Although the concept of "space data center" has been highly sought after in recent years, the number of GPUs actually deployed in space orbit is still very few. Now this situation is beginning to change, and the near-future business model of orbital computing power is gradually becoming clear.
In January this year, Canadian company Kepler Communications deployed the world's largest orbital computing cluster in orbit: about 40 Nvidia Orin edge processors are mounted on 10 satellites operating in orbit, and are interconnected through inter-satellite laser links. Kepler already has 18 customers, and on Monday it announced its newest customer, startup Sophia Space, will be testing its unique orbital computer software on the Kepler constellation.
Industry experts predict that it may take until the 2030s for the large-scale space data centers conceived by companies such as SpaceX and Blue Origin to be implemented. Until then, the first phase of commercial opportunities lies in in-situ processing of data collected in orbit to enhance sensor capabilities for a variety of commercial and government space payloads.
Kepler CEO Meena Mitri said that the company does not position itself as a "data center company" but an infrastructure provider for space applications. It hopes to become the network service layer for other satellites in orbit, as well as lower-altitude drones and aircraft, providing them with connection and computing power support.
Different from Kepler's network-oriented positioning, Sophia is developing passively cooled space computers, trying to solve one of the key challenges facing large-scale data centers in orbit: how to dissipate heat for high-performance processors without relying on bulky and expensive active cooling systems. In the new cooperation between the two parties, Sophia will upload its self-developed operating system to one of Kepler's satellites and try to complete the startup and configuration of the system on two spacecrafts and a total of six GPUs. This type of operation is already commonplace in ground-based data centers, but it is a first in an orbital environment and is seen as a key "de-risking" test before Sophia launches its first own satellite in late 2027.
For Kepler, this partnership is an important step in proving the value of its network. Currently, its constellation mainly carries and processes uplink data from the ground, or from "hosted payloads" carried on its own spacecraft. As the industry matures, Kepler is expected to gradually dock with more third-party satellites to provide network and on-orbit processing services. Mitri said that more and more satellite companies are already planning future asset layouts based on this model, especially for high-power sensors such as synthetic aperture radar, where outsourcing data processing to external computing nodes has obvious advantages. The U.S. military is an important customer for such needs. Its new generation missile defense system highly relies on satellites to detect and track threat targets. Kepler has previously completed verification of a satellite-to-ground laser link as part of a U.S. government demonstration project.
This type of "edge computing" - processing data locally at the location of data collection to improve response speed - is seen as the first scenario where orbital data centers will demonstrate value. It is this path choice that distinguishes the development direction of Kepler and Sophia from large aerospace companies such as SpaceX and Blue Origin, as well as startups such as Starcloud and Aetherflux that focus on building earth-like data center architecture and using data center-level processors.
Mitri emphasized that from a business logic point of view, the team believes that space applications will be mainly based on reasoning rather than training, so it is more inclined to deploy a large number of distributed GPUs for reasoning tasks rather than a small number of "super GPUs" with super training capabilities. “If a processor consumes several kilowatts of power but only works 10% of the time, that doesn’t really mean much,” he pointed out. “In our model, the GPU is basically running at 100% load.”
Once this type of technology is verified in an orbital environment, the imagination of space computing power will be further opened up. Sophia CEO Rob DeMillo pointed out that the state of Wisconsin in the United States just passed a bill last week prohibiting the construction of new data centers there, and some lawmakers at the federal level in the United States have proposed similar restrictions. In his view, any policy that limits the expansion of ground data centers will objectively increase the attractiveness of space data centers.
“There will probably never be another data center in this country,” DeMillo said, sarcastically. “It’s just going to get weirder.”