Google’s Suncatcher Puts AI Compute Into Orbit

• Google’s Project Suncatcher turns solar LEO swarms into AI datacenters: TPUs linked by free-space optics at 1.6 Tbps; 2027 prototypes, radiation-tolerant chips, but heat and ground links loom.

Google unveiled Project Suncatcher, a moonshot to build space-based AI compute using solar-powered constellations of small, networked satellites carrying TPUs linked by free-space optics. A sun-synchronous dawn–dusk LEO orbit would provide near-continuous power; close formations (100–200 meters) aim to deliver data center-class bandwidth. A bench demo hit 800 Gbps each way (1.6 Tbps total). Proton-beam tests of TPU v6e showed promising radiation tolerance.

Google projects launch costs could fall below $200/kg by the mid-2030s, making space compute potentially cost-competitive on energy. Key hurdles remain—thermal management, ground links, reliability. With Planet, Google plans two prototype satellites by early 2027 to validate models and optical links.

Can dawn–dusk LEO solar reliably power TPU constellations cost-competitively at scale?

• Power availability: Dawn–dusk sun‑sync offers the highest LEO solar duty cycle, but not 100%. Expect near‑continuous sunlight for months, with eclipse seasons twice a year; worst‑case eclipses ~20–35 minutes per ~95‑minute orbit. Reliable TPU uptime requires either overprovisioned arrays and buffers or battery/flywheel storage sized for those windows.
  
• Array sizing and specific power: State‑of‑the‑art flexible multi‑junction arrays deliver ~150–300 W/kg beginning‑of‑life, trending higher by early 2030s; end‑of‑life derates of 20–35% from radiation/contamination/thermal. Practical delivered electrical power after conversion and pointing losses: ~200–300 W/m² to the bus in steady sun. Meeting multi‑kW per node is feasible but drives large deployed area and precision pointing.
  
• Thermal balance for TPUs: Waste heat, not solar input, is the gating factor. Radiators at 250–320 K typically reject ~40–120 W/kg depending on coatings, view factor, and plumbing. A multi‑kW TPU stack demands radiator mass on the same order as the arrays, plus variable rejection or phase‑change buffers to ride through eclipse spikes. Thermal design is solvable but sets the mass floor.
  
• Storage to ride eclipses: Even with dawn–dusk, budget ~5–15% of orbital time in shadow across the year. For a 5 kW compute load, a 30‑minute worst‑case eclipse needs ~2.5 kWh usable. Space‑qualified Li‑ion at ~80–140 Wh/kg (end‑of‑life) implies tens of kilograms per node, plus power electronics. This materially affects cost unless loads can throttle.
  
• Lifetime and degradation: Annual power fade from radiation and UV darkening, micrometeoroid risk to thin films, and atomic oxygen erosion require margins or servicing. Expect 5–7 year economic life without refurbishment; maintaining constellation output at scale likely needs replacement launches or in‑space servicing.
  
• Formation operations impact yield: Tight 100–200 m formations add pointing constraints; thermal plumes and reflected sunlight can raise array temperatures and lower efficiency. Attitude/agility to maintain optical links costs momentum management power, slightly trimming net power budget.
• Availability strategy: Three workable options—(1) accept brief brownouts and checkpoint workloads, (2) oversize arrays and radiators to run hotter and charge during sun, (3) add storage. Option (1) is cheapest but narrows workloads; (2) and (3) add mass but keep SLAs.
  
• Cost model vs. ground energy: If launch falls to ~$200/kg and the combined power+thermal system achieves ~200 W/kg to the load, launch cost embeds at ~$1/W. Adding array/radiator/battery hardware and integration could plausibly land total power system at ~$2–4/W initial capex. Spread over 5 years at high capacity factor, the implied energy cost can be competitive with terrestrial data center electricity in premium markets, but only if node utilization remains very high and replacement cadence is efficient.
• Compute $/TOPS and link $/Gbps dominate: Even if solar LCOE in orbit is attractive, total cost hinges on TPU cost, radiation‑hardened power electronics, formation‑keep fuel, and ground relay. If optical inter‑sat links keep most traffic in‑space and duty cycle stays >85%, the energy advantage can translate to lower $/inference; heavy ground egress erodes it.
• Scalability constraints: Debris mitigation, spectrum/regulatory for crosslinks to ground relays, and optical ground station weather diversity add non‑energy costs. Mass production of arrays/radiators and standardized buses is essential to hit the $/kg and $/W assumptions.
• Bottom line: Yes, dawn–dusk LEO solar can reliably power TPU constellations if designs assume eclipse seasons, size thermal rejection correctly, and keep utilization high. Cost‑competitive at scale looks plausible in the mid‑2030s under aggressive launch prices and mass‑efficient power/thermal systems, but margins are tight and highly sensitive to radiator specific performance, storage mass, node lifetime, and ground link costs.