Masayoshi Son pours cold water: Is Musk’s space data center just sci-fi hype that defies common sense?

On June 23, 2026, at SoftBank’s annual shareholders’ meeting, Masayoshi Son poured cold water on Elon Musk’s “million-satellite orbital data center” plan. When asked by a shareholder whether SoftBank would emulate the initiative, Son explicitly rejected the idea. He pointed out that electricity costs account for only a tiny fraction of data center operating expenses, while moving to space would bring real-world obstacles such as exorbitant rocket transportation costs, maintenance difficulties, and communication latency. He stressed that the AI competitive landscape over the next few years is far more critical than what happens a decade from now, and that the decisive AI battle is destined to be fought on land. Whether orbital data centers represent the ultimate solution to breaking free from Earth’s resource constraints, or merely concept hype that defies engineering common sense, can be tested by comparing three sets of hard data on computing cost structures and the iron laws of physics. ### 71% Depreciation vs. 9% Electricity: The Cost Inversion of Space Computing One of the core premises behind Musk’s orbital data center proposal is harnessing the unlimited solar energy available in space to eliminate the high electricity bills of terrestrial data centers. But when measured against the real business model of AI computing, this logic rests on a fragile illusion. In the era of large-model AI computing, the total cost of ownership (TCO) structure of data centers has undergone a fundamental reshaping. According to SemiAnalysis’ breakdown of Meta’s 24,000-card H100 cluster, hardware depreciation dominates AI data center TCO, accounting for roughly 71%. This includes capital expenditure on core IT hardware such as GPUs, InfiniBand network switches, and optical modules. Because AI hardware iterates extremely fast and typically faces obsolescence within three to five years, this depreciation is the biggest cost sinkhole in computing infrastructure. By contrast, electricity costs—often seen as a major operational burden—account for only about 9% of overall TCO, even though they represent a large share of cash operating expenses. To save less than one-tenth of total costs on electricity while shouldering rocket launch expenses of thousands of dollars per kilogram is a complete commercial inversion. Currently, SpaceX’s Falcon 9 launch cost to low Earth orbit is approximately $2,720 to $4,000 per kilogram. A medium-sized AI training cluster weighs hundreds of tons in hardware alone, putting the launch cost to send it into space in the tens of billions of dollars. Space is filled with high-energy cosmic rays and solar particles that can cause single-event upsets in semiconductors, corrupting data integrity. To resist radiation, one must either use extremely expensive and performance-lagging radiation-hardened chips, or add heavy physical shielding. Either choice leads to soaring hardware costs and degraded performance, further accelerating that 71% depreciation burden. Enterprise procurement and industrial investors are not concerned with whether a utopia can be built in space a decade from now, but with the input-output ratio of every dollar spent today. Trading a massive capital expenditure penalty for a meager operating cost saving cannot close the loop commercially. ### 100-Ton Radiator vs. 10-Ton Hardware: The Heat Dissipation Penalty of a Vacuum AI infrastructure on Earth does face severe physical resource constraints, especially in heat dissipation and water scarcity. Moving data centers into the cryogenic cold of space at minus 270 degrees Celsius seems like the perfect solution for cooling, but this precisely violates the most basic common sense of thermodynamics. The background temperature of the space environment is extremely low, but the high-vacuum state means there is no air convection and no heat conduction medium. In this environment, waste heat can only be dissipated through infrared radiation—an extremely inefficient heat transfer process. According to engineering estimates by the World Economic Forum, a 1-megawatt orbital data center would require approximately 1,600 square meters of radiative cooling panels, an area equivalent to three standard hockey rinks. Referencing the existing thermal control system architecture of the International Space Station, the radiative radiator and its piping system needed to support 1 megawatt of computing could weigh as much as 100 tons. The computing hardware itself for the same computing power weighs only about 10 tons. The cooling system weighs ten times as much as the computing hardware. Based on Falcon 9 launch cost estimates, merely sending those 100 tons of radiators into orbit would cost around $300 million. This does not even include the enormous solar panel arrays and the necessary energy storage batteries. In space, solar energy is not available around the clock. Satellites lose sunlight when passing through Earth’s shadow, so an orbital data center must be equipped with a massive battery system to sustain computing operations—and those heavy batteries likewise incur expensive launch costs. The water-cooling constraints faced by terrestrial data centers are a “real-world problem” that can be solved through engineering optimization. Whether using cold-plate liquid cooling or immersion cooling, water’s specific heat capacity and phase-change heat transfer efficiency are far superior in engineering terms to vacuum radiative cooling. ### 1 Microsecond vs. 40 Milliseconds: Computing Utilization Locked Down by Network Latency The communication latency issue alone imposes a death sentence on orbital data centers under current AI technology architectures. Training large AI models