Deep Dive — One TPU in orbit, 1,600 Starship flights to go

Google confirmed contact with its first orbital compute satellite on Thursday, and the machine is behaving as expected: a refrigerator-sized spacecraft built by Planet Labs, riding a SpaceX Falcon 9 out of Vandenberg on the Transporter-18 rideshare, carrying four of Google's Trillium TPUs — the same accelerators that sit in its ground data centers. It is the first time a Google TPU has left the planet, and it converts Project Suncatcher from a research blog post into hardware with a heartbeat, as we noted in this morning's brief on the confirmed contact.

The launch is the easy milestone to explain. The hard part is what Google published alongside it: a peer-reviewed paper, released Thursday in the journal Joule, that walks through what it would actually take to run data centers in orbit — and concludes that the rocket cadence the economics depend on is roughly two orders of magnitude beyond anything flying today. Most companies would have shipped the satellite and buried the spreadsheet. Google shipped both.

Why anyone is doing this at all

The pitch for orbital compute is a direct answer to the two walls the terrestrial build-out keeps hitting: power and neighbors. As we argued in Deep Dive — Power, not GPUs, now sets the pace of AI, the binding constraint on AI capacity moved from chips to sockets — interconnection queues, turbines, and public consent — and the consent part keeps getting louder: Amazon acknowledged this week that more than 100 data-center moratoriums are under consideration across the US. Solar power that never sits in a zoning hearing is an attractive counterfactual, which is why Elon Musk has filed with the FCC for an orbital data-center constellation of up to one million satellites and says space data centers will beat terrestrial costs within three years, and why Jeff Bezos keeps talking about them too. Google's answer, announced last year and pre-covered here in Google is putting its first TPUs in orbit next week, is the same moonshot with a test manifest attached.

What the satellite actually does

MVP is deliberately unimpressive as satellites go. Google did not build the spacecraft from scratch: the original plan was two purpose-built compute satellites in 2027, and the company instead integrated its chips into a Platform frame that Planet Labs had already constructed, buying a year. The solar panels supply about a kilowatt of continuous power — roughly a hair dryer — and the compute duty cycle is the giveaway. The TPUs run Gemini workloads in bursts of about 15 minutes, then shut down while the radiators catch up, because that is the speed heat can leave a passive panel in vacuum. The thermal path is a layer of malleable interface material bonded to aluminum and copper heat pipes that carry the heat out to a radiator. On the ground you solve this with fans, chillers, and a utility agreement; in orbit you get one shot at the physics.

The chips themselves are standard silicon, not radiation-hardened variants. That is the real experiment. Before launch, Google blasted them in a particle accelerator at UC Davis and gave them a radiation dose larger than a five-year orbital mission would deliver — and then re-ran the tests after realizing the ground configuration shielded the chips more than flight would, which produced slightly more errors in the logic circuitry. Project Suncatcher manager Travis Beals put the outcome plainly: the resulting error rate, roughly one flipped bit per million operations, is comfortable for inference but "problematic for doing, say, some mega-scale training run where you're going to have many thousands of chips running for months." Google's own conclusion is that a satellite fleet can be trusted with serving models for a five-year lifespan, not making them.

The spreadsheet Google published anyway

Here is where this week gets interesting for anyone who follows the industry's real constraints rather than its keynotes. The Joule paper is one of the most rigorous public analyses of how compute gets to orbit, and its central number is a launch-cost target: near $200 per kilogram by 2035, which the authors argue is plausible only because SpaceX has held a price-reducing learning curve of about 20 percent a year since the Falcon 1 days. Getting there requires Starship — SpaceX's still-new super-heavy rocket — to lift on the order of 370,000 tons of payload over the next decade. At the paper's assumed 200 metric tons per flight, that is roughly 180 Starship missions a year, every year, for ten years — a figure that rounds to the "about 1,600 flights" headline the story has been carrying since publication.

For perspective: Starship completed its first orbital flight only days before this launch, and it has never flown more than five times in a single year. Musk has suggested the vehicle could reach an hourly flight rate in 2029; Google's paper, which stresses it is not an economic feasibility study, needs the rocket industry to do something it has never done, at a scale it has never attempted, on a schedule set by a decade-long cost curve. Google is also a major investor in SpaceX, which is worth remembering when weighing whether its sober math is a warning or a roadmap.

What the skeptics say

The engineer's case against orbital data centers is older than this launch and gets sharper with numbers. IEEE Spectrum's editor-in-chief argued in July that the stars and the math will not align for space compute anytime soon: there are roughly 14,500 active satellites in orbit today in total, SpaceX's record year saw 165 orbital launches, and Starlink manufactures around 4,000 satellites annually — a million-satellite data-center constellation would take a decade even at ten times every one of those rates.

Then there is heat, the constraint MVP's 15-minute bursts demonstrate in miniature. IEEE Spectrum's July cover math, from ABI Research, puts a single 700-watt GPU at 1.4 square meters of radiator area at 60°C; a 40-kilowatt rack needs about 80 square meters of radiator, and a 100-megawatt data center would need 2,500 of them — radiator area that has to point at deep space, never at the Sun, while the whole assembly stays in a stable thermal orbit. The one company that has actually flown Nvidia silicon found the limit experimentally: Starcloud's single H100 mission came home with a radiator too weak to let the chip run at full power. And the workload question cuts the same way Google's radiation data did — the analysts who see a plausible path for orbital compute expect it to serve inference at the edge, while training, with its synchronization and latency demands across thousands of chips, likely stays on the ground.

Google is racing companies, not just physics. Starcloud, a startup that has applied to the FCC for an 88,000-satellite orbital data-center constellation, has flown exactly one Nvidia H100 so far; SpaceX filed its own FCC application for up to one million orbital data-center satellites and has been showing off designs for an AI-1 satellite data center while Musk claims cost parity with ground data centers in three years; and the broader field ranges from rideshare startups catching the same Transporter manifest as Google to infrastructure investors treating orbital compute as a hedge on terrestrial siting risk. The pattern is consistent: every player's public demo is a single-digit-chip experiment, and every player's roadmap is planetary.

The contrarian read is that none of this is really an engineering argument. Terrestrial data centers fail on politics — permits, moratoriums, zoning — while orbital data centers fail on physics and launch cadence. Google is betting that physics is the more tractable problem, because physics responds to iteration and a satellite program it controls, whereas a county board does not. The counter is that Google's own paper concedes the iteration depends on an external actor — SpaceX — hitting a flight rate two orders of magnitude past current reality. The critics' closing argument is an externality one: a million-satellite compute constellation piles onto debris-collision risk, brightens the night sky further, and competes with astronomy for the spectrum — the same public-consent problem, relocated to orbit.

What to watch next

The decisive data points are already scheduled. In 2027, Google and Planet Labs plan two purpose-built compute satellites that will attempt to talk to each other over a laser link — the step that turns isolated test boxes into something resembling a distributed system, and the one Beals says matters most because bandwidth and latency between TPUs determine whether multi-rack workloads can run in orbit at all. The target Google has floated internally is a formation of about 81 satellites processing in parallel. Beyond that, watch Starship's flight rate against the paper's 180-per-year requirement, any published duty-cycle telemetry from MVP showing whether 15-minute bursts are a thermal limitation or an interim design, and whether inference-only satellite fleets start appearing on customer roadmaps. Google's own timeline is candid: years before Suncatcher turns from project to product.

Should the orbital-compute question be settled by launch-cost curves or by who wants a data center next door? Tell us in the comments.