Google to Send TPUs into Space: Launching on a SpaceX Rocket to Test Space Data Center Concept

Google is preparing to send four TPUs into orbit. Housed in a prototype satellite built in partnership with Planet, they will launch aboard SpaceX's Transporter-18 rideshare mission. This test is part of Project Suncatcher: Google wants to find out whether harnessing nearly continuous sunlight in low Earth orbit for power, and linking many TPU-equipped satellites together, can support large-scale machine learning computing in space.

This satellite will not form a "space data center" just yet. It will first test whether four TPUs can survive a rocket launch and operate under orbital conditions of radiation, vacuum, and temperature.

Lining up the official materials provided this time in chronological order makes the evolution of the plan clearer:

• The November 2025 project launch post stated: Google and Planet planned to launch two prototype satellites before early 2027, simultaneously testing orbital hardware operation and inter-satellite communications.
• The new long-form post from September 24, 2026 added a single-satellite test ahead of the dual-satellite mission. According to the post at the time of publication, the first batch of TPUs would enter orbit "next week"; this mission aims to identify hardware failure points first and apply the data to subsequent missions. Here, "next week" was the launch window Google planned at the time, not an already-completed launch.
• Google's announcement post added the single-satellite payload count: the prototype satellite carries four TPUs. Sundar Pichai's post confirmed that this satellite, built in collaboration with Planet, will fly on the SpaceX Transporter-18 rideshare mission.

Therefore, this does not mean the originally planned dual-satellite test was completed ahead of schedule; rather, Google inserted a smaller on-orbit hardware test prior to the 2027 dual-satellite interconnect: first answering whether the TPUs and cooling system can work, before answering whether two satellites can interconnect at high speed.

This clip serves to explain why the team is willing to pursue this experiment, rather than prove that the concept is already validated. The video begins by admitting that the idea sounds "crazy," then grounds the rationale in solar energy: if solar power generation is to be expanded further, deploying in a suitable orbit becomes a direction worth testing. It concludes by stating clearly that projects like this carry risk and sometimes fail, but the experimentation process may yield unexpected new findings. Thus, the video supports the research motivation and the value of trial and error; whether the four TPUs can operate in orbit must still be answered by data from this mission.

Why Move AI Compute into Space?

The reasoning starts with power supply. In Low Earth Orbit (LEO), satellites in specific orbits can receive near-continuous sunlight. Google estimates that under such lighting conditions, solar panels of equivalent surface area can generate up to 8 times the power produced on Earth.

The official video specifies what "specific orbit" means: the team compared various options, including geostationary orbit, and determined that a dawn-dusk orbit is better suited for this concept because solar arrays there can receive near-constant illumination. The video cites a power generation range of 5 to 8 times that of Earth for the same panel area. Therefore, "8 times" represents the upper bound for solar panels in a selected orbit under ideal lighting conditions, and should not be interpreted directly as the net energy advantage of an entire space data center.

If multiple solar-powered satellites can be connected via high-speed networking in the future, they could collectively handle AI computing tasks beyond the capacity of an individual satellite. However, "8 times" only reflects the power generation potential from sunlight; it does not answer whether chips can operate long-term in orbit, whether satellites can exchange sufficient data, or whether the entire system is economically viable to build. This single-satellite test addresses only the first question.

The First Satellite's Immediate Question: Can TPUs Operate in Orbit?

TPUs originally operated in data centers where temperature and power delivery are tightly controlled. Once sent into orbit, they must endure rocket vibrations, high-energy particles, and a vacuum environment where heat cannot be dissipated by air.

This single-satellite mission will record how the chips perform when subjected to launch stress, radiation, and temperature swings, which will then inform hardware and mission design revisions for the subsequent dual-satellite test.

Prior to launch, the team conducted vibration testing, proton beam irradiation, and thermal vacuum testing:

Launch Vibration: Can the Chips Survive the Ascent?

A rocket takes roughly 10 minutes from ignition to reaching low Earth orbit. During this period, the spacecraft experiences violent vibrations, enduring sustained acceleration loads of up to 10g overall; individual components, including the TPUs, can experience 50 to 100g.

• Ground testing approach: The team placed the entire satellite on a vibration table and shook it along three axes to simulate the vibration frequencies of a rocket launch. This test also evaluated whether the custom packaging designed by Planet could prevent components from sustaining damage during ascent.
• Current conclusion: Google stated that the hardware survived the three-axis vibration testing.

Orbital Radiation: How Will Compute Errors Affect the Mission?

Once in orbit, high-energy particles from solar activity and cosmic rays pose radiation risks to electronic devices. When high-energy particles penetrate semiconductors, they can cause "bit flips"—abnormal state changes in storage or compute cells (such as flipping from 0 to 1 or 1 to 0). The video divides radiation concerns into two categories: single-event upsets caused by individual particle strikes, which can typically be corrected; and Total Ionizing Dose accumulated over long-term exposure, which gradually degrades the devices.

