“AI Data Centers to Space” — Elon Musk on the “Power Wall” and the Next Winning Strategy
“Within 36 months, the 'cheapest place' to run AI will be space” — When Elon Musk makes this assertion, the conversation shifts from science fiction to a discussion of “supply constraints.” The “Orbital data centers” concept referred to here is not an idea about changing where GPUs are placed, but an argument for bypassing terrestrial bottlenecks—power, permits, equipment supply, and capital—entirely.
In this article, we will break down the logic of his statement into “what the constraints are and why space is the solution,” while also organizing “where the leaps are and what the realistic points of contention are.”
1. The “target for optimization” was never the GPU, but power
The host side raises doubts: “Energy accounts for 10–15% of a data center’s total cost of ownership, while the main cost is the GPU. If you put them in space, maintenance is difficult, so isn't it actually disadvantageous?” Elon Musk does not shift his focus from this point.
“The problem is the ‘availability’ of energy.”
“(Outside of China,) the total power supply is almost flat. But chip production is increasing exponentially. How are you going to ‘power’ those chips?”
In other words, his premise is that the essential issue is not the price or depreciation of individual GPUs, but the inability to secure “sufficient power, at sufficient speed, and at sufficient scale.”
1-1. The premise that “the ground is slow due to regulations and the grid”
What is repeated throughout the discussion is the slowness of permits and grid interconnection.
“Cover Nevada in solar? Try getting the permits.”
“Space is easier to scale than the ground. Space is a ‘regulatory play.’”
The “regulation” here includes not only environmental regulations but also the procedural costs of the entire power infrastructure, such as transmission grids, substation equipment, grid interconnection reviews, and land acquisition.
2. The “cost advantage” of space lies in solar utilization rates and the lack of need for storage
As a reason for placing them in space, he cites a difference in cost structure that goes beyond “space has stronger sunlight.”
“Solar in space is about 5 times more ‘effective’ than on the ground. That’s because you lose 30% through the atmosphere, and there are clouds, seasons, and day/night cycles.”
“In space, you don’t need batteries. Because it’s ‘always sunny.’”
In short, terrestrial solar power inevitably involves
fluctuations in generation (day/night, seasons, weather)
and energy storage/grid adjustment to fill those gaps, which he argues becomes a friction to scaling.
2-1. The logic behind the “space is 10 times cheaper” argument
He goes further, stating, “It’s not 5 times, it’s 10 times. Because batteries are unnecessary.” The point here is not the strictness of the numbers, but the assessment that terrestrial costs balloon when you include “generation + adjustment.”
And he concludes with this:
“Within 36 months (or rather 30 months), the most economically attractive place to put AI will be space.”
3. Counterargument 1: GPUs break. Can they be “maintained” in space?
This is the most intuitive and strongest counterargument. Failures are common in training, and replacement is difficult in space.
To this, Elon Musk responds that “infant mortality can be eliminated on the ground” and that “GPUs that have run for a certain period are surprisingly reliable.”
“Infant mortality can be eliminated on the ground. Once you get past the initial debugging, the chips are quite reliable.”
“So, the ‘unserviceable problem’ is not the essence. Remember what I said.”
This part is closer to a “bet” than a persuasion. It involves how to incorporate radiation, thermal design, and redundancy for space operations, and since failures cannot be reduced to zero, it requires “design based on the premise of no replacement (mass-produced redundancy, hot-swappable architecture).”
4. Counterargument 2: Can’t this be solved on the ground with private power generation?
The conversation shifts to real-world examples from xAI. There is a vivid story about how they gathered their own power generators (gas turbines) for a massive cluster (Colossus 2) and crossed state lines due to permitting issues.
“Getting (1GW) online was crazy. We scrambled for turbines and crossed state lines because of permitting issues.”
“Calculating things like ‘GB300 power consumption × number of units’ is for amateurs. You have to add cooling, networking, and reserve margins.”
4-1. “Number of GPUs → Required Power” is not a simple proportion
He explains that it is not just the servers, but also
cooling (matching the peak at worst-case scenarios)
network, CPU, and storage
and the reserve margin to account for maintenance downtime of power generation equipment
that multiply and increase.
“If you want to run 330,000 GB300s, you need about 1GW (including everything).”
He emphasizes the gut feeling that “scaling on the ground” becomes rapidly difficult in terms of power, equipment, and operations.
5. The bottleneck shifts from “turbines” to “foundries”: A chain of supply constraints
What makes the conversation interesting is that the constraints are not abstract theories but fall to the end of the supply chain.
“Turbines are sold out until 2030.”
“The real constraint is the ‘blades/vanes.’ There are only three companies in the world that can cast them.”
In other words, it doesn’t end with “just build a power plant”; the casting capacity for parts becomes clogged. This chain of supply constraints supports his conclusion that “expansion on the ground is bound not by ‘will’ but by ‘supply capacity.’”
6. If you go to space, the next constraint will be “chips and memory”
Even if you remove power constraints in space, the next wall is semiconductors. Here,ASML, TSMC, Samsung, and Nvidiaenter the conversation.
“What China hasn’t been able to replicate is not TSMC, but ASML”
“The biggest concern is memory. The rise in DDR prices is a sign of that.”
The point here isthat the space data center theory does not end as a "space story," but presents a roadmap where constraints shift from power → equipment → semiconductors → memory as it progresses.
7. The Scale Argument: Terawatts and the "Kardashev Scale"
The discussion finally leaps into the theory of civilization. Using the comparison that the average power consumption of the United States is about 0.5 TW, he states, "If AI demands 1 TW, the terrestrial power system itself will be forced to transform."
Furthermore,
"The Earth receives only a tiny fraction of solar energy. If we are to scale, we have no choice but to harvest solar power in space."
He connects this to the conclusion that "space is the only means of scaling."
This part jumps from a discussion of technology and economics to the "energy limits of civilization." However, his logic is consistent, converging on the perspective that the limits of scale are determined by physics and supply chains.
Conclusion: Space Data Centers are a Proposal for "Constraint Avoidance," Not a "Dream"
The core of this conversation is not that "space is romantic." Elon Musk says this throughout:
AI scaling on Earth will be blocked by power supply capacity and regulatory procedures.
Even if you try to escape with self-generation, you will eventually be blocked by the turbine supply chain (and further, the casting capacity for blades).
Space changes the structure of "power generation + regulation" due to solar uptime and the lack of need for energy storage.
However, once you go to space, chips and memory will become the next constraints.
In short, it is not that "space is a panacea," but a design of "where to shift the constraints."
And the provocative deadline—"36 months"—is less of a technical prediction and more of a declaration of a bet that a phase will arrive where space looks "relatively cheaper" as terrestrial constraints are exposed first.