Google's Project Suncatcher satellite successfully runs inference in orbit, generating haiku from space
Google's AI-equipped satellite booted Tensor Processing Units and executed inference with an on-board Gemma 4 model, marking the first step toward orbital data centers.
Last verified:
Google’s orbital AI experiment has moved from launch to execution. According to The Verge AI, the company’s Project Suncatcher satellite successfully activated its Tensor Processing Units (TPUs) and ran inference on an on-board Gemma 4 model while in space, marking the first functional test of AI compute beyond Earth’s atmosphere.
Suncatcher’s first inference task: space haiku generation
The satellite generated a haiku about its orbital vantage point in response to a prompt:
Blue marble below
Data streams in silent grace
Starlight feeds my core
According to The Verge, Google tested additional space-themed prompts and telemetry queries alongside this primary demonstration. The haiku output confirms that the Gemma 4 model could accept input, execute a language task, and return a response across the satellite’s communications link—validating the basic inference pipeline in an environment Google had never tested TPUs before.
Stress testing and the 2027 dual-satellite roadmap
Google’s blog post, as reported by The Verge AI, signals that this initial success is a stepping stone. The company intends to run a series of thermal and radiation-exposure tests on the TPUs over coming months to measure whether the chips degrade under space conditions or maintain reliability. Those results will inform the next phase: deployment of twin interconnected satellites in 2027, which Google has flagged as the milestone for demonstrating multi-satellite workload distribution.
The company launched the Suncatcher satellite earlier in October 2026, so the inference-generation announcement arrives roughly a month into the mission.
Why This Matters
Project Suncatcher challenges a core assumption in AI infrastructure: that large-scale compute must remain terrestrial. If TPUs can reliably execute inference workloads in orbit—and if Google can manage the latency, thermal constraints, and radiation hardening challenges—the cost of cooling and real estate for data centers could shift. Teams evaluating long-term AI deployment strategies should monitor whether the 2027 dual-satellite test demonstrates meaningful throughput and whether Google publishes performance benchmarks (latency, accuracy drift under radiation) that would make space-based inference commercially viable rather than experimental.
The inference-in-space narrative also carries strategic weight: it positions Google’s custom silicon (TPUs) as adaptable to extreme environments at a time when competition in chip design is intensifying. If the company can extend TPU utility to orbital conditions, differentiation against Nvidia’s GPUs and bespoke chips from OpenAI or Anthropic deepens—though this remains years from revenue impact.
Frequently Asked Questions
What is Project Suncatcher?
Project Suncatcher is Google's experimental initiative to deploy AI data centers in space, starting with chip-equipped satellites that run inference workloads in orbit.
What model did the satellite run?
The satellite ran Gemma 4, Google's open-weights model, which generated a space-themed haiku as a proof-of-concept inference test.
What happens next?
Google plans to test TPU performance under thermal stress and radiation exposure, followed by deployment of twin interconnected satellites in 2027.