From Ground to Orbit: The Physical AI Paradigm Revolution Behind the Compute Leap

If computing power is likened to the steam engine of the industrial age, then terrestrial data centers are the early factories, while space-based computing embeds the power unit directly into the global transportation network. The recent reaction in capital markets (sector bucking the trend with a 5.45% rise, net inflow of over 4 billion yuan) is not mere thematic speculation, but an advance pricing of a new computing paradigm.

The Physical Ceiling of Terrestrial Computing Power

The current main battlefield for AI training—large-scale data centers—is facing triple physical bottlenecks.

First is energy. Training a single large model with hundreds of billions of parameters consumes electricity sufficient to support a small city's daily life. Global data center energy consumption already accounts for 1%-2% of total global power generation and is still growing rapidly. Second is heat dissipation. Chip heat flux density approaches levels seen in nuclear reactors; liquid cooling and immersion cooling can only delay the arrival of limits, they cannot change the Second Law of Thermodynamics. Third is space. Data center site selection requires proximity to energy sources yet distance from populations; available physical space is drying up.

From an information theory perspective, the growth curve of terrestrial computing power has touched the upper limit of "energy-information" conversion efficiency. To continue scaling up, we must either find new energy supplies or deploy computing nodes in entirely new physical environments.

Three Core Advantages of Space-Based Computing Power

Low Earth Orbit (LEO) offers a nearly ideal alternative.

Energy Dimension: In space, there is no atmospheric obstruction. Solar radiation reception density is 6-10 times that of Earth, and it can be accessed uninterruptedly for 24 hours (as long as Earth's shadow is bypassed). This means photovoltaic power generation efficiency far exceeds terrestrial levels, without land-use conflicts.

Heat Dissipation Dimension: The ambient temperature on the dark side of space approaches absolute zero (2.7K), resulting in extremely high radiative cooling efficiency. Traditional data centers consume vast amounts of electricity for cooling, whereas in orbit, this problem is naturally mitigated—even enabling superconducting computing using the low-temperature environment.

Latency Dimension: For AI applications requiring real-time global response (e.g., coordinated scheduling of autonomous vehicle fleets, quantitative trading in global futures markets), the physical latency of terrestrial fiber optics (speed of light in fiber is only 2/3 of vacuum) still fails to meet nanosecond-level demands. Laser link latency between LEO satellites can be compressed to less than 1/3 of long-distance terrestrial transmission.

But behind these advantages lie enormous engineering challenges.

Tech Stack Reconstruction: From "Moving Up" to "Born Up"

Most current industry discussions on space-based computing remain stuck at the stage of "stuffing GPU racks from Earth into satellites." This is clearly not the optimal solution.

A fundamental difference: Terrestrial data centers allow for anytime maintenance and hardware replacement; once deployed in space, the average repair cost for computing nodes is over ten thousand times higher than on Earth. This demands hardware with extreme radiation tolerance and ultra-long lifespan. The traditional von Neumann architecture is unfriendly in space environments—bit error rates due to single-event upsets are much higher than on Earth.

From the perspective of Physical AI, a more rational path is to redesign computing architectures specifically for space:

• Convergence of Computing and Communication: Use laser links to form thousands of satellites into a distributed computing array, where each satellite processes local sensor data and collaborates on inference via weak consistency protocols. This is precisely the natural extension of today's distributed training techniques for large models.
• Radiation-Hardened Processing-in-Memory Architecture: PIM chips based on memristors or phase-change memory reduce data movement between storage and compute units, offering inherent resilience against single-event effects.
• Cold Atom Quantum Computing: The ultra-low temperatures of space provide an ideal platform for cold atom qubits, potentially enabling ultra-large-scale quantum computers difficult to build on Earth.

[!note] A noteworthy tech trend is NVIDIA's "Space Computing Platform" released in 2024—optimizing the packaging of Grace Hopper superchips for radiation environments, employing redundant checks and dynamic voltage regulation. But this is just the first step; a true paradigm revolution requires abandoning terrestrial thinking at the architectural level.

Industry Cycle Signals: What Is the Capital Market Pricing?

This 4 billion yuan inflow isn't buying the concept of "Space AI," but betting on a fact: as the density of LEO satellite networks like Starlink breaks through thresholds (currently over 6,000 satellites), and reusable rockets reduce launch costs by two orders of magnitude, the economic viability of space-based computing is shifting from "infeasible" to "calculable."

Specifically, launch cost per kilogram has dropped from $10,000 in 2010 to around $1,500 currently. Combined with end-user demand for low latency (autonomous driving, remote surgery, etc.), the unit computing cost of space-based power is expected to approach terrestrial edge computing levels within the next 5 years.

At the paper level, OpenAI's 2023 work Orbital Compute: A New Frontier for AI Training provided a rough estimate: At an average orbital altitude of 400km, using solar panels and radiative coolers, building a 1 EFLOPS computing cluster requires an initial investment roughly 3x that of an equivalent terrestrial data center, but annual operating costs (electricity + cooling) are only 1/5 of terrestrial levels. Considering a 10-year lifecycle, Total Cost of Ownership (TCO) can be reduced by 40%.

Unveiled Core Viewpoint

Space-based computing is not a simple migration of computing power, but a dimensional reduction reconstruction of the computing paradigm. It will change where AI models are trained (from Earth to space stations), inference scenarios (from data centers to satellite edges), and even the physical assumptions of algorithms themselves (from ignoring the environment to utilizing it). When computing units are truly embedded in every corner of physical space—including orbit—Physical AI can move from laboratories to true world models.

In summary: Computing power heading to space is essentially the inevitable path for Physical AI moving from Earth to the universe; it is a renegotiation between computing architecture and physical laws.

Original Link: https://www.tmtpost.com/8066555.html