Google's Orbital TPU Just Answered Its First Prompt

Google's Project Suncatcher prototype generated the first AI response from a data-center-class TPU in orbit this week, running Gemma 4 on four chips drawing about 1 kilowatt. It proves the hardware boots and answers in space. It does not prove orbital data centers are economical.

In this post

• What actually came back from orbit
• Four chips, one kilowatt, short bursts
• Why the radiator decides everything
• Everyone else in orbit, compared
• India is flying the same bet smaller
• The two-hundred-dollar kilogram question
• What this means for you

Google's first AI chip in orbit answered a prompt this week. Engineers uplinked a request, the model on board wrote a haiku about being in space, and the poem came back down as the spacecraft passed Svalbard, Norway. That reply is the first AI output ever produced by a data-center-class Tensor Processing Unit outside the atmosphere.

What actually came back from orbit

The satellite launched on October 1 as one of 130 payloads on SpaceX's Transporter-18 rideshare. The Falcon 9 lifted off from Vandenberg at 11:32 a.m. Pacific, deployed everything over roughly 11 minutes, and brought its booster back for a 25th flight. Built with Planet, the spacecraft is about the size of a refrigerator.

Commissioning took around three weeks. Then came the first prompt, sent during a ground-station pass.

Blue marble below
Data streams in silent grace
Starlight feeds my core.

The ground station matters more than the poem. KSAT's site at 78 degrees north is the only commercial facility that can see a polar-orbiting satellite on all 14 of its daily passes. Point your antenna there and a sun-synchronous orbit stops being a downlink problem.

After the haiku, Google ran space-themed limericks and telemetry prompts that reported the health of the payload systems. A spokesperson told TechCrunch the satellite was operating nominally. That second part is the real news. A model writing poetry proves the software stack works. A model reporting its own thermal state proves you can instrument the thing.

Four chips, one kilowatt, short bursts

The payload is four TPUs with roughly the compute of one server, fed by solar arrays that produce up to 1 kW. Google's own comparison point for that figure is a household hair dryer. The model running on board is Gemma 4, the open-weight family Google shipped in April under an Apache 2.0 licence, in sizes from 2.3 billion to 31 billion parameters.

Spec	Prototype satellite
Launched	October 1, 2026
Built with	Planet Labs
Compute	4 TPUs, about one server's worth
Power	Up to 1 kW solar
Model	Gemma 4 (open weights)
Operating mode	One chip at a time, in bursts
Orbit	Sun-synchronous low Earth orbit
Mission length	About one year

Here is the constraint that shapes everything else. The spacecraft runs one chip at a time, in bursts. NPR and Google's own write-ups put a burst at about 15 minutes. TechCrunch reported roughly 20. The exact figure is less interesting than the pattern: run, then wait for the radiators to catch up.

That duty cycle was designed in before launch. It is not a fault. But it does mean the demo says almost nothing about throughput. One Hacker News commenter put the sceptical case bluntly: "this is not an orbital data center, this is just something that will test TPU cooling functionality in orbit for only 15 minutes at a time." Another shot back that a chip processing data on TPUs is an orbital data center by definition. Both are arguing about a word. Honestly, the sceptic has the better of the physics argument, and the enthusiast has the better of the definition.

Why the radiator decides everything

On the ground you dump waste heat into air or water. In vacuum there is neither, so heat leaves as infrared from a surface, and the area of that surface sets your ceiling. An H100 drawing 700 W needs roughly 1.4 square metres of radiator, according to ABI Research's Andrew Cavalier. A 40 kW rack needs around 80 square metres.

Do the arithmetic at realistic temperatures. It is unforgiving. At 20°C with an emissivity of 0.9 you get about 600 W per square metre net in low Earth orbit once Earth's own infrared glow and reflected sunlight are counted. A single high-density AI rack drawing more than 100 kW maps to nearly 170 square metres. One rack.

Google's team has tested heat pipes and radiators in a thermal vacuum chamber, and the peer-reviewed paper in Joule names thermal management as a critical challenge. It does not publish the radiator area, the radiator mass, or the mass per unit of heat rejection. Independent analysts have noticed. Angadh Nanjangud's teardown of the paper makes the point that all the input power becomes heat, so the thermal design is not a detail, it is the system.

