliding with the physical limits of grids, cooling systems and construction timelines, even a small satellite carrying AI accelerators is a meaningful experiment in what the next layer of digital infrastructure might become.
