Google Project Suncatcher: four TPUs are in orbit, but what is proven?

Google Project Suncatcher is no longer only a paper. On October 1, 2026, Google says, a prototype satellite carrying four of its tensor processing units (TPUs) went up on SpaceX's Transporter-18 rideshare, built with Planet, and Google says it has made contact and the spacecraft is operating as expected. That settles a muddle in the coverage: one set of stories says two prototypes are due in early 2027, another says a prototype is already in orbit. Both are true, and they describe different things. What is proven so far is that four chips can be launched and talked to. Whether space can host serious AI compute is a separate question that this satellite is not built to answer.

What is true as of October 4, 2026

Google's research blog post of October 1 says a Planet-built prototype launched on Transporter-18 and was operating as expected on contact. It is one satellite. NPR's report, published October 2, describes it as refrigerator-sized, carrying four TPUs, and says Google plans to put "two more satellites into orbit to test their connection" next year. TechTimes places the launch at Vandenberg Space Force Base on a Falcon 9. Planet's own launch announcement, as summarized by GuruFocus, says the Transporter-18 flight carried 20 of its satellites, including the Suncatcher demonstration unit.

The "two prototypes by early 2027" line comes from the original announcements. Google's November 4, 2025 research post and Planet's announcement of the same date both said two prototype satellites would launch by early 2027. The satellite now in orbit arrived ahead of that, and the two-satellite mission, whose purpose is to test laser links between spacecraft, is still ahead. Our reading: the early 2027 date has not been withdrawn, but we found no fresh Google or Planet statement restating it after the launch. Treat it as a plan, not a schedule.

Two small discrepancies are worth flagging. Google's explainer page still carries a September 2026 launch date, which looks stale. And outlets disagree on whether the satellite runs Google's open Gemma model or Gemini on its 15-minute bursts: NPR says Gemma, TechTimes says Gemini. We rely on NPR.

The numbers in the paper

The underlying research, "Towards a future space-based, highly scalable AI infrastructure system design," first appeared on arXiv on November 22, 2025 and was later published in the journal Joule, per Google's launch post. Its headline figures, all from Google's own team:

• Radiation. Google put its Trillium (v6e) TPUs in a 67 MeV proton beam. The high-bandwidth memory was the most sensitive part, with irregularities beginning at a cumulative dose of 2 krad(Si). The paper compares that with an expected five-year mission dose of 750 rad(Si), and reports no hard failures attributable to total ionizing dose up to the maximum tested 15 krad(Si).
• Optical link. A bench-scale demonstrator reached 800 Gbps each way, 1.6 Tbps total, with one transceiver pair. That is a lab bench, not two spacecraft moving relative to each other, which is what the 2027 mission is meant to test.
• Constellation design. The modeled cluster has 81 satellites at 650 km altitude within a 1 km radius, flying close enough to make the optical links work.
• Cost. The economic case assumes launch prices fall below $200 per kilogram by the mid-2030s, at which point Google says launching and operating a space data center could become roughly comparable to the reported energy costs of an equivalent terrestrial one.

Notice what the radiation result does and does not say. Chips survived a dose well above what a five-year mission would deliver. That is a good result, and it is why Google can say the hardware tolerates the environment. It is a ground test of a chip, not a flight result of a system, which is exactly what the prototype is now collecting.

What the prototype can and cannot show

Reports on the satellite put its solar power at about one kilowatt, and the chips run for roughly 15 minutes at a time because the craft cannot shed heat faster. In vacuum, heat leaves only by radiation, so radiators, not solar panels, set the limit. NPR reports the chips will answer simple queries. This is a thermal and survival test, and a modest one.

The project lead is candid about the economics. Travis Beals, senior director and lead of Project Suncatcher, told NPR: "I don't see this being something where it's cheaper to do this in the next five years. I think it will take longer than that." That is the most important sentence in the coverage, because it comes from the person promoting the project.

The cost claim also leans on a number that does not exist yet. Under $200 per kilogram is a forecast that depends on cheaper, higher-cadence launch that Google's paper treats as a sustained learning rate. If it does not arrive, the comparison to terrestrial energy cost falls apart.

Orbital compute versus the power bottleneck on the ground

The pitch for orbit is that sunlight is plentiful and constant, and Google says a dawn-dusk orbit can collect several times the solar energy of a panel on the ground. The pitch for the ground is that the actual constraint on AI buildouts today is not raw energy in the abstract. As we reported in our piece on the grid connection bottleneck, the delay is getting power delivered to a site on time through interconnection queues and equipment lead times.

Orbit sidesteps that queue but swaps in different ones: launch capacity, radiators, repair (nobody can send a technician), and the cost of lost or aging hardware. Carnegie Mellon's Brandon Lucia told NPR that maintenance gets harder when the data center is circling hundreds of miles above Earth. Four chips and a kilowatt tell us nothing about those costs at scale. Even Google's context is telling: NPR notes that Sundar Pichai said the company expects $180 billion to $190 billion in capital expenditures this year.

Our position

We would rate this a real research milestone and a poor guide to near-term capacity. Google has put hardware in orbit, confirmed contact, and published peer-reviewed numbers, and that deserves credit. It has not shown a satellite doing sustained useful work, linking to a neighbor, or competing on cost.

What we would watch, in order: first, Google's and Planet's reporting on how the four TPUs fare over the coming months, including error rates, not just uptime. Google has said only that it will share more as the mission unfolds. Second, whether the two-satellite laser link test keeps its early 2027 timing. Third, any launch price actually approaching the $200 per kilogram threshold. Until those arrive, anyone planning AI capacity should keep solving the grid, because that is where the gigawatts are being built. Orbital compute is a bet on the 2030s, and its own lead says it will not be cheaper within five years.