• Ground testing approach: The team subjected TPUs to beam irradiation at the UC Davis Crocker Nuclear Laboratory proton beam facility, running AI workloads on the chips while closely monitoring how bit-flip-style errors specifically impacted computational loads.
• Current conclusion: Preliminary ground test results showed that the tested Trillium TPUs demonstrated robust tolerance, withstanding a Total Ionizing Dose exceeding the theoretical cumulative dosage for a satellite on a 5-year space mission.
• Errors observed: Engineers in the video noted that the team observed Silent Data Corruption (SDC) for the first time during beam testing. SDC is harder to detect than outright crashes: the system continues running, but computational results are silently corrupted. This is why real AI workloads were run continuously throughout the irradiation tests.

This testing also involves an engineering trade-off: if commercial TPUs combined with custom packaging, error correction, and monitoring can meet mission requirements, the team can rapidly adopt advancements from the terrestrial semiconductor industry without repeating lengthy space-grade qualification cycles for each chip generation. Current testing indicates only that this pathway is worth pursuing; it has not proven that standard commercial hardware can fly into space without modification.

Vacuum Heat Dissipation: Where Does Chip Heat Go?

Terrestrial data centers typically rely on fans blowing cold air or liquid cooling systems to carry heat away, but space is a vacuum with zero airflow, rendering all convective cooling ineffective. TPUs generate substantial heat across a compact die area that must be safely rejected to avoid severe overheating risks.

• Thermal dissipation mechanism: In the vacuum of space, the ultimate mechanism for rejecting heat outward is emitting infrared radiation into the cold cosmic background via thermal radiators. Inside the satellite, however, heat generated by the chips must first be routed from within the chassis to external radiator panels using heat pipes or pumped fluid loops.
• The challenge is scale, not physics: Radiative cooling is a mature aerospace technology. Engineers in the video cited an order of magnitude: radiator panels can only dissipate a few hundred watts per square meter, whereas a palm-sized TPU produces several kilowatts of heat, meaning each TPU could require several square meters of radiator area. This is one reason why the long-term vision of "dozens of TPUs per satellite" cannot simply be shrunk into standard satellite dimensions.
• Ground testing approach: The team placed the integrated payload into a Thermal Vacuum Chamber, simulating both vacuum and thermal conditions simultaneously while running functional tests to observe the response of the entire thermal path; actual orbital performance remains to be observed during this mission.

Vibration tables, proton beam facilities, and thermal vacuum chambers provide only ground-test results. A real orbit imposes all these conditions simultaneously on the same hardware. This launch seeks to gather on-orbit data to uncover failure modes that ground testing did not reveal.

From a Single Satellite to Constellation Interconnect: What Remains to Be Solved?

A single satellite cannot accommodate enough TPUs. To process larger AI workloads, multiple satellites must exchange data and compute collaboratively just like servers in a data center. Google envisions future satellites carrying dozens of TPUs each, with multiple satellites networked into an orbital cluster.

According to the plan, this research will advance in phases:

• This single-satellite mission: Observe the orbital performance of TPUs after enduring the challenges of launch, radiation, and thermal management.
• The planned 2027 dual-satellite mission: Test short-range, high-bandwidth laser links between two satellites.
• Long-term vision: Form clusters of multiple satellites, each carrying dozens of TPUs, to jointly undertake larger computing tasks.

Multi-satellite collaboration relies on inter-satellite laser interconnects, but existing systems cannot directly meet the demands of such compute clusters:

• Bandwidth and latency approaching data center standards: Google's video explains that thousands of chips in a future cluster must communicate both within individual satellites and across satellites; both demand high bandwidth and low latency. Most existing space laser systems target long-distance, lower-bandwidth communication, whereas compute clusters require continuous exchange of massive volumes of data over relatively short distances.
• High-precision relative positioning requirements: To guarantee the communication bandwidth required for AI computing, each satellite must know its own position and its position relative to neighboring satellites in real time, maintaining high-speed data links via lasers.
• Extreme difficulty of laser alignment: Maintaining such high-bandwidth links in orbit demands extraordinary pointing accuracy. Google describes it as: hitting a coin-sized target from several miles away while both ends are moving.

Google has completed a ground-based proof of concept: the team slightly modified off-the-shelf fiber-optic transceivers so that instead of driving optical fibers, they drive two telescopes, achieving a bidirectional link of 800 Gbps in each direction during testing. This figure demonstrates that the combination of transceivers and optical paths can transmit data in a laboratory setting; it does not mean that two satellites in motion can already maintain 800 Gbps long-term in orbit.

This laser interconnect technology is planned for field testing when two prototype satellites are sent into orbit in 2027.

What to Watch for After This Launch?

A successful launch only proves that the satellite reached orbit. The next focus is whether the four TPUs can operate continuously, how radiation errors affect AI workloads, and whether the thermal structure can keep temperatures in check. This data will be used to refine hardware for the next iteration. The dual-satellite mission will then test whether two continuously moving satellites can sustain stable, high-bandwidth laser links over the long term.

Even if both test phases meet expectations, it would only demonstrate that critical engineering milestones have cleared early validation. Scaling a handful of prototypes into sustainable, maintainable, and economically viable AI infrastructure remains an open question for subsequent research to address.