The AI Trade Can Lose Money Even If the Technology Doesn't — BofA Names the Trigger

On Oct. 16, 2026, Bank of America investment strategist Sebastian Raedler published a note arguing that AI's underlying technology can succeed while investors in AI stocks still earn poor returns — the same 'capital-cycle disappointment' that ended the 1870s railroads and the 1990s fiber buildout. The thesis landed on a tape that had already lost roughly $2 trillion in market value over the prior week and tipped the Nasdaq100 down another 2.5% on Friday.

What makes Raedler's framing different from the previous warnings this year — Bloomberg's Oct. 5 'end of easy money' note, BofA's own $8T Hartnett cash pile (Oct. 9), or Goldman's $300B revenue breakeven math (Sep. 25) — is that the failure mode is no longer hypothetical. Average AI token prices have fallen roughly 50% since the summer, the FT reported on Oct. 8 that OpenAI's annualized revenue is closer to $50B than the $70B previously signalled, and U.S. business AI spending has eased since August even as token volumes have surged. The capex-vs-revenue gap is no longer closing — it is widening.

Raedler's headline number: forward12-month consensus expectations for hyperscaler capex have roughly tripled from $330 billion in early 2025 to about $1 trillion today. That single figure is larger than the 1870s railroad buildout (peaked at 2.2% of GDP) and the 1990s tech buildout (1.1%) — current spending is running at roughly 3% of U.S. GDP, according to BofA. Goldman Sachs' baseline now sees AI capex at $765 billion in 2026 and $1.2 trillion in 2027, with the firm estimating hyperscalers need ~$300 billion of annual AI revenue to break even on the spend.

The Oct. 16 note leans on a single observation: token prices — the unit price of actually running an AI model — have fallen ~50% since the summer as competition from Chinese open-weight models intensified, while token volumes keep rising. When unit price falls faster than volume grows, revenue stalls. That is exactly the dynamic Raedler points to as 'the failure mode seen in past capital cycles,' and it is the same dynamic that broke the 1999–2002 fiber buildout.

Microsoft's fiscal 2026 capex of $115.95 billion absorbed 64% of operating cash flow — up from 47% in fiscal 2025 — and free cash flow of $66.99 billion was the lowest since fiscal 2021 in absolute terms despite a 16% jump in revenue. Amazon's situation is more extreme: fiscal 2025 capex of $131.8 billion was 17x free cash flow of just $7.7 billion. If AI capex pulls back, these two stocks face the largest FCF restoration tailwind; if it does not, they face the largest dilution to per-share earnings.

The PHLX Semiconductor Index fell 3.8% on Oct. 16, its eighth straight losing session — the longest such streak since 2008 — and erased more than $2 trillion in market capitalization from AI-linked stocks over the prior week. The SOXX ETF, which tracks the broader chip complex, is now down roughly 16% from its Oct. 6 high of $589 and about 25% below its 52-week high of $655.95. The trigger was not Raedler's note alone — the prior session (Oct. 8) saw the FT's OpenAI story, and Oct. 16 also brought hot CPI and a surge in oil — but BofA's framing put a name on the move.

Over the next 30 days, five prints will determine whether the AI-returns paradox becomes a thesis or a tantrum. Meta is first, on Oct. 21; Microsoft and Alphabet follow on Oct. 28 (the same day SK Hynix reports in Seoul); AMD on Nov. 3; and Nvidia closes the run on Nov. 17 with the Q3 FY2027 print, where management's own Aug. 26 outlook is for $108.0 billion of revenue (±2%). Every one of these reports will be read against the same question: is the 2026 capex number still the right number?

BofA's 'capital cycle' framing does not hit all AI stocks equally. The asymmetry is the story: upstream suppliers to the buildout (chips, memory, packaging) absorb the spending first and most directly; downstream buyers (hyperscalers, neoclouds, model labs) bear the monetization risk. Below the surface, the marginal bets are moving from 'will AI work?' — consensus says yes — to 'whose revenue stream actually monetizes the spend?'

Days to quarters: the next 30 days are binary catalysts, one print at a time. Meta on Oct. 21 is the highest-stakes single event because it carries both an above-consensus 2026 capex range ($130-145B) and the lowest operating margin of the four hyperscalers — so the 'even if works' argument cuts deepest there. A guide-down on capex, or a guide-down on 2027 capex, is what triggers the multiple compression in Nvidia and Broadcom that the Oct. 16 selloff only previewed.

One to three years: the capital-cycle analog is the 1999–2002 fiber buildout, where capex tripled, unit prices collapsed, and the eventual survivors traded sideways for nearly a decade. Raedler does not predict an immediate bust, only that capital cycles end in disappointment and that the closer the cycle gets to peak capex, the steeper the eventual re-rate. The token-pricing signal — prices down 50% with volumes still rising — is the early indicator that monetization is not keeping pace; if it persists through the Nov. 17 Nvidia print without a corresponding capex cut, the longer-horizon thesis becomes the base case.