Masayoshi Son Cuts Through the Noise
Masayoshi Son didn't mince words at SoftBank's recent shareholder meeting. While the tech press spent months nodding along to Elon Musk's grand vision of space-based data infrastructure, Son went straight for the throat of the hype. Building data servers in orbit won’t do much to cut costs, he argued. Even worse, it takes too long. In Son’s words, "in the battle for AI, the next few years will be far more important than what might happen a decade or so from now." He has a point. A devastating one, actually. The current generative AI arms race is happening right now, in real-time. A solution that takes ten years to launch isn't a solution—it's a distraction.
Now, Son isn't exactly a neutral observer in this fight. He is talking his own book, just like everyone else. SoftBank is heavily invested in massive, earthbound data center projects, and their portfolio relies on traditional power grids and terrestrial real estate. Still, Son playing the role of the skeptical pragmatist is the kind of reality check the industry needs. The TechCrunch Equity team pointed out the hilarious irony of Son being the one to preach caution. This is a man whose company literally threw billions at WeWork and powered some of the most speculative, high-flying bubbles of the last decade. But hypocrite or not, Son's question cuts through the noise: why are we trying to build server farms in the vacuum of space when the battle is being won on the ground?
The Cold Math of the AI Compute Arms Race
The tech industry is desperately compute-constrained. AI startups and enterprise tech giants are gasping for GPU capacity, and they are doing whatever they can to lease it. We are seeing a wild gold rush of neoclouds—companies that exist solely to rent out GPU power. The silicon boom is so intense that even bankrupt consumer brands are attempting to pivot into the compute hosting market. For instance, the shoe brand Allbirds, which recently emerged from bankruptcy, is reportedly reinventing itself as a GPU hosting provider. It sounds like a joke, but it's the reality of today's market. Timothy Fernholz detailed this bizarre transition in his interview with the company's new leadership.
Meanwhile, chipmakers like Groq are raising massive war chests—securing $650 million in new funding—to expand their inference operations. Everyone wants a piece of the AI pie, and they want it today. Building a terrestrial data center is hard. You have to secure land, navigate local zoning laws, pacify NIMBYs, and negotiate with utility companies for gigawatt-scale power. But you can pour concrete and run fiber in under a year.
You cannot launch a constellation of high-performance compute satellites in a year. The timeline is completely out of sync with the pace of software development. By the time anyone deploys a stable, multi-satellite orbital computing array, the state of the art in AI will have moved five generations ahead. The immediate bottleneck isn't space—it's execution speed.
Talking Your Own Book in Orbit
The debate over orbital compute isn't a clean academic discussion. It is a corporate street fight where every participant is trying to pump their own stock or defend their market share. Elon Musk's promotion of orbital data centers makes complete sense when you look at his cap table. Since SpaceX's IPO, the company's valuation has to be justified by more than just launching GPS satellites and bringing astronauts to the International Space Station. The public markets want to see SpaceX as a core AI infrastructure play. If you can convince Wall Street that Starship is actually a vehicle for deploying massive space-based server farms, it helps sustain that colossal valuation, which at one point briefly passed Amazon.
On the other side of the fence, you have people like Sam Altman, who has publicly rolled his eyes at the idea of orbital server farms. Altman's skepticism is partly rooted in engineering, but it's also personal—he and Musk have a long, bitter history of public spats and competing AI visions. And then there's SoftBank, which has committed billions of dollars to terrestrial energy projects and land-based GPU clusters.
None of these executives are objective observers. They are all talking their own book. When you look at the predictions coming out of the valley, you have to read them with an asterisk. Musk predicts space compute because SpaceX owns the rockets. Son predicts terrestrial compute because SoftBank owns the land and the local infrastructure.
Radiation, Radiators, and the Laws of Physics
The marketing decks for orbital data centers make it sound simple: space is cold, and solar energy is free. But the actual physics of orbit tell a much grimmer story. Space is a hostile, unforgiving environment for advanced semiconductors.
First, there is the radiation problem. Earth's atmosphere protects ground-based chips from cosmic rays and solar particles. In low Earth orbit (LEO), there is no such shield. High-energy particles constantly bombard silicon, causing single-event upsets—bits flipping in memory—and physical degradation of the processors. To prevent this, you have to use radiation-hardened chips. But radiation-hardened silicon is always several generations behind the state-of-the-art consumer GPUs in terms of training speed and efficiency.
Second, the cooling problem is a nightmare. Yes, space is cold, but it is also a vacuum. On Earth, we cool servers using conduction and convection—blowing air across heat sinks or running liquid through cooling blocks. In a vacuum, conduction and convection do not exist. The only way to reject heat is through thermal radiation, which is incredibly inefficient. To cool a dense rack of modern AI chips in space, you need colossal, heavy radiator panels that dwarf the satellite itself. It turns out that dissipating heat in a vacuum is far more expensive than cooling a warehouse in Iceland.
Finally, there is the maintenance gap. When a server chip dies in a warehouse in Virginia, a technician walks down the aisle and swaps it out. If a GPU fails in orbit, it stays broken. You either launch a replacement satellite at the cost of millions of dollars, or you watch your orbital cluster slowly decay over time.
SpaceX's Launch Monopoly Loop
SpaceX currently controls 80% to 90% of the global launch market. It is an astonishing monopoly, but it is also a self-referential one. The vast majority of SpaceX's high launch cadence is driven by Starlink launches. If you subtract Starlink's internal launch manifest, SpaceX's share of external commercial and government launches falls closer to 20% or 30%.
This is why the orbital data center concept is so brilliant for SpaceX's business model. It creates a massive, permanent internal customer for the launch division. Because these low Earth orbit satellites experience orbital decay and atmospheric drag, they must be replaced every few years. If SpaceX can convince the tech sector to host its models in space, they guarantee themselves an endless cycle of launch revenue. Every satellite that burns up in the atmosphere is another launch that needs to be booked.
It is a beautiful business loop, but it benefits SpaceX far more than it benefits the AI companies renting the compute. SpaceX is already testing these waters, signing its first post-IPO contracts to lease out space-based compute to smaller corporate clients. Whether these clients ever see a return on their investment is secondary. For SpaceX, the rocket is already paid for.
The Immediate Battleground is Here on Earth
In the end, the AI infrastructure battle will be won on the ground. The physical realities of power, cooling, and latency mean that the vast majority of training and inference will stay earthbound for the foreseeable future.
While the concept of space-based compute is a fascinating engineering exercise, it is a poor fit for the immediate needs of the industry. The next five years will determine which AI models dominate the market. During this critical window, the company that can secure 100 megawatts of terrestrial grid power and a warehouse full of liquid-cooled racks will win. The company waiting for a Starship to drop a server constellation into orbit will be left behind. SpaceX's real strength will remain its ability to act as a utility, but as Son rightly noted, the timeline for space compute simply does not match the urgency of the AI revolution.
Until someone can bypass the laws of thermodynamics and launch an orbital data center that works next Tuesday, we should keep our servers on the ground.
