Google's Project Suncatcher Sends AI Chips Into Space

Update, Oct. 1: the Suncatcher prototype is in orbit

It flew. SpaceX's Transporter-18 rideshare mission lifted off from Vandenberg Space Force Base on October 1 at 11:32 a.m. PDT, carrying 130 payloads — including Google's Project Suncatcher M1 prototype, the solar-powered satellite built with Planet Labs to test whether Google's Tensor Processing Units can survive and operate in space (per nasaspaceflight.com's launch coverage). The Suncatcher payload separated at T+1:01:25, and the Falcon 9's first stage — flying its 25th mission — touched down at Landing Zone 4 for another recovery.

What happens now is the unglamorous part: months of watching. Google and Planet Labs engineers will track how the TPUs handle launch vibration, space radiation and the brutal thermal problem of cooling chips with no air — the make-or-break questions for any future orbital data center. Google CEO Sundar Pichai framed the mission on X back on September 25: "Can our TPUs survive and operate in space? Well, we're going to find out. Project Suncatcher is hitching a ride aboard @SpaceX's Transporter-18 mission." The roadmap already extends past this flight: a 2027 follow-up mission with two satellites is planned to test high-bandwidth optical laser links — the interconnect layer you'd need before TPUs in orbit can talk to each other like a real cluster (per Gujarat Samachar).

One note of caution, from Google itself: SVP James Manyika has said publicly the technology isn't expected to be usefully operational for years. This is a science experiment with a rocket, not a product launch — and that's exactly what makes it interesting.

SPACE COMPUTING

Google's AI chips are about to leave the planet. The question the company wants answered is brutally simple — and nobody knows the answer yet.

Published September 28, 2026 · By Daniel Rao

Update, Sept. 28: October 1 confirmed — and the spec sheet is out

The launch date is now firm: October 1, 2026, on SpaceX's Transporter-18 rideshare from Vandenberg Space Force Base, with Google's refrigerator-size MVP satellite — built with Planet — carrying four Trillium TPUs on about 1 kilowatt of solar power (Dailyhunt, Trak.in, Futurum, Sept 26–28). Thermal limits mean the chips will run Gemini queries in roughly 15-minute bursts before shutting down to cool; the spacecraft is designed for a one-year mission and will deorbit in about six years (airpronews, Futurum). The competitive field is widening: Futurum estimates orbital compute could address a $1 trillion market by 2030 if launch costs collapse, and reports SpaceX's own orbital-compute effort ("Starmind AI1") would use Nvidia Vera Rubin NVL72 hardware at ~120 kW — two orders of magnitude above Suncatcher's test rig. Sharing the October 1 ride: Hyderabad-based TakeMe2Space's MOI-1a orbital-computing satellite (GadgetsNow/IndiaTimes).

Next week, a small satellite carrying some of the most valuable silicon on Earth will hitch a ride on a SpaceX Falcon 9 and go to a place no Google TPU has ever been: low-Earth orbit. The prototype mission is the first in-orbit test of Project Suncatcher, Google's research effort to figure out whether the future of AI computing might not be in a data center at all — but in space.

Google CEO Sundar Pichai framed the stakes in a single sentence on September 24, when the company announced the flight: the mission will try to answer whether Google's Tensor Processing Units can "survive and operate in space." That is the entire program. Everything else — the solar constellations, the laser links, the orbital data centers — is a bet on the answer.

A SpaceX Falcon 9 lifts off from Vandenberg's Space Launch Complex 4E — the California launch complex that hosts the Transporter rideshare missions the Suncatcher prototype will fly on. Photo: NASA/Bill Ingalls. (Shown: the 2018 GRACE-FO rideshare launch, not Transporter-18.)

What Google is actually launching

The satellite will fly on SpaceX's Transporter-18 rideshare mission, the company's dedicated smallsat carpool to sun-synchronous orbit, and was built in partnership with Planet Labs, the San Francisco satellite company that has launched hundreds of its Dove imaging satellites on similar rideshares.

On board are Google TPUs — the custom AI accelerators that power the company's Gemini models and its cloud AI services. Once in orbit, the prototype will run AI workloads while engineers measure how the chips handle the three punishments space inflicts on electronics:

• Launch forces — the vibration and acoustic violence of a Falcon 9 ascent.
• Radiation — cosmic rays and trapped particles that flip bits in memory and can degrade chips over time.
• Thermal extremes — swinging between baking sunlight and deep-shadow cold, hundreds of degrees apart, every 90-minute orbit.

Google says it has already run the TPUs through ground testing at a University of California, Davis facility, pushing them with AI workloads in conditions meant to mimic the space environment. But a lab chamber is not orbit, and the company is explicit that only real flight data will tell it what it needs to know.

Crucially, this flight is an engineering test, not a product demo. Google says the mission is designed to gather in-orbit data and identify potential failure points — not to demonstrate a working orbital data center. If you're picturing a Gemini server farm humming above the Earth next month, dial it back.

The cooling problem: no fans in a vacuum

The single most interesting technical detail in the announcement is also the least glamorous. In a terrestrial data center, cooling is a brutal, industrial affair — massive fans, chilled water loops, evaporative towers. In orbit, none of that works. There is no air to move, and water is too heavy to ship up by the ton.

So Google is testing a passive design: heat pipes paired with radiators. Heat pipes wick thermal energy away from the power-hungry chips, and radiators dump it directly into the cold of space as infrared radiation. The approach is standard on satellites, but nobody has tried to cool datacenter-class AI accelerators this way at scale. TPUs are ferociously power-dense; whether the plumbing holds up when the chips are running real ML workloads in vacuum is an open engineering question — and one of the main reasons this flight exists.

