Google CEO Pichai officially announces Google will launch Tensor Processing Units (TPUs) to orbit on October 1, expanding AI compute infrastructure to space-based platforms.

Google is about to reach for the stars.

In a lengthy post, CEO Sundar Pichai announced that Google will begin the first orbital test of its Suncatcher project on October 1st—with computing power equivalent to a single ground-based server.

Google first publicly revealed the Suncatcher project back in November of last year. It is defined as Google's "long-term research," aimed at exploring how to scale computing workloads in space. At that time, Google announced a partnership with satellite company Planet and outlined plans to launch two prototype satellites in early 2027 to validate the concept and test laser-based inter-satellite communication.

The conventional wisdom holds that computing's ultimate constraint is energy. In space, satellites can harness sustained solar power far more efficiently than ground systems while eliminating the need for physical data centers, buildings, cooling systems, and power infrastructure.

Yet beneath the surface, Suncatcher represents something deeper: the convergence of Google's founding-era idealistic vision with the commercial realities of today's AI economy.

**What is this project?**

To be precise, Google is launching a satellite codenamed MVP (Minimum Viable Product), hitching a ride on SpaceX's Transporter-18 mission aboard a Falcon 9 rocket from Vandenberg Space Force Base in California. It will enter a low Earth orbit in a dawn-dusk sun-synchronous path that keeps it almost perpetually bathed in sunlight.

The satellite is roughly refrigerator-sized. The satellite platform comes from Planet Labs, a San Francisco-based satellite imaging company. Inside are four TPUs (Tensor Processing Units) custom-built by Google, which together deliver computing power roughly equivalent to a single server in a terrestrial data center.

Power comes from solar panels producing approximately 1 kilowatt of output—roughly the consumption of a microwave or hair dryer. Google plans to operate Suncatcher for about one year, though the satellite itself can remain in orbit for up to six years before burning up during atmospheric re-entry.

According to Pichai, the mission's primary goal is testing hardware performance and thermal dissipation—determining whether TPUs can withstand the stresses of launch, radiation, and heat.

The orbital mechanics are critical. The MVP enters a dawn-dusk sun-synchronous orbit (dawn-dusk SSO), which keeps the satellite perpetually riding Earth's twilight line, receiving nearly continuous sunlight. On the ground, solar panels face constraints from cloud cover, day-night cycles, and weather. In space, however, solar radiation is far more intense—the same panel can generate up to eight times the power it would produce on Earth, with nearly uninterrupted supply.

Google plans to eventually deploy 81 satellites in a formation spanning roughly one kilometer in radius at an average altitude of about 650 kilometers, networked together via laser inter-satellite links—free-space optical communication—to function as a unified computing cluster.

Despite the vastness of orbital space, these satellites must fly in tight formation. This is because for space-based computing to match terrestrial data centers, inter-satellite communication must be both fast and reliable.

Current laser links between satellites typically operate at rates between 1 and 100 gigabits per second, and they are engineered for long-distance, low-bandwidth applications. But Google's core mission is AI training, which demands the opposite: extremely short-range, extremely high-bandwidth links.

According to Google's projections, distributed training requires aggregate bandwidth of 10 terabits per second across individual links. As distance decreases, the required receiving optical power drops dramatically—the precision equivalent to aligning two high-speed satellites to target an object the size of a coin from several miles away.

**Three challenges**

As noted, launching AI chips to space requires Google's engineers to clear three hurdles.

The first is launch itself. A rocket's ascent takes roughly ten minutes, during which satellites experience sustained acceleration. Localized components may face instantaneous g-forces of 50 to 100 times Earth's gravity, plus severe vibration. Before launch, the team subjected the satellite to three-axis vibration testing to simulate rocket-flight frequencies.

The second challenge is radiation. Cosmic rays and solar activity in space can cause "bit flips" in electronic components—individual particles strike and flip a bit from 0 to 1, causing computational errors.

To test radiation hardness, Google exposed its TPU to 67 MeV proton beams in an accelerator at UC Davis while running AI workloads. Initial results showed the chip withstood radiation doses equivalent to over five years of operation. However, some uncorrectable memory errors were observed. Google states this error rate is "probably acceptable for inference," though the implications for training require further study.

The third challenge is heat dissipation. Space is a vacuum—there is no air convection, so cooling fans are useless—yet AI chips generate intense heat.

Google's solution employs a deformable "thermal interface material" between the chip and heat pipes made of aluminum and copper, routing thermal energy to a radiator panel that dissipates heat into space. The problem is the radiator's capacity is limited. During this test, TPUs can only run intermittently—approximately 15 minutes per session—then must shut down while the radiator sheds accumulated heat. Google will run its Gemini model on this satellite for testing purposes, but cannot sustain continuous operation.

The project is led by Travis Beals, a senior director in Google's "Paradigms of Intelligence" division.

Pichai emphasized this is a stepping stone toward "orbital data centers," not a product launch. In 2027, Google plans to launch two additional prototype satellites specifically to validate laser inter-satellite links. As for when Suncatcher transitions from "project" to "product," Pichai suggested that remains years away.

**The motivation**

It is no secret that training and running large language models consume enormous amounts of electricity. Throughout 2026, every hyperscale data center operator globally has been scrambling for power—from Virginia's data center corridor to Ireland. Ground-based electricity sources are running out.

Google's thinking is straightforward: rather than compete for power, land, and water on Earth, why not seek opportunities in space?

Beyond solar energy itself, space harbors no residents, no farmland, no community opposition—and eliminates the entire apparatus of constructing data centers, connecting to power grids, and managing water cooling.

