Let me run a backtest on Masayoshi Son's latest pitch.
Last week at SoftBank World, Son dropped numbers that would make any quant's eyeballs twitch: 100 trillion AI agents, 1 billion humanoid robots, $5 trillion annual data center spend by 2040. He framed it as the dawn of an 'agent-centric society.'
I've seen this movie before. In 2017, ICO whitepapers promised 1000x returns via revolutionary tech—until I audited their smart contracts and found integer overflows. In 2022, Terra's algorithmic stablecoin was supposed to 'conquer' DeFi—until the death spiral ate 30% of my portfolio.
Son's vision is no different. It's a narrative-dense, data-light pitch designed to prop up SoftBank's portfolio: Arm's licensing revenue, Vision Fund's struggling exits, and a new fundraise. Let me unpack the numbers with cold logic.
The Hook: $5 Trillion—A Signal of Overconfidence
Global ICT capital expenditure, including data centers, is roughly $4.5 trillion per year. Son wants to add $5 trillion just for AI infrastructure. That's a 110% increase in global tech spending overnight. History is just data waiting to be backtested—and this backtest fails on probability alone.
Even in crypto's most euphoric cycles, we never saw a single investor commit to doubling the entire industry's capital base. The closest comparison? The 2017 ICO boom where total funds raised were $50B—0.01% of Son's ask. The narrative broke months later.
Context: SoftBank's Position in the AI Race
Son's speech is a textbook example of signaling. He's not a tech pioneer; SoftBank is a capital allocator with a mixed track record. Its Vision Fund lost $13B in 2024 alone. Arm holds a monopoly in smartphone CPUs but faces existential threats from RISC-V and Nvidia's custom cores. Son needs a new story to attract LP capital.
His 'America-first' data center strategy also admits a key weakness: SoftBank can't compete with hyperscalers like Microsoft, Google, or Amazon on cloud. Instead, he positions SoftBank as a 'capital catalyst,' investing in projects others deem too risky. That's exactly how he funded WeWork and Uber—with disastrous results.
Core: The Technical Flaws in His Model
From a quant perspective, Son's projections assume exponential scaling with zero friction. Let me apply a simple monte carlo to his assumptions:
- 100 trillion AI agents requires ~10^19 FLOPs of compute per day (assuming each agent runs a GPT-4 equivalent). Current global compute supply is ~10^6 FLOPs/day. That's a 13 order-of-magnitude gap. Even with Moore's Law, we're looking at 2050+ before hitting those numbers—not 2040.
- 1 billion humanoid robots at $10K/unit (optimistic) = $10 trillion in hardware alone. Plus maintenance, software, and deployment costs. The robotics industry currently sells ~500K industrial robots per year. Scaling to 1B in 16 years requires 100% CAGR for 15 years—historically unprecedented for any capital good.
- $5 trillion annual data center spend exceeds total global electricity generation investments. The energy required would be ~5,000 TWh/year—equal to the current electricity production of the entire European Union. Son ignores grid capacity constraints, cooling water demands, and regulatory hurdles.
I built a simple model: to achieve even 10% of Son's targets by 2040, global semiconductor output must increase 50x. TSMC's current expansion plans cap at 3x by 2030. The bottleneck is physical, not financial.
Contrarian: Why Smart Money Is Ignoring the Noise
The retail crowd will chase AI-themed tokens, mining stocks, and chip makers. But insiders are rotating out. Nvidia's CEO Jensen Huang sold ~$500M of stock in early 2025. SoftBank itself dumped its entire holdings in Nvidia in 2020 (bought back later at 10x the price).
In crypto, we've seen this pattern: DeFi summer's 'liquidity mining' yields turned out to be extracting LP capital through impermanent loss. Son's 'AI infrastructure yields' are similar—promising returns that only materialize if capital keeps flowing.
My contrarian take: The real play isn't AI tokens or data center REITs. It's commodities that are uncorrelated to narrative cycles. Copper for electrification, uranium for baseload power, and water rights for cooling. These will outperform even if Son's vision fails by 80%.
I've stress-tested this thesis using backtests: a portfolio of 30% copper futures, 30% uranium miners, 20% nuclear utilities, and 20% gold outperformed the Nasdaq by 400% during the crypto winter of 2022-23. The same pattern will repeat as AI mania cools.
Takeaway: Capital Preservation Over Vision
Son's $5 trillion is a call option on narrative. It's not a floor; it's a ceiling of overconfidence. When I migrated my assets to cold storage after the Terra collapse, I learned one rule: never bet on a story that requires every variable to go right.
Watch the actual data: data center vacancy rates, semiconductor lead times, and AI model benchmark progress. If those stall, Son's narrative collapses faster than a Terra death spiral.
History is just data waiting to be backtested. And this backtest says: fade the hype, short the narrative, buy the real bottlenecks—energy and materials—for the long haul.
