Google puts a TPU in orbit as its space data center plan meets a 1,800-launch hurdle

Google has launched an advanced Tensor Processing Unit into orbit for the first time, beginning a real-world test of the chips at the center of its Project Suncatcher plan for orbital data centers. The mission is intended to establish whether Google hardware can operate reliably in space before the company attempts larger, connected computing clusters.

The more immediate obstacle may not be the processor. Google’s newly published research estimates that SpaceX’s Starship would need roughly 1,800 launches over the next decade to move enough payload into orbit for launch prices to approach the level required by large-scale space computing. That projection highlights the gap between demonstrating a chip in orbit and deploying an economically viable data center there.

Google’s first orbital TPU test

The prototype satellite launched from California on a SpaceX rocket as part of a mission carrying more than 100 payloads. Planet Labs built the satellite platform, which will run a Google TPU under operating conditions that cannot be fully reproduced on Earth.

According to TechCrunch’s report on the mission, the satellite must supply about one kilowatt of continuous power, manage heat and cooling, and run a series of models while engineers monitor failures. Once commissioned, the TPU will operate in 15-minute bursts to limit pressure on the satellite’s power and thermal systems.

Travis Beals, the Google executive overseeing Project Suncatcher, told TechCrunch that ground testing cannot replicate every condition of the orbital environment. The mission therefore functions as a hardware and systems test, not as a production AI service.

Google and Planet Labs are planning a second demonstration for next year involving two satellites designed more specifically for advanced computing. Those spacecraft are expected to exchange data through laser communications, allowing Google to test whether separate orbital processors can cooperate on more demanding workloads.

Google’s longer-term concept calls for 81 satellites flying in close formation and processing workloads in parallel. Beals said communication bandwidth and latency between TPUs will be important if the system is eventually asked to handle workloads comparable to those spread across multiple racks on Earth.

The launch economics behind the 1,800-flight estimate

Google’s peer-reviewed orbital data center paper, reported by TechCrunch and scheduled for publication in Joule, examines how launch costs could change over time. The researchers explicitly say the work is not an economic feasibility study, so its cost projections should not be treated as a business case for orbital data centers.

The analysis assumes that SpaceX can continue a launch-cost learning curve of approximately 20% per year, based on the company’s history since Falcon 1. Under that assumption, launch prices could approach $200 per kilogram by 2035.

Reaching that figure would require Starship to carry about 370,000 metric tons of payload into orbit, the research suggests. At a maximum payload assumption of 200 metric tons per mission, that would amount to about 1,800 launches across 10 years, or 180 launches annually.

That is a demanding forecast for a vehicle that, according to the TechCrunch report, has not flown more than five times in a single year. SpaceX has forecast much higher flight rates, while Elon Musk has suggested Starship could eventually launch as often as hourly in 2029. Those targets remain projections rather than demonstrated operating performance.

The distinction matters for AI infrastructure planners. A terrestrial data center can scale by adding buildings, power equipment and network links to an existing site. An orbital facility must first pay to launch every processor, radiator, power system and communications component, then maintain coordination among spacecraft that cannot be repaired as easily as equipment on the ground.

Radiation appears manageable for inference, not every workload

The first satellite is also testing whether Google TPU hardware can tolerate radiation over the expected five-year life of an orbital platform. Google repeated particle-accelerator tests after discovering that the original chip configuration had more shielding than the version likely to operate in space.

The revised testing produced somewhat more logic errors, but Beals told TechCrunch that Google remains confident the chips can support substantial inference workloads. He characterized the error rate as roughly one in a million for typical inference operations, while warning that the same level of reliability would be problematic for very large training runs involving thousands of processors over months.

Those figures are statements from Google’s project leadership, not an independently validated production benchmark. They nevertheless define an important boundary for Project Suncatcher: orbital computing may initially be better suited to inference, data processing and other workloads that can tolerate occasional errors than to long, tightly synchronized training jobs.

The planned laser link between future satellites introduces another unproven element. Low-latency coordination is essential when many processors act as one system, but the available evidence covers a planned demonstration rather than a deployed multi-satellite cluster. Until that test is complete, the performance of a distributed orbital TPU system remains uncertain.

Google is also tackling the terrestrial power bottleneck

Google’s space research arrives as the company works with NVIDIA and Emerald AI on a separate infrastructure problem: how to connect expanding AI facilities to constrained electricity grids. The three companies announced the AI Energy Management Alliance, or AEMA, according to an NVIDIA Blog post.

AEMA’s focus is flexible data centers that can alter electricity consumption in response to grid conditions. The alliance describes several possible mechanisms, including shifting computing workloads, using storage, coordinating paired generation and reducing consumption during system emergencies.

The NVIDIA announcement says the coalition will pursue common technical requirements, performance measurements, operational data sharing and clearer obligations for ride-through, curtailment and contingency response. It also aims to help utilities assess whether a proposed facility can provide measurable flexibility rather than behaving only as a large, fixed electricity load.

This terrestrial initiative gives the space project useful context. Google is exploring an eventual way to place compute beyond local grid constraints, but its near-term infrastructure work is aimed at making conventional AI factories easier to connect and operate. AEMA’s claims about faster interconnection and infrastructure benefits are objectives of the alliance, not independently demonstrated adoption results; the announcement does not identify completed utility deployments or quantify delivered savings.

For enterprise buyers and AI builders, the two efforts point to different deployment trade-offs. Orbital systems could eventually offer persistent solar power and a location outside crowded terrestrial sites, but they introduce launch dependency, radiation risk, thermal limits and difficult maintenance. Flexible terrestrial facilities are closer to deployment, but they require software, storage and operating procedures capable of adjusting workloads without damaging service-level commitments.

What to watch next

The next important signal will be whether Google’s first TPU satellite completes its planned operating tests and publishes measurable information about errors, thermal behavior and workload performance. Results from the two-satellite Planet Labs demonstration will be more significant because they should test inter-satellite coordination rather than a single processor in isolation.

Launch cadence is another critical indicator. SpaceX would need to move from a small number of annual Starship flights to a sustained, high-frequency operation before Google’s 1,800-launch assumption becomes plausible. Demonstrated payload capacity, refurbishment time, launch licensing and actual cost per kilogram will matter more than aspirational flight-rate statements.

On Earth, observers should look for AEMA’s first technical specifications, utility partnerships and evidence that flexible AI facilities receive different interconnection treatment. Those developments will show whether the alliance can turn broad principles into operating rules that developers and grid operators can verify.

Creati.ai perspective

Google’s orbital launch is meaningful because it converts Project Suncatcher from a paper architecture into a hardware experiment. But the test does not yet validate an orbital data center business. It validates only that a TPU can begin operating in space under controlled conditions, while the larger system still depends on unproven launch economics, satellite networking and maintenance assumptions.

The nearer-term competitive story is on the ground. Google’s participation in AEMA acknowledges that AI infrastructure is constrained by electricity access and grid reliability today, while its space program targets a possible answer several years further out. For builders and enterprise buyers, flexible terrestrial capacity is the actionable development; orbital compute should be treated as a high-risk research path until launch cadence and multi-satellite performance are demonstrated.