Google Suncatcher Awaits 2027 as Starcloud Computes in Orbit

Google’s Project Suncatcher remains a research program rather than an operating orbital data center. Its next public milestone is two Planet-built prototype satellites targeted for early 2027, while the larger TPU constellation remains a modeled architecture without a disclosed deployment schedule.

The position has changed since the project emerged in November 2025: Starcloud has now operated data-center-class AI hardware in orbit. The company’s Starcloud-1 mission record states that its H100-equipped satellite launched that month, ran a version of Gemini in December and trained Andrej Karpathy’s nanoGPT model. That is a genuine flight result, but it is not evidence of a complete orbital data center.

Suncatcher is testing an architecture, not opening a cloud region

Project Suncatcher’s central idea is to divide machine-learning computation among compact, solar-powered satellites carrying Google Tensor Processing Units. Free-space optical links would connect the spacecraft, allowing accelerators on separate satellites to exchange the large volumes of data required by distributed AI workloads.

Google Research’s technical account describes an illustrative formation of 81 satellites at an altitude of 650 kilometers, arranged within a cluster one kilometer in radius and separated by roughly 100–200 meters. Its laboratory demonstrator transmitted 800 gigabits per second in each direction through one optical transceiver pair, or 1.6 terabits per second in aggregate. A separate test exposed a Trillium v6e TPU to a 67 MeV proton beam: memory irregularities began after 2 krad, while no hard failure was attributed to total ionizing dose through the maximum tested level of 15 krad.

Those results address individual risks under controlled conditions. They do not yet show that moving spacecraft can maintain the same optical capacity, that a complete satellite can operate reliably for years or that dozens of orbital accelerators can coordinate a large training job. The distinction is essential: Suncatcher currently supplies engineering evidence for several components, not an end-to-end demonstration.

The first mission is deliberately limited to two spacecraft

The planned learning mission reduces the problem to its smallest useful orbital test. Two satellites can examine TPU behavior, close-formation operations and a high-bandwidth cross-link without requiring Google and Planet to deploy the mass, networking complexity or collision-management systems of a full cluster.

The prototypes are also a narrow test of scalability. A stable optical connection between two spacecraft would validate one edge of the proposed network, but an 81-node system would require routing, workload scheduling and fault handling across many simultaneous links. Likewise, controlling one pair cannot establish that a dense formation will remain safe when failures or propulsion differences affect multiple satellites.

The early-2027 target remains a target, not a completed launch or guaranteed service date. The public mission information does not identify a subsequent constellation launch, customer-access program or timetable for commercial operation. Calling Suncatcher an orbital data center is therefore accurate only as shorthand for its long-term objective.

Starcloud has flight experience, not a server farm

Starcloud chose a more direct initial experiment: place a familiar data-center accelerator aboard one spacecraft and execute AI workloads after launch. That approach produces a simpler proof point than Google’s distributed design because it does not initially depend on a tightly synchronized network of satellites.

Operating an H100 in space answers an important component-level question. It shows that a powerful terrestrial accelerator can perform useful computation in orbit, and training nanoGPT goes beyond merely switching the device on. The result does not establish sustained multi-GPU training, commercial availability or reliable service over the lifetime expected from infrastructure.

A single-node demonstration also leaves the hardest system questions open. Large models require accelerators to exchange parameters quickly; customers need dependable ground communications; spacecraft must reject heat; and failed hardware cannot be replaced as routinely as equipment in a terrestrial facility. Starcloud is ahead on flight operation, while Suncatcher is aimed at the interconnect and formation architecture required to move beyond one isolated computer.

Why the proposed satellites must remain unusually close

The tight formation is a consequence of the networking requirement rather than a visual flourish. Distributed machine-learning jobs repeatedly move data between accelerators, while received optical power falls sharply as separation increases. Placing satellites hundreds of meters apart helps close the link budget needed for data-center-like communication.

That proximity introduces an operational trade-off. Differences in Earth’s gravity field, atmospheric drag and spacecraft behavior can gradually change relative positions, so the formation needs precise navigation and station-keeping. A malfunctioning satellite would also become a nearby physical hazard rather than merely a lost network node.

Even a successful two-satellite cross-link would leave scale unproven. Larger formations would add many more communication paths, possible interference points and failure combinations. Suncatcher’s first flight should consequently be judged as a test of foundational assumptions, not as a small production version of the proposed infrastructure.

Solar power improves the premise but does not solve cooling

Near-continuous sunlight is the main physical attraction of the concept. In the intended orbital conditions, solar panels could produce substantially more energy than typical terrestrial installations while requiring less battery capacity. That could reduce direct pressure on land and electricity grids if orbital compute ever reaches meaningful scale.

Space does not provide effortless cooling. Without surrounding air or water to carry heat away, a spacecraft must radiate waste heat, requiring radiator area and adding mass. More accelerators demand more electrical power and create more heat, so power collection, computing hardware and thermal management must be designed as one system.

Communication with Earth creates another constraint. Processing data near the spacecraft that generated it may reduce raw-data downlinks, but terrestrial AI users would still need reliable, high-capacity links for workloads and results. The proposed inter-satellite network solves communication inside the cluster; it does not by itself solve the connection between that cluster and customers on the ground.

The economics depend on a launch-price forecast

The published feasibility model requires launch prices to fall below $200 per kilogram by the mid-2030s before orbital compute could approach the reported annual energy cost of an equivalent terrestrial data center. That comparison is a modeled threshold, not a current launch quote or evidence that the total cost of ownership has reached parity.

Launch is only one part of the bill. Satellite production, replacement missions, ground stations, operations, communications capacity and losses would all affect the commercial calculation. Any advantage from abundant solar energy must be large enough to offset those costs and the inability to repair hardware conventionally.

The projects are proving different parts of the same thesis

The contest is no longer between a theoretical Google swarm and an unopened Starcloud payload. Starcloud has produced an early orbital-compute result, while Google has laid out a more complex route toward networked accelerators and assigned its first orbital experiment to a two-satellite mission.

Neither project has demonstrated the system implied by “orbital data center.” Starcloud has proved that one powerful GPU can execute AI work in space; Suncatcher has produced laboratory and modeling results for a distributed design but has not yet flown its prototypes. The next decisive evidence for Google will be whether the planned satellites reach orbit, operate their TPUs and sustain their optical cross-link.

Google’s Orbital AI Swarm Awaits 2027 as Starcloud Computes in Space