Google, get ready for lift-off

Google is about to send its computing infrastructure into space.

Sundar Pichai announced in a long public post that the first in-orbit test of the Suncatcher project will be launched on October 1, with total computing power roughly equal to that of a single ground-based server.

Google first publicly introduced the Suncatcher project back in November last year.

It is defined as Google's "long-term research initiative" that aims to explore the approach to deploy computing power for large-scale operation in space. At that time, Google announced its partnership with satellite firm Planet, planning to launch two prototype satellites in early 2027 to verify this concept, as well as test the feasibility of laser inter-satellite communication between the two units.

It is widely acknowledged that the ultimate bottleneck of computing power is energy. Deploying computing assets in space can not only access far stronger solar energy for extended periods, but also eliminate the need to build ground data centers to house servers.

However, all these are only superficial advantages. Behind Suncatcher lies the intersection between the ultimate romantic fantasy Google has held since its founding, and the current commercial reality built on top of artificial intelligence.

What kind of project is this?

To be precise, the satellite Google is sending to space this time is codenamed MVP (Minimum Viable Product). It will hitch a ride on SpaceX's Transporter-18 rideshare mission, launched by a Falcon 9 rocket from Vandenberg Space Force Base in California, entering a low Earth orbit that is almost permanently illuminated by sunlight.

The satellite is roughly the size of a household refrigerator, and its satellite bus is provided by San Francisco-based satellite imaging firm Planet Labs.

The satellite is equipped with 4 Google self-developed TPUs, and the sum of their total computing power is approximately equal to one server in a ground data center.

It is powered by solar panels with an output of around 1 kilowatt, which is roughly enough to run a microwave oven or a hair dryer. According to Google's plan, Suncatcher will operate for about one year, while the satellite itself can stay in orbit for up to 6 years before re-entering the atmosphere and burning up completely.

Pichai stated that the primary goal of this mission is to test the hardware performance and heat dissipation capability, verifying whether the TPUs can survive the three major challenges: launch impact, space radiation, and thermal management.

Orbit selection is also a critical factor.

MVP will enter a special orbit called "dawn-dusk SSO" (sun-synchronous orbit). It allows the satellite to fly permanently along the Earth's terminator, so it can almost always be exposed to sunlight.

On the ground, solar panels are restricted by light intensity, day-night cycles, cloudy and rainy weather. But in outer space, the situation is completely different: solar radiation is far stronger, the power output of the same panel can reach up to 8 times of its maximum output on the ground, and it can provide almost continuous power supply.

Google plans to arrange 81 satellites into an array with a radius of about 1 kilometer, at an average orbital altitude of around 650 kilometers, and connect them into a unified computing cluster using laser inter-satellite links (free-space optical communication).

Although there is vast space in orbit, these satellites still need to fly very close to each other. This is because to make space-based computing power catch up with the performance of ground data centers, the communication between satellites must be extremely fast and highly stable.

Currently, the data rate of most inter-satellite laser links ranges from 1 to 100 gigabits per second, and they are designed for "long-distance, low-bandwidth" scenarios.

But it is worth noting that Google's core goal is to train AI models, so this satellite array requires "extremely short distance, extremely high bandwidth" connections.

According to Google's estimation, to support distributed AI training, a single link needs to achieve an aggregated bandwidth of 10 terabits per second. When the distance between satellites is shortened, the required received optical power will drop significantly.

The required pointing accuracy is roughly equivalent to making two high-speed moving satellites align with a coin-sized target on each other from several miles away.

As mentioned earlier, Google engineers need to overcome three major challenges to send AI chips into space.

The first challenge is the launch process itself. The rocket will fly upward for about 10 minutes, during which the satellite will sustain continuous acceleration, some local components may instantly bear 50 to 100 times of gravitational acceleration, and the whole satellite has to survive severe vibration.

Before launch, the team tested the satellite on all three axes to simulate the vibration frequency of the rocket.

