Google развивает космическую инфраструктуру для ИИ на основе солнечной энергии Translation: Google is developing a space infrastructure for AI based on solar energy.

Google is set to develop a satellite system in low Earth orbit to harness solar energy for powering data centers. This was announced in the company’s blog.

The corporation believes that artificial intelligence is a pivotal technology with the potential to transform the world. To unlock its full capabilities, substantial computational power is required.

“The Sun is the primary energy source in our solar system, emitting energy at a magnitude of 100 trillion times that of total human production. On an optimal orbit, solar panels can be eight times more efficient than those on Earth, generating energy almost continuously and reducing the need for batteries,” the company states.

Looking ahead, space could be the ideal environment for scaling artificial intelligence, which has led Google to launch an “ambitious research project” called Suncatcher. This initiative aims to create compact groups of solar-powered satellites equipped with the company’s Tensor Processing Units (TPUs) interconnected through optical communication channels.

The proposed system will operate in a continuously illuminated area of low Earth orbit, synchronized with the Sun. In this zone, the devices will receive near-constant sunlight, maximizing energy collection.

Google emphasizes that several challenges must be overcome to realize the Suncatcher project.

Establishing inter-satellite communication links

Large-scale machine learning workloads necessitate the distribution of tasks among numerous accelerators with high bandwidth and low latency.

To achieve performance levels comparable to terrestrial data centers, communication channels between satellites must support dozens of terabits per second. Experts at Google believe this is feasible through the use of multichannel dense transceivers, along with wavelength division multiplexing and spatial multiplexing.

To achieve the required bandwidth, the levels of received power must be thousands of times greater than what is typical for conventional long-range systems.

“As input energy is inversely proportional to the square of the distance, we can mitigate this issue by placing satellites in very dense formations,” the blog states.

The team has already begun testing this approach, with a laboratory sample successfully achieving data transmission rates of 800 Gbps in both directions.

Satellite management

High-bandwidth inter-satellite communication channels require very close proximity of devices.

Experts have developed models to analyze the orbital dynamics of such clusters. Their findings suggest that placing satellites just a few hundred meters apart would require minimal maneuvers to maintain stability.

TPU radiation resistance

To ensure that the accelerators ML remain resilient, they must be capable of withstanding conditions in low Earth orbit. The company has tested the Trillium chip, yielding promising results.

Memory subsystems with high bandwidth began to show instability after accumulating a radiation dose nearly three times higher than anticipated.

Experts did not observe any significant failures.

Economic feasibility

Previously, high launch costs were a major barrier to establishing large-scale space systems. Data analysis has indicated that by the mid-2030s, launch costs could decline to below $200 per kilogram.

With such figures, a space data center could become economically viable, comparable to a terrestrial data center.

Google stressed that the primary computation does not contradict fundamental laws of physics or face insurmountable economic barriers.

“However, significant engineering challenges remain, such as thermal regulation, high-bandwidth terrestrial connectivity, and the reliability of the orbital system,” the blog notes.

To address these challenges, Google will initiate problem-solving through a training mission in partnership with Planet, planning to launch two prototype satellites by early 2027. This experiment aims to evaluate the functionality of the models and equipment in space.

It’s worth noting that in May 2025, China sent into space 12 satellites as part of a project to establish a network of orbital supercomputers.