Google’s First AI Data Center Satellite Is Online, and Its First Job Was Writing a Haiku

Google's Suncatcher satellite is in orbit testing a single TPU, and engineers' first real test was asking it to write a haiku.

Get a summary in:
View of Earth from space at night with city lights visible
Follow Teck Hustlers on Google
Add Teck Hustlers as a preferred source to see more of our stories in Google Top Stories.
Add as preferred source on Google

Google’s first AI data center satellite is online in orbit, and the first thing engineers had it do was write a haiku about being in space.

TL;DR: The Suncatcher prototype launched earlier in October on a SpaceX Falcon 9 alongside more than 100 other spacecraft, and it’s now communicating with ground teams during a three-week testing campaign. It carries a single Google Tensor Processing Unit that generates up to 1 kilowatt of power, enough to run for only about 20 minutes at a stretch. This week engineers prompted the satellite’s onboard Gemma 4 model for a haiku, and it wrote one about streaming data and starlight. It’s a small, slightly whimsical milestone inside a genuinely long-shot bet on running AI computing from orbit.

Illustration of a satellite in orbit above Earth

The actual haiku, for the record

Engineers sent Gemma 4, preloaded on the satellite, a simple prompt asking for a haiku about being in space. It wrote back:

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

The team also tried a few space-themed limericks and asked the satellite to report back on the orbital conditions around it, treating the exercise as much a communications test as a creative one. Getting any output at all back down to the ground, successfully, through an active link is the actual achievement here. The poetry is just the fun part people will remember.

Why a satellite can only “think” for 20 minutes

The hardware constraint is the real story. A single TPU generating up to 1 kilowatt sounds modest because it is modest, especially compared to the megawatts a single rack in a terrestrial AI data center can draw. That power ceiling means the chip can only run for roughly 20 minutes before it needs to cool down or recharge, which complicates something as basic as sending a prompt and getting a response back when the satellite has to manage its own power budget around every request. Engineers are currently three weeks into testing the satellite’s various subsystems, with more work ahead to see how the TPU holds up under thermal stress and in the radiation environment of low orbit.

Radiation is a genuinely tricky problem here. It can cause what engineers call “bit flips,” where a stray particle corrupts a piece of data without tripping any of the usual monitoring tools that would catch an error on the ground. Google’s pre-launch testing suggested that’s a minor issue for straightforward inference tasks, but a bigger one for anything resembling training, which needs much more data integrity over much longer stretches of computation.

The actual point of this experiment

Nobody at Google thinks one satellite with one chip is going to meaningfully move the needle on AI compute. The real goal is proving out the idea of flying constellations of satellites close enough together to coordinate on larger inference workloads, essentially treating orbit as free, infinite real estate for data centers that never have to fight a local community over power hookups or water use. If this single Suncatcher prototype performs well through its testing campaign, Google plans to launch two more satellites with more advanced TPUs in 2027, this time with higher power generation and actual optical links between them to test coordination between multiple orbiting units.

Google isn’t the only one chasing this

A handful of other companies are working on versions of the same idea. Starcloud is betting on high-speed inter-satellite links so its spacecraft can split up larger workloads between them, which is the more ambitious version of what Google is testing with Suncatcher. SpaceX is reportedly also planning laser-linked data center satellites of its own, though the orbits involved are expected to be too far apart for the satellites to meaningfully share processing load with each other. Starcloud and another company, Cowboy Space, are focused more narrowly on running inference on individual satellites rather than networking them together, a more conservative approach than what Google and SpaceX are each separately chasing.

There’s a tangled corporate thread running through all of this too. Alphabet holds a meaningful investment stake in SpaceX, and both companies’ long-term orbital data center ambitions depend on Starship becoming fully and rapidly reusable, since nobody is launching a useful constellation of compute satellites on today’s per-launch economics.

Why bother with orbit when ground power is the real bottleneck

The timing of this experiment isn’t a coincidence. Terrestrial AI data centers are running into real limits on power availability, water for cooling, and local community pushback over grid strain, the same pressures showing up in stories about flooded construction sites and canceled utility hookups elsewhere in the industry this year. Orbit sidesteps all three problems at once: solar power is effectively unlimited up there, there’s no water to source, and there’s no neighborhood to object to the noise or the runoff. The tradeoff is that everything about building and maintaining hardware in space is dramatically harder and more expensive than doing the same thing on the ground, which is exactly what the economics below are about to make clear. On Earth, those same pressures are showing up in Nvidia weighing a financial backstop for a flood-prone data center site and in AWS’s own trillion-dollar spending ambitions, both of which are about squeezing more capacity out of land that has real neighbors and real limits.

The economics are still brutal

A peer-reviewed analysis of the Suncatcher concept published earlier in October estimated that SpaceX would need to fly Starship somewhere around 1,800 times to achieve the kind of cost reductions through repetition that earlier rockets saw over their operational lifetimes. That’s a staggering number of launches, and it’s the clearest sign that this is a 2030s infrastructure bet, not a near-term fix for the terrestrial data center capacity crunch that’s currently driving up chip and memory prices.

The bottom line

A chip writing haiku from orbit is the charming headline, but the real story is Google quietly testing whether space can ever be a practical place to put AI compute at all. The power constraints, the radiation risk, and the sheer number of rocket launches needed to make the economics work all point to the same conclusion: this is a moonshot in the literal sense, one that won’t pay off for years even if everything goes right.

Follow Teck Hustlers on Google
Add Teck Hustlers as a preferred source to see more of our stories in Google Top Stories.
Add as preferred source on Google
About

The Teck Hustlers Editorial Team covers technology, business, finance and gaming news for readers who want the story behind the headline. Every article is researched, written and edited by our in-house team, checked against company announcements, filings and reputable reporting, and updated when new facts emerge. We focus on what each development means for founders, professionals, investors and everyday users. Have a tip or a correction? Reach us through our contact page.

Up next: Reddit Is Pushing Users to Log In as AI Search Dries Up Its Google Traffic