The End of the Compute Race May Lie in Space

Looking at industry cycles, when a technology starts extending from the ground into space, it often means it has entered the deep waters of "involution" (intense competition). AI computing power is currently standing at this critical point.

Recently, discussions about "space data centers" have heated up. This isn't sci-fi plotline, but a realistic game played by the industry under computing power hunger. Let's break this down; what it reflects behind the scenes is an extreme expression of the supply-demand contradiction in AI computing power.

From "Computing Power Hunger" to "Computing Power Frontier"

Ground-based data centers are undergoing unprecedented expansion. Nvidia's GPUs, once high-end gaming hardware, have become "strategic resources" for global tech companies today. Major cloud providers and AI companies are frantically buying and hoarding computing power. But an overlooked issue is that this expansion has physical boundaries.

Power bottlenecks are the most direct. The power consumption of a large AI training cluster is almost comparable to a small-to-medium city. And electricity costs are becoming an increasingly larger proportion of computing power costs. Heat dissipation, land, and water resource consumption are forcing practitioners to think: Where does computing power go next?

Space became a seemingly crazy option. The solar energy density in low Earth orbit is almost several times that of the ground, without day-night or seasonal interference. More importantly, heat dissipation efficiency in a vacuum environment is incomparable to ground-based air cooling or water cooling. These physical "advantages" have led some pioneers to seriously evaluate: Is stacking computing power in space worth it?

From the perspective of semiconductor geopolitics, space computing power has a deeper implication. It could become a way for certain countries or enterprises to bypass ground-based physical limitations and build "computing power islands." When Nvidia chips face export controls, space computing power might offer a "physical isolation" alternative. This is no longer just a technical issue, but an extension of geopolitical maneuvering.

Crunching the Numbers: Feasibility of Space Computing Power

I tried to make a simple calculation from a cost-benefit perspective. Currently, SpaceX's Starship or other heavy-lift rockets have reduced launch costs to thousands of dollars per kilogram. But even at this level, compared to ground data centers, the initial construction cost of space computing power is still orders of magnitude higher.

But changes in cost structure are key. The bulk of ground data center costs are electricity, O&M, and land. The bulk of space computing power costs are launch and hardware radiation hardening. If future AI demand continues to expand to the point where ground power supply truly cannot meet it, the "total cost of ownership" for space computing power might reach a tipping point.

A preliminary conclusion is: In the short term (3-5 years), space computing power cannot replace ground data centers. But it might pioneer viable business models in specific scenarios first.

For example, for global communications requiring extremely low latency, or "edge computing" scenarios with extremely high data sovereignty requirements, space computing power has natural advantages. Also, for AI training tasks requiring long-term, stable, uninterrupted operation, such as foundational large model training, placing them in space can avoid issues like ground power fluctuations and network attacks.

Challenges and Layout: Who is Betting

Of course, we must also see the huge challenges. Radiation in the space environment severely shortens the lifespan of chips and memory. Thermal management in a vacuum cannot be solved by simple heat sinks. More critically, how to achieve efficient, low-latency communication between space and ground computing power? This requires a globally covering laser communication constellation.

Currently, there aren't many real industrial players. Some startups and a few aerospace giants are doing early validation. But the real variable lies with cloud providers. If Amazon, Microsoft, Google, etc., start seriously evaluating migrating some computing power to low Earth orbit, that would mean this track has truly entered the "investable" stage.

From an investment perspective, I believe space computing power is still in the "0 to 1" stage. The real opportunity might not lie in the computing power itself, but in the infrastructure supporting this ecosystem, such as radiation-hardened chips, high-power laser communications, and new thermal materials. These secondary market targets are worth paying attention to in advance.

Original link: https://www.tmtpost.com/8071099.html