What happened
Son used a public forum Tuesday morning Tokyo time to put a number on what he thinks AI infrastructure will cost the world by 2040. The figure: $5 trillion a year, every year, roughly the size of Japan's entire GDP. He explicitly rejected the comparison to the 2000 dot-com peak, arguing that today's AI spend is backed by real revenue at the hyperscaler layer, not just equity issuance. CryptoBriefing reported the remarks, which come as SoftBank continues to lean into Arm Holdings and its OpenAI-adjacent stakes.
The framing matters because Son is the same executive who called the 1999 internet buildout correctly on direction and catastrophically wrong on timing. His track record cuts both ways. But the specific dollar figure is what desks will remember. $5 trillion annually by 2040 is roughly twice the current global military spend and about ten times what hyperscalers collectively deployed on AI capex in 2025.
Why it matters
For crypto, the readthrough runs two ways. First, the AI-token complex (Fetch.ai, Render, Bittensor, Akash) has been trading as a beta on the broader AI narrative rather than on protocol fundamentals. A senior industry principal putting a $5T floor on the multi-decade trade gives that beta a fresh anchor. Second, and less discussed, is the energy side. Son's number implicitly requires a power buildout that dwarfs anything in the current grid pipeline, and Bitcoin miners with locked-in power contracts (Core Scientific, TeraWulf, Iris Energy) have been repricing as HPC hosts for exactly this reason.
The contrast paragraph belongs here. Son's confidence looks good on a slide. The actual bottleneck isn't dollars, it's transformers, permits, and gigawatts. Every hyperscaler CFO on the last earnings cycle flagged power availability as the binding constraint, not capex authorization.