Project Suncatcher: Google wants to put its TPUs in orbit — solar satellites for AI compute at 8x terrestrial power

Project Suncatcher: Google wants to put its TPUs in orbit — solar-powered satellites for AI compute at 8x terrestrial power

🔎 AI compute leaves Earth

Google has just crossed a milestone that science fiction never dared to imagine: running artificial intelligence workloads in space. On October 1, 2026, a Project Suncatcher prototype was placed into orbit aboard SpaceX's Transporter-18 mission. The satellite, built in partnership with Planet Labs, is operational. It's the first building block of an infrastructure that could, eventually, move a significant portion of foundation model training and inference out of the atmosphere.

The idea is radical: in sun-synchronous orbit at 650 km, solar panels capture up to 8 times more annual energy than on the ground. No night, no clouds, no seasons. Energy becomes virtually infinite and free — the real bottleneck for AI compute today. Google is betting that launch costs, which have been in free fall for a decade, will make the economic equation viable by the mid-2030s.

The Essentials

• First prototype in orbit since October 1, 2026 (SpaceX Transporter-18 mission), contact confirmed, satellite operational.
• 8x higher solar energy in sun-synchronous orbit (~650 km) vs ground installations — no day/night cycle, no atmosphere.
• Radiation-hardened TPU Trillium v6e: 67 MeV proton beam tests show resistance to ~3x the expected dose over 5 years; memory remains the weak link.
• 1.6 Tbps bidirectional free-space optical links validated on the bench; 81-satellite constellation modeled.
• Economic goal: launch costs < $200/kg by the mid-2030s (vs ~$3,600/kg today on a reusable Falcon 9) → orbital compute competitive with terrestrial data centers.
• Two TPU prototypes expected in early 2027 with Planet Labs.

Recommended Tools

Why space? Energy, and more energy

The answer comes down to a single number: 8x. In sun-synchronous orbit at 650 km altitude, a solar panel receives up to eight times more cumulative energy over the course of a year than an identical panel on the ground. No darkness, no cloud cover, no atmospheric absorption. For Google, whose data center energy bill already runs into the billions of dollars, the equation is simple: orbital energy is free once the infrastructure is deployed.

The system design document published by Google Research describes an architecture in which each satellite carries its own TPUs, its deployable solar panels, its radiators, and its optical communication terminals. Energy feeds the compute directly — no DC/AC conversion, no distribution losses across kilometers of cables. The overall efficiency of the system is mechanically improved as a result.

This is a break from terrestrial data centers, where Power Usage Effectiveness (PUE) still hovers around 1.1 to 1.3 among the best operators. Every watt consumed by a GPU or a TPU on the ground drags along with it an extra 10 to 30% in overhead for cooling, distribution, and redundancy. In orbit, the vacuum is your radiator. Cooling is done through thermal radiation — passive, with no pumps, no water, no maintenance.

TPU Trillium v6e: silicon hardened for space

The heart of the system is the TPU v6e "Trillium". Google didn't originally design them for space — these are the chips that train Gemini 3.1 Pro and power Google Cloud's production workloads. But the Suncatcher team subjected them to a brutal test: a 67 MeV proton beam at NASA's radiation laboratory.

The result: the chips withstand roughly 3 times the total ionizing dose (TID) expected for a 5-year mission before showing permanent data corruption. That's a comfortable safety margin. The identified Achilles' heel? Memory. Registers, caches, and HBM are more sensitive to single-event upsets (bit flips caused by a single ion) than pure compute logic.

Google doesn't say more about the mitigation techniques — error-correcting codes, memory scrubbing, triple modular redundancy — but the figure of one error per 10 million AI requests deemed "viable" by the constellation analysis gives an idea of the reliability level being targeted. For inference, that's acceptable. For training models with billions of parameters, stronger guarantees or fault-tolerant algorithms will be needed.

Two full TPU prototypes are expected in early 2027, once again with Planet Labs. These will be the first true real-world test benches: orbital thermal cycling (-100 °C to +100 °C), continuous radiation, launch vibration, hard vacuum.