Radiation looks more solved than expected. Google ran a Trillium v6e TPU under a 67 MeV proton beam at UC Davis while it executed AI workloads. It survived a total ionizing dose of 15 krad with no hard failures, about twenty times the 750 rad a five-year mission needs. Memory was the soft spot: uncorrectable errors began around 2 krad, which works out to roughly one failure per 10 million inferences. Fine for inference. What it does to a long training run is still unmeasured.

There is a quieter risk worth watching. If the run window shrinks over the year, that means the thermal path is degrading. That would cost more compute than any bit flip.

Everyone else in orbit, compared

Suncatcher's architecture is the odd one out. Google wants 81 satellites inside a 1 km cluster radius, flying a couple of hundred metres apart at 650 km in a dawn-dusk orbit, talking over free-space lasers. A lab demonstrator moved 800 Gbps each way, 1.6 Tbps in aggregate, against a target of 10 Tbps per link. Everyone else is mostly putting inference on single spacecraft.

Player	Hardware	Power	First flight
Google (Suncatcher)	4 TPUs, laser-linked cluster planned	~1 kW today	Flew Oct 2026
Starcloud	Nvidia H100, then Blackwell	~8 kW on Starcloud-2	Nov 2025
SpaceX (Starmind AI1)	Nvidia Vera Rubin NVL72	175 kW average, 250 kW peak	Q4 2027 target
Pixxel + Sarvam	Datacenter-class GPUs	Not disclosed	Q4 2026 target
TakeMe2Space	Nvidia Orin NX	~5 kW target	Flew alongside Google

Starcloud got a GPU up first. Starcloud-1 flew in November 2025 with an H100 and 80 GB of memory, ran a version of Gemma, and trained nanoGPT on the complete works of Shakespeare. A second Nvidia chip on the same satellite failed during launch. Its bigger follow-up slipped into next year.

SpaceX is aiming far higher. Starmind AI1 is a 75-metre wingspan satellite with 175 kW of average compute power and a stated efficiency of 75 kW per tonne. Musk says the first ones fly in the fourth quarter of next year. The company has asked the FCC for up to a million data-center satellites; that filing drew 1,500 objections, and Amazon warned it would need 550 replacement satellites every day.

One awkward detail ties the two together. Alphabet holds a stake in SpaceX worth more than $82 billion, which makes Google a part-owner of the company whose rocket it needs and whose constellation competes with its own.

India is flying the same bet smaller

India is not waiting. Pixxel and Sarvam announced Pathfinder in May: a satellite in the 200 kg class carrying datacenter-class GPUs plus a hyperspectral camera, with Sarvam's models doing training and inference on board, targeting late this year.

TakeMe2Space put MOI-1A on the very same launch. It weighs under 50 kg and carries an Nvidia Orin NX. The company says it has 23 customers across agriculture, mining, supply chain and insurance, and it plans a 50 kW orbital data centre by 2029. Its first satellite was lost when ISRO's PSLV-C62 failed in January.

Indian effort	What it is	Target
Pixxel + Sarvam	GPUs plus hyperspectral camera	Late this year
TakeMe2Space	Small satellite, Orin NX, model hosting	5 kW now, scaling up by 2029
NeevCloud + Agnikul	Compute on Agnikul's upper stage	Pilot this year
ISRO	Feasibility study on in-orbit processing	Preliminary only

ISRO's position is worth quoting because it is the most honest one in the whole sector. Chairman V Narayanan told the Times of India: "As part of building futuristic technologies, we are evaluating data processing in space. At this juncture, only preliminary work has happened."

For anyone reading this in Bengaluru or Hyderabad, the near-term business is not training frontier models above the atmosphere. It is filtering Earth-observation data before it hits a congested downlink. A radar satellite can produce about ten gigabytes of raw data every second. Sending down an answer instead of the raw frames is a real saving, and it does not need cheap launches to make sense.

The two-hundred-dollar kilogram question

Google's paper projects launch to low Earth orbit at $200 per kilogram or less by the mid-2030s, down from the low thousands today. Independent modelling by Tim Harper puts that in perspective, and no surprise, it lands well below the headline number. At that price a reference design weighing 41.4 tonnes per megawatt costs $8.28 million per megawatt to launch, and about $8 million per megawatt to build the spacecraft around the computers. A typical terrestrial shell and core runs between ten and fourteen million dollars per megawatt.

