On October 1, another Falcon 9 lifted off from Vandenberg on the Transporter-18 rideshare, carrying about 130 payloads at once. Among the cubesats and Earth-observation satellites was a fridge-sized spacecraft that Google internally calls simply MVP. Inside are four server-grade AI chips, the same ones that run Gemini models in Google Cloud.
The headlines say "Google launched a neural network into space", and the obvious question is: why? Neural networks work perfectly well on the ground. The answer is more interesting than it looks. Google is not testing a model, it is testing a hypothesis about data centers. And the main problem this satellite has already surfaced is not the one everyone worried about before launch.
What actually flew
The spacecraft is the first prototype of Google's research effort Project Suncatcher. The bus was built by Planet, the company that already operates hundreds of Earth-imaging satellites. Google is responsible for the payload.
| Launch | October 1, 2026, Falcon 9, Transporter-18, Vandenberg |
| Chips | 4 TPU Trillium (v6e), a v6e-4 configuration, roughly one ground server |
| Power | Solar panels, about 1 kW |
| Orbit | Sun-synchronous dawn-dusk, about 650 km |
| Lifetime | About a year of planned operations, up to six years in orbit |
| Status | Contact established, the spacecraft is "operating as expected" (Travis Beals, project lead at Google) |
The orbit is not random. In a dawn-dusk sun-synchronous orbit the satellite rides the terminator line, and the Sun almost never sets for it. By Google's own estimate, a solar panel there produces up to eight times more energy than the same panel on the ground: no night, no clouds, no atmosphere.
Radiation: an exam the chips passed on Earth
The obvious fear is space radiation. Server chips are designed for data centers, not orbit, and what usually flies is radiation-hardened electronics that lag generations behind in speed. If a commercial chip dies within a month, the whole idea dies with it.
So Google irradiated the chips first. At UC Davis's Crocker Nuclear Laboratory, Trillium TPUs were run under a 67 MeV proton beam, under real AI workloads rather than idle. The results were published as a preprint in November 2025 and later in the peer-reviewed journal Joule:
- 15 krad(Si): the maximum dose tested, with no permanent failures
- ~0.75 krad(Si): the dose a shielded chip accumulates over five years in orbit
- 2 krad(Si): the point where HBM memory started to show errors, the most vulnerable component
- 1 silent data corruption event: a wrong result that nothing flagged
That is a twentyfold margin over the five-year dose. Most bit flips are cleared by resetting the chip. The last item is the unpleasant one: an error that does not announce itself cannot be fixed by a reset. It has to be caught by checks in software. More on that below.
The real enemy is heat
Here is what looks like the main lesson of the first week. According to TechTimes, once in orbit the mission is defined not by radiation but by cooling. The chips run short Gemini queries in bursts of roughly fifteen minutes, after which compute has to stop so the radiators can shed the accumulated heat.
It feels counterintuitive: space is cold. But cold does not mean easy to cool. On Earth, air or water carries heat away from a server: convection. In a vacuum there is no convection. Only radiation is left, with the radiator glowing in the infrared. That is slow. Dissipating a kilowatt takes a lot of panel area, and every square metre is mass you have to launch.
In one line: energy in space is abundant, getting rid of heat is hard. An orbital data center is limited not by the power outlet but by the radiator.
A caveat: Google has not disclosed details of the duty cycle, and the fifteen-minute figure comes from TechTimes. But the physics behind it is unambiguous, and every orbital compute project runs into the same wall.
Why bother: the $200-per-kilogram economics
The motive is simple. Training and serving large models is bottlenecked by electricity: new data centers wait years for grid connections, and power systems cannot keep up with demand. In orbit the Sun shines almost around the clock, there is no land to pay for and no interconnection permits.
Google itself names the condition under which this works: launch costs must fall to about $200 per kilogram, roughly by the mid-2030s. Only then does an orbital data center match a ground one on energy cost. Today we are far from that number, and the whole bet rides on fully reusable super-heavy rockets like Starship.
Google's roadmap:
- 2026: one MVP satellite, to see whether a commercial TPU survives real flight
- 2027: two satellites with Planet, testing free-space laser links between spacecraft
- Later: clusters; the research models a swarm of 81 satellites within a one-kilometre radius, linked optically
Laser links are the second key question after heat. Large models train on thousands of chips that constantly exchange data. If satellites cannot talk to each other at speeds comparable to cables inside a data center, the cluster becomes a collection of loners.
Who else is in the race
Google is neither the first nor the only one. Over the past year several players have reached orbit, each with its own approach.
| Player | In orbit | Approach and plans |
|---|---|---|
| Google + Planet | Project Suncatcher: 1 satellite, 4 Trillium TPUs, ~1 kW (October 2026) | A decade-long research effort: two satellites with laser links in 2027, then clusters of 81 |
| Starcloud | Starcloud-1 with an NVIDIA H100 (November 2025) | First data-center GPU in space, first model trained in orbit. Raised $170M at a $1.1B valuation (March 2026) and $250M more at $2.3B (August 2026) |
| SpaceX | A filing so far | In January 2026 filed with the FCC for up to a million data-center satellites at 500–2,000 km, linked via Starlink |
| Axiom Space + Kepler | 2 orbital data center nodes (January 2026) | Processing other satellites' data in orbit, Kepler optical network at 2.5 Gbps |
| ADA Space + Zhejiang Lab (China) | 12 satellites of the "Three-Body Computing Constellation" (May 2025) | 744 TOPS per satellite, target of 2,800 spacecraft |
The approaches split into two camps. China and Axiom build compute for space: satellites process their own imagery and data in orbit instead of pushing raw terabytes down a narrow link. That is pragmatic and pays off today. Google, Starcloud and SpaceX aim at something else, compute in space for Earth: moving the data centers that train and serve models into orbit. That is the $200-per-kilogram bet.
Starcloud stands apart: the startup not only flew the first data-center GPU, it also ran Google's Gemma model on it. In a sense, Google was beaten to "a neural network in orbit" with its own model.
What the skeptics say
Not everyone shares the enthusiasm. Commenting to TechTimes, astrophysicist Jonathan McDowell of the Harvard-Smithsonian Center for Astrophysics pointed to the cost of deorbiting: thousands of burning-up satellites leave combustion products in the upper atmosphere, and nobody really knows how that affects atmospheric chemistry. Space debris expert Moriba Jah of the University of Texas warns that mass deployments raise the risk of Kessler syndrome, a cascade of collisions that renders orbits unusable.
There are plain engineering questions too. Solar panels degrade by 0.5–0.8% a year. A broken chip in orbit cannot be replaced, while a ground data center swaps hardware every few years as faster generations ship. An orbital cluster ages together with its chips.
My take
What interests me most here is not space but how it changes the requirements for software. When hardware can return a wrong bit at any moment and goes off to cool every fifteen minutes, the program has to survive that: checkpoint, re-verify critical computations, resume calmly after a reset. On Earth we are used to assuming a server "just works". In an orbital data center that assumption is gone, and that single silent data corruption from the lab test says it louder than any number of kilowatts.
The MVP itself is not a data center. It is a check that the basic assumptions hold in real flight. It has answered the first question: the chips are alive. The second, what it costs to throw away heat, turned out to matter more than it looked on paper. The third, launch price, is decided not at Google but on SpaceX's launch pads. A real data center in orbit is probably a decade away if everything goes to plan. But for the first time it is not a slide deck, it is hardware that flies and answers queries.
Originally published at klukyanov.ru.
Shorter weekly write-ups (in Russian) — on Telegram.
Top comments (0)