The second challenge is space radiation.

Cosmic rays and solar activities in space may cause "bit flip" in electronic components. This refers to the phenomenon that a single high-energy particle hits the chip, flipping a bit from 0 to 1, which will further cause calculation errors.

To test radiation resistance, Google placed the TPU in the accelerator at the University of California, Davis, bombarded it with 67 MeV proton beams while running AI workloads on the chip.

Preliminary test results show that the chip can withstand a radiation dose equivalent to more than 5 years of mission operation.

However, some uncorrectable errors still occurred in the video memory. Google stated that this error rate "is roughly acceptable for inference scenarios", while its impact on model training still requires further research.

The third challenge is heat dissipation.

Space is a vacuum environment with no air convection, so cooling fans are completely useless, while AI chips generate extremely high heat during operation.

Therefore, Google's solution is to use a layer of deformable "thermal interface material" to connect the chip to heat pipes made of aluminum and copper, then transfer the heat to a heat sink, which radiates the excess heat directly into space.

The problem is that the heat dissipation capacity of the heat sink is limited. So in this test, the TPU can only run intermittently for about 15 minutes each time, then it has to shut down and wait for the heat sink to dissipate all accumulated heat.

Google will run the Gemini model on this satellite for testing, but it cannot support continuous operation of the model.

The project is led by Google Senior Director Travis Beals, whose department is named "Paradigms of Intelligence".

Pichai emphasized that this is a stepping stone towards the vision of "orbital data centers", rather than a formal product launch.

In 2027, Google also plans to launch two more prototype satellites specifically to verify the performance of laser inter-satellite links.

As for when Suncatcher can evolve from a "project" to a formal "product", Pichai said it will still take many years.

What is the core motivation behind all this?

The extremely high power consumption of training and running large AI models is no longer a secret to the public.

In 2026, all hyperscale data center operators around the world are scrambling for power resources across the globe, from the data center corridor in Virginia, the United States, all the way to Ireland. Almost all available power resources on the ground have been nearly fully occupied.

Google's logic is that instead of competing for power, land and water resources on the ground, it is better to explore new opportunities in outer space.

Apart from abundant solar energy, there are no residents, no farmland, no community opposition in space, which also eliminates the whole set of troubles including building data centers, connecting to the power grid, and consuming massive water for cooling.

Google stated in its published paper that the total power output of the Sun is more than 100 trillion times of the total global human power generation.

Another major feature of Suncatcher is its modular design.

In many sci-fi works, space research centers are usually "megastructures" with extremely large volume, filled with a huge number of computers inside.

But Google did not choose this approach. The paper specifically discussed this "monolithic" scheme, and concluded that such structure requires on-site assembly by humans or robots in space, which brings more difficulties in collision avoidance, and the structural strength requirement will also push up the total weight and system complexity significantly.

Google's choice is to use a fleet of much smaller satellites, flying in a very close formation.

The advantage of this approach is excellent scalability. When more computing power is needed, Google can launch more satellites. Theoretically, the computing power can be expanded infinitely like building blocks, until the entire dawn-dusk orbit belt is fully filled.

The fact that the scheme makes theoretical sense does not mean it is economically viable.

Google calculated that currently, the cost to send 1 kilogram of payload into low Earth orbit is roughly between 1500 and 2900 US dollars. But this price is only a hypothetical estimation, the actual cost will only be higher depending on the specific launch demand of each mission.

Google stated that to make Suncatcher economically feasible, the launch cost needs to drop to around 200 US dollars per kilogram.

Why 200 US dollars per kilogram?

Because the electricity cost for data centers in the United States is roughly 570 to 3000 US dollars per kilowatt per year. If the launch cost can be reduced to 200 US dollars per kilogram, then when the launch cost is amortized over the entire service life of the satellite, the total cost per kilowatt may be roughly equivalent to the energy cost of ground data centers.

