AI compute has reached orbit—but one haiku is not a data centre

A prompt went up and a haiku came back

Google now says its Project Suncatcher prototype completed an AI inference in orbit. During a ground-station pass, engineers uplinked a prompt to an onboard Gemma 4 open model running on what the company calls a data-centre-class Tensor Processing Unit. As the satellite passed Svalbard, Norway, the team downlinked a generated haiku. Google says it then tried space-themed limericks and prompts asking the payload to report its technical status.

That is a real change from the 1 October launch update, which established contact but reported no completed workload. It verifies a minimal end-to-end path: command, model execution and returned text. Google does not disclose the prompt or model size beyond the Gemma 4 name, token throughput, end-to-end latency, power draw, temperature, memory errors, retries, uptime or the number of unsuccessful attempts. The evidence is also company-reported, without raw telemetry or independent observation.

For people weighing near-term data-centre land, grid and water decisions, the responsible conclusion remains narrow. A short inference can prove that relevant hardware turns on and returns output in orbit. It does not show that high-density accelerators can run continuously, reject heat, survive radiation for years, communicate as a cluster or deliver useful capacity at competitive cost.[5]

The larger claim remains a research programme

Google confirmed on 1 October that a Project Suncatcher prototype built with Planet had launched aboard SpaceX's Transporter-18 mission. The later haiku update now adds one completed inference. The satellite's broader purpose is still to collect data on how a TPU and supporting hardware respond to physical stress, radiation and thermal extremes.

The distinction matters because ‘AI compute has reached orbit’ can mean several different things. A short workload can run; a cluster can communicate; or a commercial service can deliver dependable capacity. Google's update establishes the first step on one prototype. It does not establish the networking, endurance, fault recovery or economics needed for the others.

Google describes Project Suncatcher as long-term research into solar-powered AI infrastructure in space. The prototype is therefore a sensor-rich engineering experiment, not proof that a fleet of orbital data centres is technically, economically or environmentally viable.[1][2][5]

The peer-reviewed paper models an 81-satellite cluster

Alongside the launch, a Google-authored systems paper became available in the peer-reviewed journal Joule. It develops an illustrative low-Earth-orbit architecture rather than reporting a clinical-style trial or a deployed service. All nine listed authors are affiliated with Google, which has an obvious commercial interest in the concept.

The worked example places 81 satellites inside a formation with a one-kilometre radius at a mean altitude of about 650 kilometres. In the model, next-nearest-neighbour distances oscillate between roughly 100 and 200 metres. The authors explore how such a tightly packed formation might keep relative positions while maintaining solar orientation and avoiding unacceptable collision risk.

Those simulations are useful because they expose the engineering problem: data-centre accelerators need not only power, but also frequent, low-latency communication. The paper estimates that a data-centre-scale system could require roughly 10 terabits per second of aggregate inter-satellite capacity. That is a system requirement, not a measured performance result from space.

A formation of dozens of high-value satellites would also have to tolerate failed thrusters, degraded sensors, space weather and conjunctions with other objects. The paper treats formation dynamics, but it does not demonstrate years of safe autonomous operation. A successful single launch cannot supply that evidence.[3][4]

The fastest link and radiation results were obtained on the ground

For communications, the paper reports a bench demonstration of 800 gigabits per second in one direction and 1.6 terabits per second bidirectionally over a short free-space optical path. The authors discuss combining wavelengths and spatial channels to move toward the much larger aggregate requirement. But the experiment was not a link between satellites passing through orbital vibration, pointing error and thermal cycles.

Google's public roadmap calls for a separate two-satellite optical-link experiment in 2027. That planned test is the more relevant milestone for whether very high bandwidth survives real formation flight. Until results are published, the ground link should be read as a component demonstration, not evidence that a distributed orbital accelerator cluster can communicate at data-centre scale.

The radiation evidence is also a ground test. A Trillium v6e TPU was exposed to a 67 megaelectronvolt proton beam. The paper estimates a shielded five-year mission would accumulate about 750 rad(Si) of total ionising dose. One chip showed no hard failure attributed to total ionising dose through 15 kilorad(Si), although its high-bandwidth memory was the most sensitive component and irregularities began after a cumulative two kilorad(Si).