The network: 1.6 Tbps over free-space optics

Distributed compute is useless without an interconnect. Google has bench-validated free-space optical (FSO) links at 1.6 Tbps bidirectional per link. That's the same order of magnitude as NVLink/NVSwitch interconnects or 800G/1.6T InfiniBand links in terrestrial racks — but without fiber, without connectors, through the void.

The orbital dynamics analysis covers a constellation of 81 satellites. At 650 km altitude, each satellite sees its neighbors during windows of a few minutes per orbit (period ~97 minutes). The topology changes constantly. Routing must be adaptive, tolerant to outages, and capable of reassembling fragmented tensors across the mesh.

That's where the shoe pinches: latency. Even at the speed of light, a satellite-ground-satellite round trip adds 4 to 8 ms depending on elevation. Between satellites with direct line of sight, it's sub-millisecond. But for a synchronized training workload (all-reduce every 10-100 ms), orbital jitter forces a rethink of data parallelism algorithms. Google hasn't published details on the software stack — Pathways? A modified JAX? An adapted Megatron-LM? — but the challenge is as much software as it is hardware.

The Orbital Economy: The Launch Learning Curve

The tipping point is not technological. It is economic.

The curve follows a classic learning curve: each doubling of launched volume drops the unit cost by 15 to 20%. Starship, with its 150 t LEO capacity and full reusability (stage + booster), breaks the glass ceiling. At < $200/kg, the marginal cost of putting a kilogram of TPU, solar panel, and radiator into orbit becomes comparable to the amortized cost of a square meter of terrestrial data center — without the electricity bill, without the real estate, without the cooling.

NextBigFuture estimates that at this threshold, orbital compute becomes competitive with terrestrial data centers for high energy density workloads: foundation model training, massive inference, physics simulation. Low-density workloads (web servers, OLTP databases) will remain on the ground — latency to the end user is unforgiving.

The scarcity of compute: the strategic context

Project Suncatcher doesn't come out of nowhere. It fits into a compute war where every tech giant secures its capacity as a strategic resource.

Google recently rationed Gemini access for Meta — a clear signal: compute is the new scarce currency. Meanwhile, the company is launching Antigravity 2.0, its agent-first suite taking aim at Cursor and Claude Code, and rolling out Gemini Spark, a 24/7 AI agent meant to become a "second brain" for developers. These products consume compute at an unprecedented scale. Continuous inference, autonomous agents, multi-step reasoning: all of this multiplies demand by factors that terrestrial data centers struggle to absorb.

Suncatcher is the supply answer to this structural demand. Not a gadget, not a laser communication experiment. Production infrastructure. If the 2027 prototypes validate reliability, the next phase will be an operational constellation — perhaps as early as 2028-2029 for inference, early 2030s for training.

Risks and unknowns: what Google doesn't say

Space debris and regulation

A constellation of 81 satellites at 650 km is a lot of mass in orbit. ITU/FCC regulations require deorbiting plans (the 25-year rule, soon to be 5 years in the US). Google will have to demonstrate a clean end-of-life — deorbit propulsion, or an orbit low enough for a rapid natural atmospheric reentry. Deployable solar panels increase the cross-sectional area: more drag, but also more collision risk.

Physical and cyber security

A data center in space is physically inaccessible. No technician to swap out a disk, a board, a cable. Redundancy must be n+2 at minimum. On the cyber side, the attack surface includes the ground segment (control stations, gateways), the space segment (optical links, telemetry RF), and the supply chain (chips, embedded software). A hostile state actor could attempt optical jamming, fault injection via telemetry, or a supply chain attack on the TPUs before launch.

Unit economics: satellite CAPEX vs rack

An 8-GPU HGX H100 rack costs ~$300,000 (2024 prices). A Suncatcher satellite carrying the equivalent in TPUs + panels + radiators + propulsion + bus + optics: how much? Google isn't saying. But if the satellite costs $10M and lasts 5 years, the monthly amortization is ~$167k/month — versus ~$8k/month for the rack (amortized over 3 years). The energy savings ($0/kWh vs ~$0.06–0.10/kWh industrial) and the absence of PUE overhead have to make up the difference. Over 5 years, a 10 kW rack consumes ~4.4 GWh → $260–440k of electricity. That delta doesn't cover the CAPEX gap. The math only works if launch costs drop drastically — hence the <$200/kg target.

