SoftBank's Masayoshi Son: Super AI Requires $9 Trillion and 400 Gigawatts
SoftBank Group Chairman Masayoshi Son stated that to create a super artificial intelligence that is about 10,000 times smarter than the human brain, approximately 400 gigawatts of power and a total capital expenditure of about $9 trillion for hardware and data centers are needed, mentioning a scale of around 200 million accelerator chips.
At the Future Investment Initiative in Saudi Arabia, he said that while $9 trillion seems excessive to many, he believes it is still reasonable and may even be on the low side. If AI only rewrites about 5% of the global economy, the annual output could match this investment; profits will concentrate in a few companies, with individual firms potentially generating trillion-dollar annual profits. He also stated that NVIDIA is undervalued due to the larger workloads expected in the future.
He envisions a narrative from 2025 to 2026 focusing on "physical superintelligence": the bottleneck will no longer be algorithms, but chips, power, and data centers. At a shareholder meeting, he set a target for net asset value to reach 1,000 trillion yen, about $6.2 trillion, within ten years, referring to himself as a "golden egg-laying swan." SoftBank considers Arm and OpenAI as dual engines and continues to invest heavily in OpenAI; large investments and bridge loans related to this are expected to be in the tens of billions of dollars for the fiscal year 2025.
On the power side, he later mentioned that mid-term data centers will still rely on natural gas, with fusion potentially taking over in fifteen years, and he discussed the vision of data center capacity reaching 3 terawatts by 2040. He framed the idea of recouping costs in nine years, with the four major companies sharing $4 trillion in net profits, and projected NVIDIA's market value in ten years to be in the hundreds of trillions of dollars as a law of the market extrapolated from his scale, not based on disclosed company profit forecasts.
Buyers are platforms that need to secure power and chip quotas, while sellers are the power grid, wafers, and land. Funding flows from SoftBank's leverage and Gulf funds to data centers. The beneficiaries are a few giants that can turn power supply into a moat; those under pressure are the power grids that interpret industrial loads based on residential electricity prices. This is a speech of founder-level magnitude, not a signed $9 trillion order.
Source: Public Information
ABAB AI Insight
Masayoshi Son frames his vision as a power equation: the intelligence multiplier first translates into 400 gigawatts and 200 million chips. The premise of recouping $9 trillion in "one year" is that 5% of global output is captured by this infrastructure, with profit margins close to software. This treats capital expenditure directly as revenue's inverse, omitting electricity prices, depreciation, model failures, and regulation. SoftBank's own actions are more concrete: using bridge loans to acquire OpenAI and chip assets, leveraging Arm as a licensing layer, and supporting stock prices with net asset value stories.
The capital path is vision financing. The motivation is to seize the physical layer after algorithmic differences diminish—whoever occupies substations and packaging capacity first will bill the labs. The four giants sharing trillion-dollar net profits replicates the concentration of the platform era into the power era. NVIDIA's market value extrapolation assumes constant profit margins, ignoring clients' self-developed chips and power limits.
This is similar to how he once wrote a 300-year vision into the Vision Fund, akin to how Bezos frames everything as a flywheel. The industry is transitioning from model expansion to power control. Whoever can turn gigawatts into financeable assets will collect tolls during the construction phase before superintelligence is achieved in the lab.
The essence is capital concentration. The mechanism uses an unverifiable intelligence multiplier to price current power stations and wafers. The larger the multiplier, the easier it is to portray current capital expenditure as cheap. The recouping story serves the next bridge loan, not the already connected 400 gigawatts.