Google and Planet Labs Pursue Orbital AI as Grid Queues Stretch

The case for putting AI data centers in orbit was never supposed to be that space is cheap. As CNBC reports, orbital compute costs roughly four times the equivalent terrestrial setup in 2026, and no serious analyst argues that changes soon. The more interesting argument is about what is actually scarce. Earth's power grid is reportedly becoming the binding constraint on AI expansion, not cost per computation: grid connection queues in primary markets like Northern Virginia stretch to seven years, and a 1 MW AI cluster generates roughly $70 million per year in cloud GPU revenue, meaning a five-year delay represents around $350 million in foregone revenue per megawatt. Against those numbers, expensive orbital infrastructure starts to look like a rational hedge.

Google and Planet Labs are the collaboration pushing this furthest under the public name Project Suncatcher, which Google describes as a research moonshot to scale machine learning in space. The design uses solar-powered satellites carrying Google Tensor Processing Unit chips connected by free-space optical inter-satellite links. A Google research paper envisions a potential 81-satellite cluster within a 1km radius. As SpaceNews reported, Planet will build and operate the platform, with two spacecraft carrying Google TPUs slated to launch by early 2027 to test TPU performance in the space environment and demonstrate high-bandwidth inter-satellite links.

SpaceX is the potential launch partner, and the project's long-run economics are explicitly tied to Starship's cost trajectory. Google's analysis holds that if launch costs to low Earth orbit reach $200 per kilogram, launch costs amortized over spacecraft lifetime become roughly comparable to data center energy costs on a per-kilowatt basis. Google theorizes that threshold could be reached by 2035 if Starship launches 180 times per year. That is a very large conditional, and one the CNBC analysis treats with appropriate caution.

What the reporting does not fully address is the workload fit question. Analysts note that LEO satellites add roughly 4.6 to 7.4 milliseconds of round-trip latency, described as potentially disqualifying for real-time inference workloads but irrelevant for AI training and batch inference, which account for the majority of GPU hours. The two-spacecraft test in 2027 will provide early data on whether TPUs survive and perform in orbit, but Google itself notes that significant technical and logistical hurdles still exist at cluster scale. The bet here plays out over a decade, and only if Starship's launch cadence actually materializes.