Google notes in its research paper that the sun's output power exceeds humanity's total electrical generation by a factor of more than one million billion.

Another defining feature of Suncatcher is its modularity. In many science fiction works, space research installations are imagined as massive megastructures filled with computing equipment, assembled by humans or robots on-site—an approach Google explicitly rejected. Their paper discusses this "monolithic" approach and concludes it requires on-orbit assembly by humans or robots, complicates collision avoidance, and escalates structural requirements, weight, and complexity.

Instead, Google chose a fleet of smaller satellites maintaining close formation flight.

The advantage is scalability: more computing power simply requires launching additional satellites, theoretically allowing infinite expansion—like stacking blocks—until filling an entire dawn-dusk orbit corridor.

Theory is one thing; economics are another.

Google calculated that launching one kilogram to low Earth orbit currently costs between $1,500 and $2,900. Yet this is a nominal figure; actual costs depend on specific mission requirements and typically run higher. For Suncatcher to become viable, launch costs must drop to roughly $200 per kilogram.

Why $200?

Because U.S. data center electricity costs approximately $570 to $3,000 per kilowatt annually. If launch costs fall to $200 per kilogram, amortizing those costs over a satellite's lifetime and dividing by kilowatt output produces energy costs comparable to terrestrial data centers.

Google states achieving $200 per kilogram requires two conditions: SpaceX's Starship becoming operational and achieving 180 launches annually.

Therefore, Google estimates reaching cost parity by 2035 at the earliest.

By then, the combined launch and operational costs of space-based data centers, measured per kilowatt annually, could match energy costs of equivalent ground-based facilities. Of course, by 2035, terrestrial data center electricity costs will have shifted as well—potentially higher or lower.

**The space race begins**

Though 2035 remains distant, the space race is already underway.

Elon Musk has long stated that SpaceX will build space data centers. After Pichai's announcement of the MVP launch, Musk responded with two rocket emojis. He later commented: "The total compute in space obviously converges to 100% of all compute."

Amazon founder Jeff Bezos made similar remarks, predicting gigawatt-scale data centers in space within a decade. Former Google CEO Eric Schmidt acquired rocket company Relativity Space, also planning to deploy data centers in orbit. Startup Starcloud has already placed a satellite equipped with NVIDIA H100 GPUs in orbit and claims plans for a 5-gigawatt space data center spanning a 4-kilometer solar array.

**Google's moonshot**

If Suncatcher were understood purely as a business venture—spending billions to launch chips into orbit simply to save on electricity—there would be little to discuss.

But commerce is merely the surface. The deeper root enabling this project traces back to a cultural principle embedded at Google's founding: the "moonshot" mentality.

The story begins in 2005. That year, Stanford professor Sebastian Thrun led a team of students in building an autonomous vehicle named Stanley to compete in the DARPA Grand Challenge—a race requiring an unmanned car to traverse a 132-mile off-road course through California.

Stanley completed the full distance and won, becoming instantly famous. Among the spectators, Google founders Larry Page and Sergey Brin attended—reportedly in disguise.

They were there not only to watch but to recruit. In 2007, Page brought Thrun to Google, along with virtually unlimited resources.

Thrun delivered. The 360-degree camera system he mounted on autonomous vehicles became Google Street View. In January 2009, he launched Project Chauffeur, the autonomous driving initiative that eventually evolved into today's Waymo. Page created an unprecedented role for him: Google's first "Director of Other"—charged with managing "things investors don't understand, but are cool."

But Thrun alone wasn't enough. Page also recruited Astro Teller, an AI Ph.D. from Carnegie Mellon University with a track record of entrepreneurship. Before joining Google, Teller founded BodyMedia, a wearable company, and Cerebellum Capital, an AI-driven investment fund. Page was drawn to Teller's proven ability to "turn crazy ideas into businesses" and brought him aboard as "Director of New Products."

In September 1962, President John F. Kennedy stood at Rice University and declared: "We choose to go to the moon in this decade." By 1969, Apollo 11 had reached the Moon. From that moment forward, "moonshot" in English came to mean "a bold, time-bound, technically audacious goal with massive resource commitment that actually gets done."

Both Page and Brin attended Montessori schools. Page once said: "We both went to Montessori schools. I think that culture of questioning rules and authority, self-motivation, constant questioning, and doing things differently from others—I think that made us a bit different."

In 2010, Google created a semi-secret internal laboratory called Google X, led by Teller and Thrun, to institutionalize this "moonshot" culture. Teller explained that the Apollo program served as X's spiritual prototype—turning the impossible into achievable reality.

He articulated a counterintuitive principle: convincing people to pursue something 10 times better is easier than getting them to improve by 10 percent. Incremental improvement keeps you competing on the same field as everyone else; a 10x goal forces you off the beaten path entirely.

Google X eventually rebranded as "X, The Moonshot Factory," earning Teller the title "Captain of Moonshots."

X differs fundamentally from ordinary corporate research labs. Typical labs exist to improve a company's core business; X's mission is solving difficult problems outside Google's core operations. It bears no revenue targets and operates under no product deadlines.

With this context, Suncatcher suddenly makes complete sense. Google's official blog explicitly positions it alongside two predecessors: quantum computing, which launched over a decade ago, and autonomous driving, begun fifteen years prior. Suncatcher is Google's moonshot tradition's newest member. Like its predecessors, it appeared audacious at birth. Yet this "try the impossible first, worry about returns later" culture is precisely why Google dares send servers to space.