Google said that achieving the 200 US dollars per kilogram target requires two prerequisites: SpaceX's Starship is successfully put into commercial operation, and it can achieve 180 launches per year afterwards.

Therefore, Google estimates that this goal will not be achieved until at least 2035.

By that time, the launch and operation cost of space-based data centers, calculated per kilowatt per year, may be comparable to the energy cost of ground data centers of the same scale.

Of course, by then, the electricity cost of ground data centers will also change, which may be higher or lower than the current level.

Although 2035 is still far away from us, the space computing race has already begun.

Elon Musk has long stated that SpaceX "will build" space data centers. After Pichai posted the announcement of the MVP satellite launch, Musk replied to him with two rocket emojis.

Shortly afterwards, Musk added: "The total amount of computing power in space will obviously converge to 100% of all computing power eventually."

Amazon founder Jeff Bezos also made similar remarks, saying that in 10 years, gigawatt-level data centers will appear in space. Former Google CEO Eric Schmidt acquired the rocket company Relativity Space, and he also plans to send data centers into orbit.

Startup company Starcloud has already sent a satellite equipped with NVIDIA H100 GPU into orbit, and stated that it will build a 5-gigawatt space data center with a 4-kilometer wide solar array one day in the future.

Google's MoonShot

If you only understand Suncatcher from a commercial perspective, there is not much to discuss. After all, spending huge sums of money to send chips into space, is it just for saving some electricity bills?

Because commercial value is only the superficial layer of this project. The fundamental reason why this project can be established is the "moonshot" culture that has been deeply rooted in Google since its founding.

The story dates back to 2005.

In that year, Stanford professor Sebastian Thrun led a group of students to build an autonomous vehicle named Stanley, and participated in the Grand Challenge held by the Defense Advanced Research Projects Agency (DARPA) of the United States.

The competition required the autonomous vehicle to traverse a 132-mile off-road track located in California.

Stanley completed the whole course, won the championship, and became famous overnight.

Among the audience of that competition were Google's founders Larry Page and Sergey Brin, who reportedly even disguised themselves to attend the event.

They did not just come to watch the race, but also planned to recruit talents from the participating teams.

In 2007, Page invited Thrun to join Google, and offered him almost unlimited resources to support his work.

Thrun did not disappoint them. The 360-degree in-vehicle camera system he installed on the autonomous vehicle later evolved into Google Street View. In January 2009, he launched the autonomous driving project Project Chauffeur, which eventually grew into today's self-driving business Waymo.

Then Page gave him an unprecedented position — Google's first "director of other", who was specifically responsible for managing those projects that "investors cannot understand, but are extremely cool".

Thrun alone was not enough. Page also took a fancy to another person, Astro Teller.

Teller is a PhD in artificial intelligence from Carnegie Mellon University. Before joining Google, he had serial entrepreneurial experiences, founded wearable device company BodyMedia, and also established Cerebellum Capital, an investment fund that uses AI for investment decision-making.

Page once stated that what he valued most was Teller's track record of "turning crazy ideas into viable businesses", so he recruited him to serve as the "director of new products".

On September 12, 1962, then US President John F. Kennedy delivered a speech at Rice University, saying "We choose to go to the moon in this decade". In 1969, Apollo 11 successfully landed on the moon.

Since then, the word "moonshot" in English has represented those grand goals that are publicly proposed, with a clear deadline, technically seemingly impossible, but achieved through concentrated resource mobilization.

Both Page and Brin attended Montessori schools. Page once said: "We both went to Montessori schools, and I think it was that training of 'not obeying rules and orders, being self-driven, constantly asking questions about the world, and doing things differently from others' that made us a little different."

In 2010, a semi-secret laboratory emerged inside Google, which was Google X, led by Teller and Thrun, specifically built to carry this "moonshot" culture.

Teller said that the Apollo program is the spiritual prototype of X, which turns the impossible into reality.