The Ghost in the Vacuum
The Ghost in the Vacuum
Five hundred kilometers above the Arctic Ocean, a refrigerator-sized satellite fired up an array of terrestrial data-center silicon and wrote a haiku.
Data streams in silent grace
Starlight feeds my core. ✨
The verse itself is unremarkable—the sort of polite, competent prose large language models generate millions of times a day on Earth. But the hardware executing the matrix multiplications was floating through the thermosphere at 17,000 miles per hour.
The spacecraft is the lone vanguard of Project Suncatcher, a joint experimental testbed built by Google and satellite imaging operator Planet. Launched on October 1 aboard SpaceX’s Transporter-18 rideshare mission, the satellite carries four Google Trillium Tensor Processing Units (TPUs). It also carries an instance of Gemma 4, marking the first time a frontier open weights model has generated tokens natively in the hostile vacuum of low Earth orbit.
Silicon in space is nothing new, but the silicon currently operating aboard Suncatcher breaks nearly every convention of aerospace engineering.
Escaping the 1990s Compute Trap
For half a century, computing in space has traded speed for paranoia.
Spacecraft avionics are notoriously archaic. The RAD750 processor that drives NASA’s Perseverance rover across Mars is essentially a hardened PowerPC 750 chip from a 1997 Apple iMac G3, fabricated on a 150-nanometer process. It runs at an asthmatic 200 megahertz, commands tens of thousands of dollars per unit, and sips single-digit watts.
Aerospace engineers tolerated the abysmal clock speeds because survival mattered more than throughput. In deep space and low Earth orbit alike, heavy galactic cosmic rays and high-energy solar protons tear through silicon lattices. In a microscopic modern transistor, a single stray proton can flip a bit from a zero to a one, corrupting memory, crashing flight control systems, or permanently burning out gate oxide layers. Radiation-hardened chips survived because their internal features were broad, rugged, and slow.
Google and Planet are pursuing the opposite thesis: What happens if you take modern, dense, commercial data-center accelerators—etched with microscopic geometries measured in nanometers—and throw them straight into the radiation storm?
Prior to launch, Google engineers strapped Trillium silicon into a beamline at the Crocker Nuclear Laboratory at UC Davis, bombarding the chips with 67 MeV protons while driving sustained inference loads. The test demonstrated that the processors could survive the equivalent total ionizing dose of a five-year mission without permanent failure while researchers characterized bit-flip errors.
Suncatcher is the real-world validation. And Gemma 4 is the canary in the coal mine.
For half a century, computing in space has traded speed for paranoia. Modern orbital infrastructure asks whether probabilistic weights can simply absorb the storm.
Twenty Minutes of Consciousness
Operating an AI accelerator in a vacuum is an exercise in thermodynamic physics.
In an air-cooled or liquid-chilled server hall in Council Bluffs or The Dalles, a cluster of Trillium chips runs around the clock, with massive cooling towers carrying away megawatts of waste heat via continuous convection.
In the vacuum of space, convection does not exist. There is no air to sweep heat off a heat sink. Every watt of heat generated by the silicon must be physically conducted away to the chassis through embedded heat pipes and dumped into the void purely as infrared radiation from exterior thermal panels.
Because of this physical bottleneck, the Suncatcher satellite operates under severe thermal and electrical rationing:
- The Solar Envelope: The satellite’s solar arrays deliver approximately 1 kilowatt of power—enough to run individual test sequences, but far short of a continuous compute load.
- The Single-Core Duty Cycle: While four TPUs are mounted inside the hull, the satellite powers only one TPU at a time.
- The Operating Window: Gemma 4 does not idle quietly in background memory. The TPU is brought online in isolated bursts lasting approximately 15 to 20 minutes per pass. As the silicon works, heat accumulates faster than the spacecraft can continuously reject it, requiring the accelerator to cycle down so the thermal radiators can dissipate the thermal load into space.
Gemma 4 is not up there to answer customer queries or analyze surveillance imagery. Ground controllers uplink prompts as the satellite comes into line of sight with a polar ground station in Svalbard, Norway. By running short creative generations alongside telemetry-related prompts, engineers are directly evaluating accelerator behavior, operational stability, and thermal dissipation under live orbital conditions.
Tolerating the Glitch
The more radical engineering question Suncatcher seeks to answer is computational: Can a modern transformer tolerate a bit flip?
Traditional flight software cannot tolerate a single corrupted bit; a flipped sign bit in an orbital mechanics equation can send a probe tumbling out of control. But transformer models operate on probabilistic vector weights.
By running repeated inference workloads through Gemma 4, engineers can track whether radiation exposure causes silent data corruption inside the TPU registers or memory arrays. Does a single bit flip in an attention weight go unnoticed in the noise of floating-point rounding? Or does it trigger an attention collapse that sends the model into gibberish loops?
The answers will determine whether orbital data centers are feasible, or merely an expensive science fiction fantasy.
A flipped sign bit crashes an orbital thruster. But in a neural network running billions of floating-point parameters, a single bit flip might simply dissolve into the noise.
The 2027 Horizon
If Suncatcher's single-satellite experiments succeed over the coming months, the program moves quickly to its next phase. In 2027, the team plans to launch twin satellites equipped with free-space optical laser communications.
The longer-term goal is the far more difficult problem of distributed model parallelism across open space—interconnecting separate spacecraft flying in tight formation with high-bandwidth optical links to share machine-learning workloads.
For now, the solitary satellite continues its silent loop from pole to pole. Every few orbits, as the Arctic ground station rises over the horizon, a single Trillium chip powers up, dumps its heat into the dark, drops a packet of verified tokens down to Earth, and goes cold again.
Aaron Rose is the founder and publisher of Tech Reader Magazine, covering system architecture, cloud platforms, and silicon engineering.