Google's First Orbital TPU Launches a Year Behind Nvidia's H100

Google says its first Project Suncatcher hardware launches to orbit within days, riding a SpaceX Transporter rideshare mission developed with satellite maker Planet, according to an update Google published Sept. 24. The mission's entire job is to find out whether a Google Tensor Processing Unit survives getting to space at all; it is not running any workload once it arrives.

That distinction matters because Google is not the first company to put an AI chip in orbit. Nvidia-backed Starcloud launched a satellite called Starcloud-1 on Nov. 2, 2025, carrying a single Nvidia H100, and by December had it running inference and returning responses from Google's own Gemma model, according to Data Center Frontier. Google's chip, arriving roughly ten months later, is still at the stage of proving the hardware works, not the software running on it.

The number that matters is $200 per kilogram, not eight times the sunlight

Project Suncatcher's pitch, laid out in a Nov. 4, 2025 preprint by Travis Beals, senior director of Google's Paradigms of Intelligence group, and eight co-authors, rests on a genuinely large physical advantage: a solar panel in the right low-Earth orbit can be up to eight times more productive than one on the ground, with near-continuous sunlight that cuts the need for batteries. That number gets repeated in every writeup of the project, including Google's own, because it is the most quotable figure in the pitch.

It is also not the number that decides whether any of this ships at scale. Google's own preprint puts that number at $200 per kilogram, the launch price the company's analysis says historical and projected pricing trends could reach by the mid-2030s, the point at which orbital compute becomes roughly comparable to a terrestrial data center's energy cost on a per-kilowatt-year basis. Every other figure in this story, the radiation tolerance, the vibration survival, the optical bandwidth, only matters if that cost curve actually bends the way Google's model assumes it will.

Radiation held up; a ten-minute rocket ride is a different test

Google tested its Trillium chip, the sixth-generation Cloud TPU also known as v6e, in a 67 MeV proton beam at UC Davis's Crocker Nuclear Laboratory, running live AI workloads while monitoring for bit-flip errors. The chip's high-bandwidth memory, its most radiation-sensitive component, only began showing irregularities above a cumulative dose of 2 krad(Si), nearly three times the 750 rad(Si) Google expects a shielded satellite to absorb over a five-year mission. No hard failures showed up at all up to the maximum tested dose of 15 krad(Si) on a single chip.

1. Expected 5-year mission dose (shielded)0.75 krad(Si)
2. First HBM irregularities appear2 krad(Si)
3. Maximum tested, no hard failures15 krad(Si)

Source: Google Research preprint, Nov. 4, 2025, accessed Sept. 24, 2026

Radiation, though, is only one of the failure modes a launch imposes. Google's Sept. 24 update says the rocket ride itself subjects a satellite to roughly 10 times the force of gravity sustained over about 10 minutes, with individual components, including the TPU chips themselves, seeing spikes of 50 to 100 g. Google says it shook a satellite on all three axes to mimic those launch frequencies and that the hardware held up, but the company's own language, calling itself "pleasantly surprised" that the test went as planned, is a tell that the outcome was not a foregone conclusion internally.

Cooling has no atmosphere to work with

A TPU generates concentrated heat in a small area, and on Earth that heat leaves through moving air. In orbit, Google's update notes, there is no airflow at all, so heat can only leave through radiators, a fundamentally different cooling architecture than any data center Google operates today. The company says it has tested a combination of heat pipes and radiators in a thermal vacuum chamber that simulates both the temperature swings and the vacuum of orbit, but has not yet run that cooling system in space itself, which is one of the specific things next week's flight is meant to establish.

Engineering hurdle	Test completed	What's still unverified
Radiation tolerance	Proton beam, UC Davis Crocker Nuclear Lab	Multi-year cumulative exposure in situ
Launch vibration	Three-axis shake test to 10g sustained, 100g peak	Actual rocket flight loads
Cooling	Thermal vacuum chamber	On-orbit radiator performance
Inter-satellite bandwidth	1.6 Tbps bench demo, single transceiver pair	Formation flying at kilometers-scale range

On the connectivity side, Google's research team says it has already demonstrated 800 gigabits per second in each direction, 1.6 terabits per second total, across a single bench-scale transceiver pair, the kind of bandwidth a modeled 81-satellite constellation would need at a mean altitude of 650 kilometers with satellites flying as close as 100 to 200 meters apart. None of that has been tested with satellites actually in relative motion in orbit; Google's own next milestone for that specific test is a two-satellite mission with Planet planned for early 2027, not this week's single-chip survival flight.

What would change the read

Starcloud's head start does not settle which architecture wins. A single H100 running inference on borrowed compute time is a different bet than Google's plan for gigawatt-scale constellations of purpose-built, tightly clustered satellites carrying dozens of TPUs each, and Google's preprint frames this week's launch as the first of several deliberate steps rather than a race to be first. What would change the read is which company reaches a genuinely useful compute density in orbit first, at a real dollar-per-kilowatt figure rather than a modeled one: Starcloud has an actual chip taking real queries today, while Google's number, $200 per kilogram to orbit, remains a projection for the mid-2030s. Google's own silicon has spent years compounding cost-per-token gains inside terrestrial data centers, including the Snapdragon-class edge chips and NPUs this site has tracked chasing similar efficiency gains on the ground; whether that compounding curve translates to orbit, or whether launch economics simply refuse to cooperate, is the question neither company's hardware survival test can answer yet. A more immediate tell will be whether Google's own reverse-engineered efficiency gains on terrestrial NPU silicon end up mattering more to near-term AI economics than anything scheduled to leave the atmosphere this decade.