is not simple web hosting or data storage; it is a microsecond-level synchronization war among tens of thousands of GPUs. In a cluster of ten thousand cards, forward propagation and backpropagation are distributed across different nodes, and gradient updates of model parameters require global synchronization through All-Reduce operations. Such synchronization operations demand extremely stringent network latency and bandwidth. Modern AI clusters relying on RDMA (Remote Direct Memory Access, a technology that allows direct memory data exchange between different computers without CPU involvement) and InfiniBand (a high-bandwidth, low-latency dedicated network protocol) typically require end-to-end latency between 1 and 5 microseconds. Only at this latency level can GPUs rapidly complete parameter exchanges during computation gaps and maintain high computing utilization. The physical latency of low Earth orbit satellite communications typically ranges between 20 and 40 milliseconds. One millisecond equals 1,000 microseconds, meaning satellite link latency is nearly ten thousand times higher than that inside a data center network. The physical limit of the speed of light determines that this gap cannot be bridged. Low Earth orbit satellites are roughly 500 to 2,000 kilometers from the ground. Even if satellites interconnect using laser communications, the physical distance and routing hops of inter-satellite links dictate that latency can never approach the microsecond level. Under such latency conditions, distributed training would leave expensive GPU clusters idle for long periods while waiting for network transmissions. Computing utilization would plummet from over 60% in terrestrial clusters to single digits. The computing power that enterprises spend hundreds of millions of dollars to purchase would spend the vast majority of its time waiting for data packets to travel from one satellite to another. ### PUE 1.09: The Engineering Limits and Solutions of Terrestrial Computing Infrastructure At the shareholders’ meeting, Masayoshi Son emphasized that SoftBank will focus on building terrestrial data centers. This is not only SoftBank’s judgment but also the consensus of the entire hyperscale computing industry. Terrestrial data centers do face challenges such as site selection difficulties, community protests, and grid interconnection queues, but these problems all have clear engineering solutions. Through continuously optimized cooling technologies and energy management, the PUE (Power Usage Effectiveness, the ratio of total data center energy consumption to IT equipment energy consumption; the closer to 1, the higher the efficiency) of terrestrial hyperscale data centers has been compressed to extremely low levels. Google reports that the annual average PUE of its large data centers has reached 1.09. This means that less than 10% of total electricity consumption is used for non-IT equipment such as cooling and lighting, with the vast majority of electricity directly converted into computing power. This has been achieved through the introduction of AI-driven cooling control systems, efficient heat recovery technologies, and large-scale liquid cooling deployment. In terms of site selection logic, computing infrastructure is shifting toward high-latitude cold regions and areas rich in clean energy. By signing long-term power purchase agreements (PPAs) combined with facility-level battery energy storage systems, data centers can break free from dependence on a single grid and build microgrid models. Water consumption is also evolving toward closed-loop systems, with immersion liquid cooling nearly eliminating evaporative water loss. The evolution path of terrestrial infrastructure is clearly visible, and the cost curve is predictable. Investors and industry observers can accurately calculate the marginal cost per FLOP of computing power over the next three years—a certainty that is an indispensable foundation in the AI race. ### The Decisive Three-Year Window: The Time Horizon of Computing Infrastructure Supporters of Musk often pin their hopes on Starship reaching mass production and bringing launch costs below $100 per kilogram, thereby fundamentally altering the business logic of orbital data centers. This long-term optimism overlooks the cruelest reality of the AI industry: the time window. Masayoshi Son pointed out that the AI competitive landscape over the next few years is far more critical than what happens a decade from now. AI is currently in an explosive phase of evolution from large language models to multimodal and reasoning models, with a massive computing power gap. Whoever can deploy hyperscale computing clusters on land faster and more cost-effectively will seize the initiative in model capability and commercial deployment. The capital expenditure cycles of large-model companies are measured in quarters, and hardware depreciation is limited to two or three years. Waiting for Starship to mature, for orbital heat dissipation technology to break through, and for radiation-hardened AI chips to be developed—these frontier explorations measured in decades—cannot solve today’s computing hunger. Enterprise procurement needs 800G InfiniBand networks and liquid-cooled H100 clusters that can go online tomorrow, not an orbital constellation that might materialize a decade later. Orbital data centers violate the cost structure of AI computing TCO, breach the thermodynamic constraints of a vacuum environment, and are locked down in efficiency by communication latency under the speed-of-light limit. Within the foreseeable engineering future, the physical destination of AI computing power is inevitably on land.

Masayoshi Son pours cold water: Is Musk’s space data center just sci-fi hype that defies common sense?

Popular Articles