The radiation question is equally real. Space radiation doesn't just wear chips down; it causes bit flips — single-event upsets where a particle strike corrupts data in memory. Terrestrial AI clusters already obsess over error correction. In orbit, the error budget gets much tighter, and Google's bench tests — including, according to earlier reporting, proton-beam exposure of TPU v6e hardware — can only approximate what the flight will encounter.

A decade of patience: what Suncatcher actually is

This flight didn't come out of nowhere. Google Research unveiled Project Suncatcher in November 2025 as a "research moonshot," describing compact constellations of solar-powered satellites carrying TPUs and linked by free-space optical connections — essentially lasers — that would let the constellation behave as one distributed computer.

The orbital mechanics are the pitch. In a dawn-dusk sun-synchronous orbit at around 650 kilometers, satellites can stay in near-continuous sunlight, and Google has claimed solar panels in that configuration can be up to eight times more productive than the same panels on the ground. The original plan called for two prototype satellites with Planet by early 2027 to validate the hardware and the optical links. This new Transporter-18 flight is an earlier, smaller check — a hardware rehearsal before the real dress rehearsal.

What comes next is already on the calendar. Google says it plans to launch two satellites in 2027 to test the high-bandwidth laser links needed to connect future computing clusters — the networking backbone of the whole concept. Google has demonstrated 800 Gbps one-way transmission on a bench with a single transceiver pair, but beaming data between satellites flying hundreds of meters apart at 7.5 kilometers per second is a different sport.

And if you want to calibrate your expectations, listen to James Manyika, Google's senior vice president for research. Asked about timelines, he told the New York Times: "We don't expect, to be perfectly frank, that we'll have anything usefully operational in the next few years." Then he added the line that should be taped to every Suncatcher slide deck: "Remember how Google was researching for like 15 years, before anything showed up? I think this is going to look like that."

Planet Labs Dove satellites deploying from the International Space Station in 2014. Planet — Google's partner on Suncatcher — has launched hundreds of small satellites on rideshare missions and will build the prototype spacecraft. Photo: Steve Jurvetson / NASA (CC BY).

The race overhead: SpaceX, Starcloud, and the new space compute rush

Google is not the only one looking up. SpaceX CEO Elon Musk has filed with the Federal Communications Commission for a vast constellation to support orbital data centers — reporting puts the request at up to a million satellites. Jeff Bezos' Blue Origin has reportedly filed its own FCC request under the name "Project Sunrise."

The startups are already flying. Starcloud, a Redmond, Washington startup that came out of Y Combinator, raised $170 million at a $1.1 billion valuation (per TechCrunch) and launched its Starcloud-1 demonstrator in November 2025 carrying an Nvidia H100 — the first data-center-class GPU operated in space. During that mission it reportedly ran inference on Google's own Gemma model and trained a small language model entirely in orbit. The company says bigger satellites, with Blackwell GPUs and eventually Starship-class data center spacecraft, are on its roadmap.

China is in the game too: startup ADA Space launched a dozen satellites in 2025 toward what it calls an orbital supercomputer constellation. The contest for the space compute layer is forming the same way the cloud contest did — early, crowded, and American-led, with the serious players racing to de-risk the physics before anyone has to defend the economics.

Why it matters for the US: the energy wall

The deeper reason Suncatcher exists has nothing to do with space romanticism. It is about electricity. Training and serving frontier AI models consumes gigawatts, and American utilities are straining to approve new data center power at the speed the AI labs want to build. In several states, new AI facilities are waiting years for grid interconnections, and the industry's water use has become a political issue of its own.

Space offers, in principle, a way around all of it: near-limitless solar power, no land use, no local grid queues, and a coolant — the vacuum itself — that never runs dry. That is the pitch Google, SpaceX, and Starcloud are all making, each in their own way. The honest version is that nobody knows whether the economics survive contact with reality. Launch costs, satellite manufacturing bottlenecks, and the engineering constraints of keeping thousands of chips alive in a radiation bath are formidable — and experts quoted in Reuters' coverage stress the concept remains years from commercial viability.

But the sequencing matters. Google is starting with the hardest, cheapest questions first: can the chips survive the ride, can they be kept cool, can they be kept honest in a radiation environment. If the answers are yes, the 2027 laser-link satellites will test whether a constellation can actually think as one machine. If the answers are no, Google learns it for the price of one rideshare slot instead of a constellation.

That is the right way to run a moonshot — and the reason this unglamorous test flight is worth watching. The space data center era, if it ever arrives, will have started the week a handful of Google TPUs took the elevator to orbit and phoned home to say how the ride went. The White House has been pushing hard on American AI leadership and infrastructure — and Suncatcher is what the private-sector half of that race looks like when it points its rockets upward.

When does the Suncatcher satellite launch?

October 1, 2026, on SpaceX's Transporter-18 from Vandenberg — the current target; schedules can move.

What chips is Google flying?

Four Trillium TPUs — about one server's worth of a regular data center, on ~1 kW of solar power.

Is this a working space data center?

No — an engineering test: 15-minute compute bursts, then cooldown. One-year design life; deorbits in ~6 years.

Sources

Sources: Reuters: Google plans first test of AI chips in space under Project Suncatcher (Sept 24, 2026) · RuntimeWire: Google will test its TPUs in orbit on SpaceX's Transporter-18 (citing Google's technical update, Sundar Pichai's Sept 24 post, and Ars Technica) · Inc.: Google Is About to Launch Project Suncatcher (background, Manyika/NYT quotes) · TechCrunch: Starcloud raises $170M Series A (March 2026) · Dawn: Google plans first test of AI chips in space under Project Suncatcher · The Hindu BusinessLine: Google to test AI chips in space with first Project Suncatcher mission