Google's Suncatcher launch puts one server's worth of TPUs through a cooling test in orbit

When finance forwards the Suncatcher story and asks whether orbit belongs in next year's capacity plan, the answer is on the spec sheet. MVP is about the size of a refrigerator and holds about as much computing as one data center server, according to Digital Trends ^[3]. A data center on the ground runs thousands of the same chips ^[4].

The pitch runs well ahead of the payload. Google says solar panels in the right orbit can produce up to eight times more energy than they would on Earth, Digital Trends reported ^[6]. Elon Musk and Jeff Bezos have pitched orbital data centers as an alternative to divisive terrestrial ones, according to Ars Technica ^[7]. SpaceX and Blue Origin said this year they are seeking regulatory approval for orbital compute clusters ^[8]. On the satellite itself, four TPUs share about 1kW of solar output ^[2]. Split evenly, each chip gets roughly 250 watts ^[21].

Much of what could go wrong was tested on the ground first. Google ran the TPUs through a proton beam at UC Davis's Crocker Nuclear Laboratory while they ran AI workloads ^[11]. "Initial results have shown that our Trillium TPUs hold up remarkably well, and can survive a radiation total ionizing dose greater than what they would receive during a five-year space mission," Travis Beals of Google said ^[10]. Digital Trends reported that a simple restart usually fixed any glitches ^[22].

Heat is what the flight actually tests. Digital Trends called cooling Google's biggest headache ^[12]. Space has no air to carry heat off, so Google built a layered path from the chips through a putty-like material, aluminum and copper to a radiator panel ^[12]. The team has tried it only in a thermal vacuum chamber so far ^[24]. "We'll see how our new TPU cooling system works in space and refine our designs as we learn more," Beals said ^[13]. When results come back, the first figure to check is whether the 15-minute run limit gets longer ^[5]. The reports do not say how long each cool-down lasts.

A single satellite cannot test what a cluster depends on. Satellites in a cluster have to talk to each other, and Beals said Google's lasers "need to operate at very high bandwidth over extremely short distances" ^[14]. He compared holding that link to "hitting a coin-size target from miles away while both points are in motion" ^[23]. "We'll test our work on this in 2027 when we put two satellites in orbit," he said ^[15]. Neither the spacecraft nor the launch matches what a full network would use. The bus is a Planet Labs build ^[16], MVP rides a rideshare ^[1], and Ars Technica expects the eventual networks to need expensive dedicated launches ^[17].

The sources disagree on how long MVP will work. Digital Trends says it will handle simple AI queries for a year and could stay in orbit for as long as six years before burning up on reentry ^[18]. Ars Technica says it will operate for just a few months ^[19].

For anyone fielding the capacity question, sort orbital compute claims on two axes: whether the hardware runs continuously or in bursts, and whether it links to other compute at high bandwidth or works alone. MVP is bursty and standalone ^[5]. It is an experiment. The 2027 pair is meant to prove the links, the other axis ^[15]. A continuous satellite with no links would be one remote server. Only hardware that is continuous and linked is capacity a planner can put a number on. I'd keep orbit at zero in capacity plans until a Google flight reaches that box. The cost of that choice is small for now, because Manyika's own timeline puts useful operation years out ^[20], though a team that ignores the 2027 result entirely will start late if the links work.