End-user latency

For interactive inference (chat, code, search), the user-satellite-user round-trip latency is prohibitive: 20–40 ms minimum (ground → satellite → ground) + processing. Starlink manages it for IP traffic because user terminals talk directly to the satellite. For AI, the model has to be in the satellite, and the user has to also be connected to the satellite — or accept the latency. Google is probably targeting asynchronous workloads (training, batch inference, pre-computing embeddings, synthetic data generation) where latency doesn't kill the UX.

Common Mistakes

Mistake 1: Confusing Suncatcher with Starlink

What's wrong: Thinking that Google is building a communication constellation like Starlink.
The reality: Suncatcher is a compute infrastructure. The optical links are inter-satellite (and ground-to-satellite for ingestion/results), not for serving user terminals. No user "phased array", no consumer Ku/Ka band.

Mistake 2: Believing that space solves heat "for free"

What's wrong: Imagining that the vacuum dissipates heat without constraints.
The reality: Thermal radiation follows the Stefan-Boltzmann law (∝ T⁴). To dissipate 10 kW at 300 K, you need ~20 m² of radiators with emissivity 0.9. Solar panels generate heat in addition to electricity. Thermal design is a headache: deployable radiators, heat pipes, loop heat pipes, surfaces with variable optical properties. It's not "free" — it's different.

Mistake 3: Underestimating software complexity

What's wrong: Believing that JAX/Pathways runs "as-is" on a mobile constellation.
The reality: The topology changes every 97 minutes. Links go up and down. Bandwidth varies. Failures are the norm (radiation, eclipse, debris). You need a space-native orchestrator: frequent checkpointing, elastic training, fault-tolerant all-reduce, priority-based scheduling based on orbital visibility. Google hasn't published this stack — it's the real differentiator to come.

Frequently Asked Questions

When will we see Suncatcher compute in production?

Not before 2028-2029 for batch inference, early 2030s for training. The two 2027 prototypes will validate hardware reliability. The operational constellation requires a sustained launch cadence (Starship) and a mature software stack.

Which AI models run on the TPU Trillium in orbit?

The same as on Google Cloud: Gemini 3.1 Pro, Gemini 3 Pro Deep Think, and Google's internal models. The TPU architecture is identical — only the radiation qualification changes. Inference of open-source models (Llama, Gemma) is technically possible but not announced.

How much does an AI query processed in space cost?

Unknown today. Google does not publish marginal costs. The working hypothesis: at < $200/kg launch, orbital $/token becomes competitive with terrestrial $/token for high energy density workloads. No public pricing schedule.

Does user data leave Earth?

No, not in the current architecture. Data ingestion and result delivery go through dedicated ground stations (probably co-located with Google Cloud PoPs). The space segment does not store persistent user data — only model weights, temporary activations, and training checkpoints.

What is the net environmental impact?

Positive if the avoided terrestrial energy is fossil-based. Every kWh produced in orbit replaces a terrestrial kWh. But you must account for: satellite manufacturing (aluminum, carbon, rare earths), launch (kerosene/methane for Falcon 9, methane/oxygen for Starship), and end of life. A complete life cycle assessment (LCA) does not publicly exist. The carbon break-even depends on the terrestrial electricity mix being replaced.

✅ Conclusion

Project Suncatcher is not a laser communication experiment nor a PR stunt. It's an industrial bet: the wager that the learning curve of space launch (Starship, full reusability) will cross the exponential demand curve for AI compute before 2035. The prototype is in orbit. The chips are holding up to radiation. The optics are delivering 1.6 Tbps. What remains to be proven is that the economics hold up — and that the software knows how to dance on a topology that changes every 97 minutes.

If Google succeeds, the data center of the future won't be on the banks of a river or in the desert. It will be above our heads, powered by a star that never sets.