AI Industry & Business

Google’s Orbital Data Center Wrote a Haiku, but Its Own Energy Chief Says Space Costs Up to 100x Earth

Google's Suncatcher satellite ran Gemma 4 in orbit. The same company's energy chief says space data centers cost up to 100x Earth, so the economics are still the story.

By John Paul November 2, 2026 7 min read

Google Project Suncatcher team celebrating the first AI response from a TPU in orbit

Google says its Project Suncatcher prototype has become the first data center class TPU to answer a prompt from orbit. On 30 October it uplinked a request, booted a TPU, ran Gemma 4, and received a haiku back over Svalbard, Norway.

The milestone is real but narrow. The satellite carries four TPUs and about 1 kilowatt of power. Google can run one chip at a time for roughly 20 minutes, so this is a proof of concept, not an orbital cloud region.

The bigger news arrived a day later. The Information reported on 1 November that Google’s head of data center energy said space data centers could cost up to 100 times as much as those on Earth, and that launch prices would need to fall below about $200 per kilogram by the mid-2030s to make the math work.

Quick answer: Yes, Google put an AI model to work in orbit. No, that does not make space compute affordable or available to developers. The engineering milestone is confirmed by Google’s blog and Sundar Pichai. The 100x figure is a reported statement, not a published cost model, and the public summary does not show the calculation.

What Google actually confirmed

Google’s 30 October post is specific about the sequence. The team sent a prompt to the onboard Gemma 4 model, the satellite downlinked the answer during a ground station pass, and Google called it a “Hello World” moment for a data center class TPU in space. It will now test thermal stress and radiation, then move to a 2027 milestone with two interconnected satellites.

The prototype launched on 1 October aboard SpaceX’s Transporter-18 rideshare mission and was built with Planet. Google’s claim needs a qualifier. It is the first AI response from Google’s own data center class TPU in orbit, not the first AI inference in space. Starcloud ran an Nvidia H100 in orbit in 2025, as NPR noted. Google’s achievement is narrower and still meaningful: its own accelerator, its own open model, and a complete round trip from the ground to orbit and back.

A ‘hello world’ milestone for Project Suncatcher 🚀

After launching our prototype satellite a few weeks ago, we’ve now successfully sent a command up to orbit, booted up the TPU, ran the prompt, and received our first complete response back on earth – a haiku! pic.twitter.com/kOFmETUskm

— Sundar Pichai (@sundarpichai) October 30, 2026

The cost catch that landed the next day

The Information’s report is the colder story. It quotes Google’s head of data center energy saying orbital data centers could cost up to 100 times more than terrestrial ones today. It also says the US has enough power but permitting is the bottleneck, and that launch costs need to fall below roughly $200 per kilogram by the mid-2030s for space compute to approach parity.

That threshold is not new. Google’s own research paper projected a path to about $200 per kilogram by the mid-2030s, when space capacity might approach terrestrial energy costs on a per-kilowatt-year basis. The difference now is the source. It is the company’s own energy lead, speaking to a reporter, while Google is still flying the prototype.

Travis Beals, the Google senior director behind Suncatcher, had already told NPR in October that he does not expect space compute to be cheaper than Earth compute within five years. The 100x figure and the five-year caveat point in the same direction. The engineering is moving. The economics are not.

Space versus Earth: what the numbers actually say

The prototype and a terrestrial AI campus are not in the same class. The table uses Google’s published figures, TechCrunch’s reporting on the payload, and The Information’s cost report. It compares a test article with a commercial facility, not a like-for-like price sheet.

Factor	Project Suncatcher prototype	Terrestrial AI data center
Compute	Four Google TPUs, one active at a time	Thousands of accelerators across a campus
Power	About 1 kW	Grid-connected, megawatts per campus
Solar input	Up to 8x more than Earth at a suitable orbit, per Google	Grid mix, limited by weather and night
Cooling	Heat pipes and radiators, no air	Air or liquid cooling
Network	Ground station passes; laser links planned for 2027	Fiber and public cloud regions
Cost	Reported up to 100x Earth today	Baseline
Launch parity	About $200 per kilogram by the mid-2030s, projected	Not applicable
Maturity	One-year research mission	Commercial service

Confirmed versus not confirmed

Several claims are solid. Several are not, and the difference matters for product planning.

