SoftBank founder Masayoshi Son said global AI spending will need to reach $5 trillion per year by 2040, and rejected the argument that today's capital buildout amounts to a bubble. The figure, more than triple the current annual capex of the largest US hyperscalers combined, frames Son's thesis that the AI infrastructure race is still in its opening chapter rather than nearing a peak.
Son's $5 trillion projection sits well above consensus. Current-generation forecasts from Wall Street analysts for combined AI capex across Microsoft, Google, Meta, Amazon, and Oracle top out closer to $500 billion annually by the late 2020s. Son is projecting an order-of-magnitude expansion from there over the following decade.
The framing matters because SoftBank has already put itself at the center of that spending curve. The firm is a lead financial backer of Stargate, the OpenAI and Oracle-led US data center program valued at up to $500 billion over four years, and Son has spent 2025 lining up further AI infrastructure commitments across chips, power, and compute.
Key facts
- 01SoftBank founder Masayoshi Son projected AI will require $5 trillion in annual spending globally by 2040.
- 02Son dismissed the framing that the current AI capital buildout constitutes a bubble.
- 03The comments come as SoftBank has anchored multiple large AI infrastructure commitments, including Stargate alongside OpenAI and Oracle.
Bubble concerns have grown louder as the total announced spend across hyperscalers, sovereign programs, and private consortia has climbed past $1 trillion in cumulative commitments over the past 18 months. Skeptics point to circular financing structures, where chip vendors invest in customers who then buy chips, and to the gap between announced revenue attributable to AI products and the compute bills being run up to serve them.
Son's counter is a demand-side argument. If AI systems eventually handle a meaningful share of global cognitive labor — coding, analysis, customer service, drug discovery, logistics — the compute required to run those systems at population scale dwarfs anything currently deployed. In that framing, the constraint is not capital but power, chips, and land, and $5 trillion a year is what it takes to keep pace.
That thesis is not universally shared inside the AI industry itself. Executives at frontier labs have sparred publicly over whether the marginal returns on ever-larger training runs are holding up, and whether inference costs will fall fast enough to justify current infrastructure commitments. Sam Altman recently characterized rival Elon Musk's plans for orbital data centers as a short-term sales pitch, a reminder that even the industry's most bullish operators disagree sharply on where the buildout is heading.
Son's track record on outsized projections is uneven. His 300-year vision speeches and past forecasts on the internet economy have sometimes proved directionally right on scale and wrong on timing, and SoftBank's Vision Fund investments in the late 2010s produced heavy writedowns before the AI thesis re-rated the firm's portfolio. Nvidia, WeWork, Arm, and OpenAI have followed very different trajectories from the same balance sheet.
The bubble question also depends on what is being counted. Semiconductor capex, hyperscaler data center capex, sovereign AI programs, private lab funding, and utility investment in generation and transmission are not the same category of spend, and lumping them together produces headline numbers that mask very different risk profiles. A $5 trillion annual figure would necessarily include grid buildout and power generation, not just GPUs.
Skeptics inside finance argue the more immediate risk is not that AI fails to deliver value but that the value accrues to fewer companies than the current spend implies. If a handful of frontier labs and hyperscalers capture the productivity gains while the long tail of AI-application startups burns through capital, the aggregate return on trillions of dollars of infrastructure could still disappoint even in a world where AI works as advertised.
For SoftBank, the $5 trillion number is as much a positioning statement as a forecast. Son is signaling to co-investors, sovereign partners, and lenders that the firm intends to keep writing large checks into AI infrastructure well past the point where more cautious capital taps out. Whether that turns out to be prescient or premature depends less on the trajectory of model capabilities than on how quickly the revenue side of the AI economy catches up to the compute side — and on that question, Son is effectively betting that time is on his side.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