The researchers estimate approximately one uncorrectable memory error per 50 rad under the tested conditions—translated, with their workload assumptions, to roughly one event per ten million inferences. They caution that training could behave differently and that more data are needed. Beam exposure cannot reproduce every particle, manufacturing variation, thermal interaction or single-event effect encountered over years in orbit.[2][3]

Solar abundance does not remove the heat problem

The attraction is straightforward. In a suitable sun-synchronous orbit, solar panels can receive near-continuous sunlight, while land use and local water consumption associated with terrestrial facilities could in principle be reduced. Google says a panel in that setting could receive up to eight times the solar energy of a panel at a mid-latitude ground site.

More available sunlight is not the same as eight times more usable computing. Conversion losses, battery or ride-through needs, radiation degradation, orientation constraints and the mass of power electronics still matter. Above all, high-performance accelerators convert most electrical input into heat, and space has no air or water to carry that heat away.

Radiators must reject heat by infrared emission. They add area and mass, affect attitude control and can face unwanted sunlight or Earth radiation. The prototype can inform component temperatures, but the launch post does not report a sustained, data-centre-level thermal load. Cooling therefore remains one of the project's central unanswered questions rather than an advantage already demonstrated.[2][3]

The cost comparison depends on a steep fall in launch prices

The paper's economic discussion asks when the annualised cost of launching power-generating hardware might approach terrestrial data-centre energy costs. For one illustrative satellite, it estimates about $14,700 per kilowatt per year at a launch price of $3,600 per kilogram, falling to roughly $810 per kilowatt per year at $200 per kilogram. The terrestrial comparison range cited is approximately $570 to $3,000 per kilowatt per year.

The lower result is conditional on launch prices falling below about $200 per kilogram in the mid-2030s. A more optimistic scenario in the paper reaches $60 per kilogram under specific assumptions about future Starship-derived capacity and learning curves. These are projections, not quoted launch prices available today.

Nor is launched power the full cost of compute. Accelerators, radiation shielding, radiators, optical terminals, propulsion, replacement spacecraft, insurance, ground networks, operations, failed launches and de-orbit capability would all affect a commercial comparison. Launch emissions and upper-atmosphere effects also need lifecycle assessment; moving electricity demand off Earth does not make the system impact-free.

A fair comparison would specify equal compute delivered, reliability, latency, lifespan and environmental boundaries. The current paper is most useful as a sensitivity analysis showing where launch cost becomes decisive. It is not a business case proving orbital computing is cheaper.[3]

What the prototype could change for people—and what it cannot yet promise

If the engineering eventually works, space-based capacity could add another location for highly parallel workloads and reduce some pressure on land, electricity grids and cooling water near terrestrial data centres. It might also concentrate supply in the hands of organisations able to finance launch fleets, spectrum, ground stations and specialised hardware.

For communities near proposed data centres, the prototype does not yet justify claims that future AI demand will leave the ground. Terrestrial facilities are being planned now, while a scalable orbital alternative would require many successful technology and regulatory milestones. Grid planners, water authorities and residents should evaluate current projects on current evidence.

There are space-side externalities too. Dense formations would add objects, manoeuvres and failure modes to an already shared orbital environment. Operators would need transparent conjunction procedures, reliable disposal, coordination with regulators and evidence that a failed spacecraft could not endanger neighbours. The public sources do not provide an end-to-end debris and collision-risk assessment for a commercial-scale fleet.

The prototype is nonetheless a meaningful experiment. One returned model response replaces the earlier assumption that the payload could execute an AI workload with a small observed milestone, while future telemetry can test component temperature, radiation effects and mechanical survival. The responsible conclusion is therefore neither dismissal nor a declaration of data-centre success: a company has run a short inference in orbit, while the proposed infrastructure remains unproven.[1][2][3]

What evidence would change the assessment

First, Google should publish the prototype's prespecified tests and denominators: operating hours, TPU workload type, power draw, throughput, temperature range, corrected and uncorrected memory errors, resets and any component degradation. Results should include failures and missing telemetry, not only selected successes.

Second, the planned two-satellite experiment should report link distance, duration, availability, bidirectional throughput, latency, pointing loss and performance through thermal cycles. Independent reproduction or third-party review would strengthen a result produced by the system's developer.

Third, a multi-satellite demonstration must show safe formation control and useful distributed computing, including recovery from a failed node. A credible economic and environmental assessment should count the complete system, launch and replacement cadence, ground infrastructure, atmospheric effects and end-of-life disposal against an equivalent terrestrial service.

Until those milestones arrive, the wording should remain precise. A prototype has completed a small AI inference in orbit. An orbital AI data centre has not been demonstrated.[1][2][3]