Google AI Achieves First Response in Space as Data Center Competition Intensifies

After a month since launch, Gemma 4 is operational in orbit... One TPU can only run for 20 minutes at a time.

Big tech turns to space due to power shortages... Cost, cooling, and radiation remain challenges.

Laser connections between two satellites planned for 2027... SpaceX and Starcloud also in competition.

[Block Media Reporter Lee Hye-yeon] Google has successfully received its first response from an artificial intelligence (AI) system operating on a satellite launched into space. The AI model performed calculations directly in space and sent the results back to Earth.

As the power demand for AI data centers surges, the infrastructure competition among big tech companies is extending into space. However, operating AI in space and generating commercial profits are two separate issues. High construction costs and securing cooling and communication technologies remain key challenges.

According to Cryptopolitan, Google received its first AI response from a prototype satellite of Project Suncatcher on the 30th (local time). The satellite, equipped with Google DeepMind's open model Gemma 4, sent a haiku themed around space.

The satellite was launched on the 1st aboard SpaceX's Falcon 9 rocket. Google collaborated with Earth observation satellite company Planet on this project.

After about three weeks of system checks, Google's research team conducted the first AI computation. They sent commands while the satellite was passing over the Svalbard ground station in Norway and received a response. The three-line poem sent back by the satellite is as follows:

"Blue marble below
Data streams in silent grace
Starlight feeds my core."

Google also tested a limerick themed around space and telemetry commands to check the satellite's operational status. The significance of this experiment lies in confirming that AI computations can actually operate in space and transmit results back to Earth.

However, it is still in the early stages; the satellite carries four Tensor Processing Units (TPUs) dedicated to Google's AI, but only one can operate for about 20 minutes at a time. The power generated by the satellite is limited to a maximum of 1 kilowatt (kW).

AI Power Crisis Fuels Space Data Center Competition

The background of this experiment is the surging power demand of AI data centers.

According to a study published in the international journal Nature, the power consumption of AI data centers from six major big tech companies is estimated to reach about 118 terawatt-hours (TWh) by 2024. By 2030, it is projected to increase to between 239 and 295 TWh, accounting for about 1% of global power demand. Goldman Sachs also predicts that the power demand for data centers worldwide will increase by 170% from 2025 to 2030.

Securing land and power grids for building data centers is becoming increasingly difficult. Global real estate services firm JLL analyzed that connecting data centers to power grids can take 2 to 10 years depending on the region. As of 2025, Amsterdam and Tokyo are expected to take an average of 10 years, the longest among major markets.

In space, there is no need to connect to terrestrial power grids, as solar power can be generated. Particularly in sun-synchronous orbits, which circle along the line where the sun never sets, sunlight can be received almost continuously. Google explains that solar panels in this orbit can generate up to eight times more power than on the ground. This could reduce the constraints of power supply and land acquisition faced by terrestrial data centers.

However, not all AI computations can be moved to space. JLL identifies AI training and batch processing, as well as processing data collected in space, as suitable tasks for space data centers. In contrast, AI services that require real-time transactions or immediate interaction with users are better suited for terrestrial data centers. Servers need to be located close to users to ensure response speed.

There are differing views on AI training. The Boston Consulting Group (BCG) analyzes that large-scale foundational AI model training will likely remain on the ground due to the need for tightly connected, high-density computing equipment. It is expected that space data centers will evolve to take on specific tasks rather than replace terrestrial infrastructure.

Positive Market Outlook but Economic Viability Remains a Challenge

Regarding the market potential of space data centers, both positive forecasts and cost burdens are discussed.

BCG predicts that in the most likely scenario, space data centers could capture 10-15% of the global AI data center market by 2040. This translates to annual revenues of $240 billion to $320 billion.

However, this outlook is based on the premise of technological advancements and cost reductions. Currently, the construction costs of space data centers are estimated to be 2.5 to 3 times higher than those on the ground.

BCG anticipates that even in a realistic scenario where launch costs decrease by 67% and satellite weight is reduced by 50% over the next 5-10 years, the cost gap will remain between 1.5 and 1.8 times. If the satellite failure rate remains at 30%, the gap will be 1.8 times, and it will need to drop to 10% to narrow to 1.5 times. Unlike on the ground, in space, if one component fails, the entire satellite may need to be decommissioned, making failure rates critical to profitability.

There are also significant technical obstacles. One of the biggest issues is cooling. In space, there is no air to circulate heat as on Earth. To dissipate heat from chips, large heat sinks must be attached. BCG estimates that a 100 kW satellite will require about 400 square meters of heat sinks. As heat sinks become larger and heavier, launch costs also increase.

Radiation is another concern. Space radiation can affect semiconductors, causing bit errors in memory or damaging computational results. The U.S. Government Accountability Office (GAO) has also pointed out that space radiation can unpredictably damage data and degrade hardware performance.

The difficulty of repairing faulty equipment, as easily as on Earth, is also a burden. Currently, satellites are typically removed from orbit after about 5-7 years of operation, and if the replacement method continues instead of repair, operational costs and space debris issues may increase.

From an environmental perspective, there are aspects to consider. Space data centers have the advantage of reducing power consumption and carbon emissions during operation by using solar power. However, a study published in 2025 in the American Computer Society (ACM) journal presented different results. Analyzing the entire process from rocket launch to space operation and satellite re-entry, it found that even under optimistic conditions, the carbon emissions from space computing could be up to 10 times higher than those of terrestrial systems. The primary cause is the carbon emissions generated during launch and re-entry. This means that simply using solar power does not automatically classify space data centers as environmentally friendly infrastructure.

Starcloud Also Competing... Satellite Connection Test Planned for 2027

Google is not the only company developing space data centers.

Starcloud launched a satellite carrying NVIDIA's H100 graphics processing unit (GPU) in November 2025. This satellite trained a small language model, NanoGPT, and also performed inference using Google's open model, Gemma. In March 2026, it secured $170 million in Series A funding, valuing the company at $1.1 billion.

SpaceX is also accelerating its AI infrastructure expansion. On the 26th of this month, SpaceX announced that it secured about $40 billion in AI chip financing through Apollo Global Management. This funding will be used to purchase NVIDIA's next-generation chip, 'Vera Rubin', for terrestrial data centers.

Each company has a different approach. Cowboy Space is pursuing a method of processing AI inference with a single satellite. In contrast, Google envisions operating multiple satellites in close proximity to share computational tasks.

The paper published by Google's research team in the international journal Joule includes a design for operating 81 satellites densely within a radius of 1 km at an altitude of about 650 km. The goal is to enable them to operate like a single data center by exchanging large volumes of data via laser communication between satellites.

If the prototype tests proceed smoothly, Google plans to conduct follow-up experiments connecting two satellites in 2027. In the next phase, they will secure more power and test optical communication connections between satellites.

The successful creation of a haiku confirms the first step in executing AI models in space, but it remains uncertain whether an AI that can write poetry can also fulfill the role of a data center. To prove commercial competitiveness, it will need to achieve not only technical success but also cost-effectiveness and operational efficiency.