Masayoshi Son Pours Cold Water on Elon Musk
On June 23 local time, Masayoshi Son, the founder of SoftBank Group, publicly poured cold water on the concept of space data centers advocated by Elon Musk at the annual general meeting of shareholders of SoftBank Mobile, its telecommunications subsidiary. He believes that although this concept seems to be able to reduce electricity costs through solar energy and the space environment, in fact, its overall economic efficiency is far inferior to that of ground-based solutions, and the gains do not justify the losses.
When responding to relevant questions during the shareholder Q&A session, Masayoshi Son said that the core selling point of space data centers is to reduce electricity costs, but this advantage accounts for a very limited proportion in the overall cost structure of data centers. The data he provided shows that in the total operating cost of the entire life cycle of current AI data centers, electricity expenditure only accounts for about 7%. What really accounts for an absolute proportion is the capital expenditure on hardware devices such as GPUs and servers, as well as the supporting investment in land, computer room construction, and network bandwidth.
Masayoshi Son said that Musk is an "outstanding innovator," but the space environment can only save part of the electricity cost, and high hidden costs need to be paid. The transportation cost of sending all equipment into space, the on-orbit maintenance cost, and the communication delay problem caused by data transmission will all become non-negligible obstacles. He said bluntly that it will take at least ten years for the commercialization of space data centers, while the AI competition will be decided in the next three to five years.
Based on this judgment, Masayoshi Son is betting SoftBank's chips on the ground. He revealed that SoftBank has invested about $65 billion in OpenAI and has become an important shareholder. It plans to invest 45 billion euros in France to build the largest data center cluster in Europe and intends to invest more than $500 billion in Ohio, the United States, to build the "largest-scale ever" AI infrastructure. Relying on nuclear power energy, land resources, and policy support, these projects attempt to reduce the unit computing power cost through economies of scale. Masayoshi Son said, "We will build a powerful computing power network on the ground to provide strong support for AI enterprises."
As the main promoter of the concept of space data centers, SpaceX, owned by Musk, has drawn a grand blueprint in its prospectus: using Starship to send million-ton computing equipment into orbit, and creating a low-cost and highly reliable source of computing power through the infinite solar energy in space and the vacuum low-temperature environment. According to the official application document submitted by SpaceX to the US Federal Communications Commission in January this year, the company plans to deploy up to one million computing satellites in the low Earth orbit at an altitude of 500 to 2,000 kilometers to build a complete on-orbit data center network. The system is powered by solar energy, realizes high-speed data transmission through inter-satellite laser links, and is interconnected with the existing Starlink satellite Internet system. SpaceX said bluntly in the application document that this plan is the first step towards a "Kardashev Type II civilization," with the goal of making full use of solar energy and breaking through the physical limitations of the Earth's local power supply on the development of AI computing power.
Currently, SpaceX has completed the design of the first-generation AI satellite "AI1." Each satellite is equipped with interchangeable computing modules, large-area solar panels, and an expandable liquid-cooled radiation heat dissipation system, which uses the space vacuum environment to achieve efficient passive heat dissipation. The schedule given by Musk is to strive to achieve an annual deployment rate of 1 gigawatt of space AI computing power by the end of 2027, and then gradually increase the deployment scale as the reusable technology of Starship matures. Musk once claimed that if one million tons of payload can be sent into orbit every year, in theory, a solar AI satellite cluster of 100GW can be built, and even a computing power scale of 1 terawatt (TW) can be achieved.
In fact, Masayoshi Son is not the first person to pour cold water on space data centers. Sam Altman, the CEO of OpenAI, publicly stated as early as February this year that Musk's concept of orbital data centers is "unrealistic." The launch cost, economic and technological challenges, and the difficulty of on-orbit maintenance of components such as GPUs make it "difficult to achieve in the next ten years." However, Altman admitted that orbital data centers have future potential, but we have not reached that stage yet.
Currently, the industry generally holds a cautious attitude towards the economic viability of space data centers. Most analysts believe that in the foreseeable future, ground-based data centers will still be the absolute main force of AI computing power. Even if space computing power is implemented, it is more likely to be first applied in scenarios such as offline training that are not sensitive to latency and have extremely high requirements for computing power scale, and it is difficult to fully replace ground facilities.
This article is from the WeChat official account “Jiemian News”. Author: Li Kefeng. Republished by 36Kr with permission.