Google's Space TPU Is Up. The Starship Math Says About 1,800 Flights Come First.

Google's first orbital computing satellite is in space, and the paper behind it puts a number on what comes next: roughly 1,800 Starship launches before space-based data centers could reach competitive launch costs. It is a useful reality check on the orbital data center hype, and it comes with large "ifs."

What Launched

Google launched its prototype Project Suncatcher satellite on Oct. 1, 2026, aboard a SpaceX Falcon 9 on the Transporter-18 rideshare mission from Vandenberg Space Force Base, per Tech Times. The satellite carries four Trillium-generation TPUs, the same type of chip Google deploys in terrestrial data centers, and was built in partnership with Planet Labs, which developed the spacecraft bus.

The design is deliberately modest. Tech Times reports the chips will run Gemini inference in bursts of about 15 minutes before the satellite must pause computing and let its radiators shed accumulated heat. TechCrunch reports a second-generation demonstration with two purpose-built satellites is planned for next year, linked by lasers. Google's Travis Beals called Suncatcher a "long-term moonshot."

1,600 or 1,800? The Number

The figure to use is about 1,800. TechCrunch's original headline said 1,600, and the outlet later corrected it to 1,800 in an editor's note. Google's own Suncatcher paper states: "reaching <$200/kg by ~2035 would require launching ~370,000t additional cumulative mass, equivalent to ~1800 Starship launches."

The arithmetic checks out. At roughly 200 metric tons of payload per Starship, 1,800 flights is about 360,000 tons, in line with the 370,000-ton figure. Spread over a decade, TechCrunch puts it at about 180 launches a year, or roughly one every two days. It also notes that Starship has never flown more than five times in a single year.

What the Number Means, and What It Doesn't

The paper's framing is a cost curve, not a launch order. It assumes launch prices keep falling at a rate of about 20% for every doubling of cumulative mass launched, a "learning rate" borrowed from manufacturing economics. FourWeekMBA's reading of the paper is that 1,800 launches is the cumulative mass consistent with sustaining that learning rate to reach about $200 per kilogram, and that the paper treats the target as ambitious. For context, that analysis cites about $3,600 per kilogram for reusable Falcon 9 today, roughly 18 times higher than the target.

The paper's economic test is the cost of power. Using a Starlink v2 mini-class satellite (575 kg, about 28 kW), FourWeekMBA reports the target price works out to about $810 per kilowatt-year, which falls inside the $570 to $3,000 per kilowatt-year range of terrestrial power costs. Only two of the four satellite designs the paper tested clear that bar. The paper also says it does not constitute a full economic analysis and leaves out compute hardware, networking, and maintenance.

Radiation Was the Worry. Heat Is the Problem.

Google's paper reports radiation tests on the TPUs that were reassuring: high-bandwidth memory was the most sensitive component, yet it survived total ionizing dose testing to 2 krad(Si), well above the 750 rad(Si) the mission required, and no hard failures were attributed to total dose up to 15 krad(Si). Tech Times argues that radiation is no longer the binding constraint and that thermal management is, which is why the first satellite is built around a 15-minute duty cycle. Whether heat can be rejected in vacuum at scale is what the mission is meant to find out.

A Note on Sources

Coverage describes the analysis as a peer-reviewed paper in the journal Joule. The version we were able to retrieve and read in full is Google's own Project Suncatcher research paper (marked © 2025), which contains the 1,800-launch figure. We could not retrieve the Joule article itself, so the figures above come from Google's paper and from reporting on it.

Why It Matters

• The hype now has a launch cadence attached. About 180 Starship flights a year, for a decade, is far beyond anything flown so far.
• The math is conditional. The paper's 1,800 is tied to a 20% learning rate and a $200/kg target, and it excludes most non-launch costs.
• Cooling is the open engineering question. The first prototype runs in 15-minute bursts for a reason.
• Starship's cadence is now an AI story. Our coverage of Starship's first orbital flight and our explainer on space-based data centers cover the vehicle and the case for putting compute in orbit.