In recent years, the artificial intelligence (AI) industry has witnessed a rather unconventional transformation.

On one hand, ground-based AI data centers are expanding at an unprecedented rate, with NVIDIA GPUs in high demand among global tech firms, and nations continuously investing in computing infrastructure. On the other hand, some enterprises and research institutions are setting their sights on low Earth orbit, pondering a question that once belonged solely to the realm of science fiction: Will a fraction of future data centers be constructed in space?

This concept is gradually stepping into the industrial limelight. Yet, the greater the enthusiasm, the more intense the debate.

In the United States, Starcloud is making efforts to launch high-performance computing equipment into orbit for environmental validation, while Google has unveiled Project Suncatcher to explore the feasibility of deploying AI computing capabilities in space. China, too, is advancing its explorations in areas such as computing satellites and spatial intelligent computing, from the "Xingsuan" initiative by GX Space to the space-based intelligent computing verification conducted by the Zhejiang Lab's "Three-Body Computing Constellation," all seeking to answer the same question: Can satellites in the future not only collect and transmit data but also engage in computing?

Proponents argue that AI's appetite for computing resources is insatiable, and space offers unique conditions that are unattainable on Earth. Skeptics, however, contend that this is merely a concept hyped up by capital. They point out that ground-based data centers already boast a mature industrial chain, and sending servers into orbit would entail overcoming a series of challenges, including energy supply, heat dissipation, maintenance, and cost.

So, what exactly is space computing power? Is it the next frontier in computing infrastructure, or is it still a distant technological dream?

01 The Next Battleground for AI Computing Power

Over the past few years, large models have continuously pushed the envelope of capabilities, propelling AI into a stage that is even more reliant on computing power.

AI's demand for computing resources is escalating. Training large models necessitates tens of thousands of GPUs operating continuously; after deployment, these models must also handle massive real-time requests from users worldwide.

As model scales continue to grow, the competition in AI has begun to extend to the infrastructure that supports these models. The computing power a single GPU can deliver hinges on the entire infrastructure behind it: data centers, power supply, cooling systems, network facilities, and land resources.

AI is reshaping data centers.

In the era of the traditional internet, data centers primarily managed web access, application services, and data storage, with relatively stable computing loads.

However, data centers in the AI era are vastly different. A multitude of GPUs must operate under high loads for extended periods, server density is on the rise, and the amount of heat generated per unit area is increasing, placing higher demands on energy and cooling.

This is why global tech giants are vying for new energy and data center resources. Microsoft, Google, Amazon, and others are continuously ramping up their investments in AI infrastructure, and in some parts of the United States, data center construction is even being constrained by grid capacity.

As more and more GPUs are brought online, new questions emerge: How can sufficiently stable and cost-effective infrastructure be provided to keep these computing capabilities continuously effective?

02 Why Look to Space?

If the issue were merely a lack of energy, increasing power generation capacity might seem like a straightforward solution. However, AI data centers require not just electricity but also suitable land, stable networks, and long-term operational conditions.

At the same time, satellites are emerging as new data sources. Traditionally, satellites collected data, while the ground handled processing and analysis—a model that worked well when the number of satellites was limited. However, with the proliferation of remote sensing and communication satellites, the volume of satellite data is skyrocketing, and transmitting all raw data back to the ground is facing new pressures.

Take remote sensing satellites, for instance: a single satellite can capture a vast number of surface images daily, but the truly valuable information is often just a fraction of them. For example, whether a forest fire has broken out, whether abnormal vessels have appeared in maritime areas, or whether environmental changes have occurred in a certain region.

Transmitting all raw data back to the ground through limited space-ground links would not only consume significant communication resources but also slow down response times.

Therefore, endowing satellites with certain computing capabilities to perform data screening, target recognition, and preliminary analysis in orbit is becoming the most practical application direction for spatial computing.

