Masayoshi Son says AI will cost $5tn a year by 2040, and calls bubble talk absurd

Masayoshi Son said building artificial intelligence could cost about $5 trillion per year (800 trillion yen) through 2040. The comment was made at SoftBank’s annual corporate conference in Tokyo, but no specific funding plan or timelines for how costs would be met were provided.

Analysis

This reads less like a fundamental update than an attempt to re-anchor expectations for the AI capital cycle. The market takeaway is not the absolute dollar figure; it is that the financing layer of AI is becoming as important as the model layer. If capital intensity keeps rising, the beneficiaries shift from pure software to picks-and-shovels: semicap equipment, advanced packaging, power delivery, cooling, and data-center infrastructure. That is where incremental spend usually clears into visible orders with better pricing power.

For SoftBank, the second-order issue is balance-sheet optionality rather than near-term earnings. The company can stay relevant if it can monetize stakes and recycle capital into AI infrastructure assets, but that strategy is highly dependent on capital markets staying open and on outside investors tolerating long-duration payoffs. If public-market AI names de-rate, the leverage in that ecosystem can turn from strategic advantage into forced asset-sale risk. Watch for any sign that the narrative shifts from "investing in AI" to "funding AI capex"—those are very different risk profiles.

The contrarian view is that the consensus may be underestimating ROI compression in AI infrastructure. If every participant chases the same buildout, return on incremental capex can fall faster than revenue expands, which would pressure high-multiple AI software and some model-layer names first. The cleanest falsifier is a slowdown in hyperscaler capex guidance or a widening gap between AI spend announcements and monetization metrics over the next 1-3 quarters; that would quickly deflate the thesis and shift leadership back toward cash-generative large-cap tech rather than the more levered AI ecosystem.

Key Decisions for Investors

• No immediate directional trade in SOBKY; treat this as a watch item until there is a disclosed financing vehicle, asset sale, or committed AI spend plan. Time horizon: 1-3 months. Falsifier: no concrete capital deployment or monetization catalyst.
• Long a basket of AI infrastructure enablers on pullbacks versus software beta: ASML/AMAT/LRCX or a semicap proxy over IGV, 6-12 months. Risk/reward favors the names that convert capex into orders and pricing power first.
• Pair trade: long power and data-center infrastructure beneficiaries such as VRT/ETN/PWR versus short a high-multiple AI software basket, if hyperscaler capex keeps expanding without commensurate revenue conversion. This expresses the idea that physical bottlenecks monetize before model-layer optimism does.
• If trading SoftBank exposure via ADRs, use rallies only after evidence of asset monetization or external funding support; otherwise fade strength. The risk is not missing AI upside, but funding friction if the capital cycle tightens.