Claim	Status
Google’s TPU ran Gemma 4 in orbit and returned a response	Confirmed by Google’s blog and Pichai’s post
The prototype launched on 1 October on Transporter-18 with Planet	Confirmed by Google
Four TPUs, about 1 kW, one chip at a time for roughly 20 minutes	Reported by TechCrunch, not stated in Google’s 30 October post
Space data centers cost up to 100x Earth today	Reported by The Information; calculation not public
Launch costs below $200 per kilogram by the mid-2030s	Google modeling and a reported executive estimate, not a market price
Twin satellites with laser links in 2027	Google-stated plan, not yet flown
“First AI in space” in the broad sense	Not confirmed; Starcloud ran an Nvidia H100 earlier

The next phase of Project Suncatcher is scheduled for 2027.

We’ll send two satellites into orbit to test whether we can perform AI computations distributed across multiple satellites so that we can one day run more complex workloads. pic.twitter.com/AJNL9u2EzD

— Google (@Google) October 30, 2026

Why 100x is plausible, and why it is still a guess

Space does not remove the hard costs of computing. It moves them. A satellite needs a launch, a bus, solar arrays, radiators, radiation-tolerant packaging, ground stations, and a replacement plan when a chip fails. Google’s own facts page describes vibration loads of 50 to 100 g on components, proton beam tests at UC Davis, heat pipes and radiators for cooling, and laser links that must hold a connection while both endpoints move.

A terrestrial operator can replace a failed server in hours. An orbital operator may have to wait for the next launch and accept a degraded constellation in the meantime. The 100x number is best read as a ceiling, not a median. “Up to” can cover a worst case with low utilization, short satellite life, and expensive launch.

It does not tell us the cost per useful token, which is what developers actually pay. That comparison would include launch, insurance, replacement, ground segment, idle time, and the energy penalty of running one TPU at a time because the power budget cannot support four. Google has not published that number.

What it means for Indian developers

Nothing in this week’s news changes Gemini API pricing, Cloud TPU quotas, or the capacity available in Google’s India regions. Indian teams will keep renting terrestrial capacity in Mumbai, Hyderabad, Chennai, Singapore, or the United States. Orbital inference still needs a ground station and a terrestrial network path, so it does not bypass latency, data residency, or the last mile.

The useful signal is about power and permits, not satellites. India’s AI buildout faces the same land, power, and water constraints that make space look attractive in the first place. Space is a research option for the 2030s, not a shortcut for a startup in 2026.

For Indian hardware and aerospace suppliers, the shopping list is more interesting than the headline. Radiation-tolerant packaging, thermal systems, optical links, and satellite buses are the parts Google has explicitly not solved yet. Those are real engineering markets. Model serving is not one of them, at least not yet.

The earlier launch is covered in this site’s Project Suncatcher launch report. For the terrestrial side of the same cost problem, see the Nvidia DSX POD container data center, SpaceX’s GPU debt bet, and Google’s TorchTPU cost benchmark.

Frequently asked questions

Did Google really run an AI model in space?

Yes. Google says its prototype satellite ran Gemma 4 on a TPU and returned a complete response on 30 October. The company published the haiku and named the ground station pass over Svalbard. The sequence is confirmed by Google’s blog and Pichai’s post, though this is a company announcement, not an independent audit.

How much did the Suncatcher satellite cost?

Google has not disclosed the cost of the prototype, the launch, or the TPU payload. The 100x figure compares space data centers with Earth data centers, not the price of this one satellite. Any specific mission cost is a guess unless Google or Planet publishes it.

Is space data center compute cheaper than Earth compute?

Not today, according to Google’s energy executive as reported by The Information. The claim is up to 100 times more expensive, with launch costs needing to fall below about $200 per kilogram by the mid-2030s for parity to become plausible. Beals has also said he does not expect a cost advantage within five years.

When will Project Suncatcher be commercially available?

Google has not given a commercial date. The next milestone is two interconnected satellites in 2027 to test distributed AI compute and laser links. Commercial service would require a larger constellation, lower launch costs, and years of reliability data that do not exist yet.

Should Indian developers plan for orbital inference?

No. Treat it as a research program, not a procurement option. The near-term levers for Indian teams are still terrestrial: reserved capacity, smaller models, quantization, caching, and workload scheduling. Watch the 2027 laser test and the radiation data, not the haiku.