03 How Does Computing Power Reach Space?

Launching servers into space is not inherently difficult in terms of launch technology. Rockets can deliver equipment into orbit, but the real challenge for space computing power is how to keep these devices operating there long-term.

Ground-based data centers have a well-established system: stable power grids, large-scale cooling, equipment maintenance, and continuous upgrades. However, in the space environment, computing devices must independently address issues such as energy supply, heat dissipation, space radiation, and long-term reliability.

Therefore, space computing power will not leap directly to "orbital data centers" but will evolve along a gradual development path.

In the current exploration paths, the industry is more inclined to first enable satellites to possess computing capabilities and then gradually expand the computing scale. From computing modules on a single satellite to a spatial computing network composed of multiple satellites, and then to the potential orbital computing infrastructure that may emerge in the future, this process can be roughly divided into four stages.

The earliest stage involves computing payloads. Simply put, this means adding computing capabilities to traditional satellites. Satellites still undertake primary tasks such as remote sensing and communication, but with the addition of AI chips and computing modules, they can complete some data processing in orbit and transmit only more valuable information. The Aurora series of space computers by CAS Space is an exploration in this direction. At this stage, satellites begin to possess "their own brains."

The next stage of development is computing satellites. At this point, computing is no longer just an additional function but becomes a core objective of satellite design. Satellites require stronger energy sources, higher-performance computing chips, and more sophisticated thermal management capabilities—essentially creating computing nodes in orbit. The "Xingsuan" plan by GX Space explores this direction.

Taking a step further is the computing constellation. The computing capabilities of a single satellite are always limited, but if multiple computing satellites are interconnected through inter-satellite communication, a spatial computing network similar to a server cluster can be formed. Different satellites can perform task scheduling and resource allocation. The "Three-Body Computing Constellation" by Zhejiang Lab, the "Xingsuan" plan by GX Space, and the explorations around spatial intelligent computing by PIESAT all reflect this trend.

However, these projects are currently still in the technical verification and industrial exploration stages, and there is a long way to go before achieving true "space cloud computing."

A more long-term direction is orbital data centers. This represents the most imaginative form of space computing power: deploying large-scale computing infrastructure in orbit to provide computing capabilities akin to ground-based data centers for AI tasks. Explorations such as Google's Project Suncatcher and the Beijing Space Data Center fall into this direction.

However, this remains a long-term goal. Because orbital data centers face not just technical challenges but also commercial issues. Whether launch costs are low enough, whether energy supply is sufficiently stable, whether heat dissipation problems can be solved, and whether equipment can operate long-term will all determine whether it can ultimately succeed.

These four stages are not mutually exclusive. In the future, computing payloads will still serve remote sensing and communication satellites, computing satellites will become spatial computing nodes, computing constellations will further form networks, and orbital data centers will represent the possible direction of larger-scale spatial infrastructure.

04 Is Space Computing Power Reliable?

The most alluring aspect of space computing power is also the aspect most prone to skepticism.

The orbital environment does offer some conditions that are unattainable on Earth. Satellites can tap into continuous solar energy, low Earth orbit networks are rapidly expanding, and more and more data is being generated by spatial devices such as satellites. If computing capabilities can be brought closer to this data, many tasks would not need to rely on large-scale data transmission back to the ground.

However, from a practical standpoint, sending computing equipment into space is not as straightforward as imagined.

Ground-based data centers already have a mature system, with established solutions for server upgrades, equipment maintenance, energy, and cooling. Computing devices in orbit, however, must independently address energy, heat dissipation, reliability, and cost issues.

Therefore, space computing power is unlikely to replace ground-based data centers in the short term.

A more realistic path is to first address the issues inherent to space itself: enabling satellites to possess computing capabilities, forming computing networks among satellites, and then exploring larger-scale orbital computing infrastructure.

Over the past few decades, satellites have helped humanity see the world and connect the world. Space computing power is now exploring a third possibility: enabling satellites to begin understanding the world they see, moving from "seeing" and "transmitting" to "computing."