So space is not cheaper at that number. It simply stops being obviously absurd. Harper's model suggests broad competition needs something closer to $50 to $150 per kilogram, plus high utilisation and several reliable years in orbit.

Reaching the Google figure needs about 370,000 tonnes of cumulative launch, or roughly 1,800 Starship flights at two hundred tonnes apiece. Starship had completed 13 integrated flight tests as of September and has not yet demonstrated routine commercial delivery and reuse. Everything in this category is a bet on that rocket working, cheaply, on schedule.

Two more things get less airtime. A Saarland University paper called "Dirty Bits in Low-Earth Orbit" estimates an orbital data centre could produce an order of magnitude more emissions than a ground one once launch and reentry burn-up are counted. And Google's own leadership has been unusually direct: "We don't expect, to be perfectly frank, that we'll have anything usefully operational in the next few years."

What this means for you

Look, your cloud bill does not change because of a haiku. What changed is what is now provable. Two years ago this was a preprint with a cost curve. Today there are four chips in orbit that answered a prompt, and a year-long mission that will return thermal and radiation data nobody on Earth has.

Watch three signals. First, whether Google publishes real operating numbers from the year-long run instead of a second press release. Second, whether the twin pair holds a laser link at close range while both spacecraft move. Third, whether anyone buys TPUs for orbit rather than building their own silicon. ABI Research forecasts 1.5 GW of effective orbital compute by 2035, which is a rounding error next to terrestrial capacity and a serious industry if it lands.

My read: the poem is a genuine engineering first and a good press moment at the same time. Both can be true. What it is not is a data centre. If you are planning infrastructure spend over the next two years, plan it on the ground.

People Also Ask

Did Google really put a data center in space?

No. The prototype carries four TPUs, draws about 1 kW, and runs one chip at a time. Google calls it a research prototype on a one-year mission, not a service.

How fast are the links between these satellites?

In a 2025 lab demonstration, a single optical transceiver pair moved 800 Gbps in each direction, 1.6 Tbps combined. The in-orbit laser test between two satellites is scheduled for 2027.

Why put AI in space at all?

Sunlight. In a dawn-dusk sun-synchronous orbit a solar panel sees near-continuous sun, which Google says yields up to eight times more power than the same panel on the ground, with no land, water or grid connection to negotiate.

Will orbital compute replace ground data centers?

Not on any near timeline. Launch cost, radiator mass, radiation damage and the impossibility of repairing a failed chip all favour the ground for the workloads you and I use.

Key Takeaways

• An off-the-shelf TPU booted, ran an open model, and answered from orbit. That is a real first.
• The mission is a hardware stress test in space, not a cloud region with a bill attached.
• Cooling sets the ceiling. Compute was never the hard part.
• The economics turn on launch prices falling roughly tenfold, and on Starship delivering.
• India's version of this bet is small, specific and already flying on the same rockets.

You might also like

• Nvidia's $4,999 DGX Spark: 100B AI Models on Your Desk
• Meta Runs Google TPUs in Production: Nvidia's Moat Cracks
• Oracle Puts Google TPUs on OCI: The Chip Menu Is Open

📚 Sources

#	Source	Used for	Link
1	Google Blog (Suncatcher first response)	Haiku text, Svalbard pass, first TPU response in orbit	blog.google
2	Google Blog (Behind Project Suncatcher)	8x solar figure, vibration loads, proton beam test	blog.google
3	Google Research / Joule paper	Launch cost projection, 81-satellite cluster, radiation dose figures	research.google
4	TechCrunch	Four TPUs, 1 kW, 20-minute bursts, Starship flight count	techcrunch.com
5	NPR	Short run windows, refrigerator size, thermal limits	npr.org
6	The Verge	Confirmation of first response, twin-satellite plan	theverge.com
7	SpaceX mission page	130 payloads, launch time, booster 25th flight	spacex.com
8	CNBC	Alphabet's $82B SpaceX stake, launch details	cnbc.com
9	Tim Harper analysis	Per-megawatt cost model, launch price threshold	timharper.net
10	ABI Research	Radiator sizing, 1.5 GW orbital compute forecast by 2035	abiresearch.com