Ground-based AI compute is hitting limits in power, cooling, and land. Google has launched Project Suncatcher, which will send a prototype satellite carrying four TPU chips into orbit for the first in-orbit AI chip validation. The test will assess chip resilience to launch shock, space radiation, and extreme temperatures, laying groundwork for a two-satellite laser communication network in 2027. Google is not alone: SpaceX, Starcloud, and Guoxing Aerospace are also pursuing orbital compute, opening a technology race in space data centers.
Ground data center expansion faces clear bottlenecks. China's ICT Institute reports domestic data center power consumption hit 196 billion kWh in 2025. The main constraint on new compute infrastructure is no longer chip supply but insufficient power, transformer queues, and slow grid approvals. Energy, cooling, and land pressures limit ground expansion. Low Earth orbit offers near-continuous solar power, bypassing grid limits, attracting multiple orbital compute efforts. Google's Project Suncatcher aims to validate the feasibility of large-scale space AI infrastructure.
The mission uses Google's sixth-generation Trillium TPU, a custom ASIC machine learning chip. The satellite is about the size of a refrigerator, carries four TPUs, and delivers roughly the compute of one ground server. It runs on onboard solar panels at about 1 kW in a dawn-dusk sun-synchronous orbit for stable sunlight. Four technical challenges must be overcome. During launch, rocket vibration subjects chips to 50–100G acceleration loads, requiring structural integrity. Space radiation can cause bit flips and data errors; Google has run ground-based radiation AI workload tests, but only in-orbit testing proves real reliability. In vacuum, traditional air cooling fails, so Google uses heat pipes with radiator panels for conduction cooling. A single satellite has limited compute; the long-term goal is a satellite compute cluster linked by high-bandwidth inter-satellite laser communication. Google plans to launch two satellites in 2027 to validate high-speed laser links. It is also developing a next-gen server chip codenamed Frozen v2, expected to deliver 6–10x the token throughput per watt of current TPUs. The satellite will launch via SpaceX Transporter-18 rideshare with satellite imaging firm Planet Labs. The mission's core purpose is collecting orbital operating data and diagnosing hardware faults, not delivering a commercial orbital data center.
Orbital compute players fall into three groups. SpaceX, representing the rocket track, proposed an orbital data center concept in late 2025, using LEO satellites and solar power to bypass ground constraints. With reusable rocket cost advantages and Starlink cash flow, SpaceX is also building a Terafab chip factory for vertical integration, though its prospectus flags commercialization uncertainty for orbital AI data centers. TrendForce forecasts the global satellite industry could reach $447 billion in 2027, growing 14% annually. Startups like Starcloud focus directly on LEO data centers powered by continuous solar energy. Chinese firms are accelerating: Guoxing Aerospace launched its Star Compute Plan for a 2,800-satellite compute constellation and, with Shanghai Jiao Tong University, demonstrated space-based compute controlling a ground robot. Aerospace Hongtu, with 12 in-orbit Nuwa constellation satellites, is advancing space compute satellites carrying onboard GPUs, FPGA intelligent payloads, and inter-satellite laser communication. Supply chain players are moving too: Shiyun Circuit's customers plan large solar plants for a vehicle-storage-satellite integrated energy system, and its commercial space PCB products are in R&D sampling.
Industry views say orbital data centers remain far from large-scale commercialization, with launch costs, engineering difficulty, and satellite manufacturing capacity as real obstacles. But orbital and spectrum resources are scarce, pushing leading firms to secure early advantage. Google, with its own TPUs, can integrate chips, large models, and constellations—making it one of the few full-stack players. Compute always revolves around energy and cooling. Space compute is still in early validation, and upstream segments like laser communication terminals, onboard compute payloads, and launch services may see orders land earlier than complete orbital data centers.
ICgoodFind: Google's Project Suncatcher begins in-orbit TPU validation, accelerating space-based AI chips and inter-satellite laser links as firms worldwide race into orbital compute.
