Musk wants to put AI servers in orbit. SoftBank's founder argues the real AI buildout will be won on Earth in the next two to three years, long before any space based constellation can scale.
Masayoshi Son has built a career on betting big on the next era, but the SoftBank founder is now telling shareholders the AI compute race will be decided in the next few years, not from orbit. The argument is not that orbital data centers are physically impossible. It is that they are timing-blind: a satellite constellation that takes a decade to deploy and deprecate arrives too late to matter in a market where compute pricing power is being captured right now, on the ground, in 2026.
Son made the case at SoftBank's 46th annual shareholders meeting, telling investors that in the AI race, "the next few years will be far more important than what might happen a decade or so from now," according to coverage in the Japan Times and Seeking Alpha, and as documented in SoftBank's official 46th shareholders meeting deck. His target was Elon Musk's orbital data center pitch: a plan, now in front of U.S. regulators, to put AI compute on a constellation of satellites circling the Earth.
The category matters here. AI compute is the processing power that trains and runs large AI models, and the market is currently constrained: there is not enough of it. The companies that control working AI compute capacity between now and roughly 2029 will set the price floor for the entire industry. Anyone building capacity that only arrives in 2034 or 2036 is, in Son's framing, building for the wrong race.
The orbital plan, in its current form, is real but aspirational. In January 2026, SpaceX filed with the Federal Communications Commission for a million-satellite orbital data center constellation, as SpaceNews and SatNews reported, and the FCC Space Bureau accepted the application for processing in DA-26-113. That is a regulatory step forward, not a deployment. The pitch from Musk is that putting compute in orbit will eventually undercut the cost of building ground-based data centers, in part because solar power in space is continuous and cooling is free.
Son's skepticism is not the only thing worth weighing. On the TechCrunch Equity podcast, reporter Sean O'Kane offered the sharpest economic rebuttal to the orbital pitch: a satellite constellation that has to be refreshed every few years is not necessarily a cost saver. It is, in his framing, "guaranteeing that much more business" for SpaceX itself. If the orbital compute bet functions as a multi-decade capital commitment whose hardware has to be replaced on a near-constant cycle, the savings the pitch promises may never reach the customer. They may just stay inside SpaceX's launch and manufacturing backlog.
That counter keeps the critique honest. Son's argument is timing, not physics, and it lands harder than a typical CEO quote because Son is not a typical skeptic. SoftBank's 2016 takeover of chip designer Arm Holdings and the group's broader AI compute push make Son one of the more aggressive capital allocators in this market. The TechCrunch Equity podcast's Kirsten Korosec labeled the skepticism "very ironic" given that track record; that is editorial framing, not a neutral claim, but the underlying standing is real.
The wider context is also part of the story. The same week Son was speaking, SpaceX signed a compute deal with Reflection AI, an open-source AI lab. That is SpaceX already selling compute capacity, not just pitching orbital infrastructure. The broader market is in the middle of a scramble for AI capacity, with new entrants, sometimes called neo-clouds, raising fresh rounds. Neo-clouds are smaller GPU-leasing companies that rent capacity to AI firms which cannot get enough from the big hyperscalers (the largest cloud platforms, including Microsoft Azure, Amazon Web Services, and Google Cloud). Son's framing fits that scramble: the bottleneck is now, on Earth, not in 2035.
What to watch next: whether SoftBank itself puts more capital into ground-based AI compute through Arm or its other chip holdings, and whether the FCC's review of SpaceX's million-satellite filing advances beyond procedural acceptance to substantive spectrum and orbital-debris review. The orbital data center pitch is not going away. The question Son has put on the table is whether, by the time it arrives, anyone will still be waiting for it.