Reader Response to "AI Overinvestment" (www.mbi-deepdives.com)

Reader Craig Melillo responded to a column on possible AI overinvestment with a detailed, grounded scenario analysis: hyperscalers (AMZN, META, GOOG, MSFT) have been spending heavily on AI capacity in a “tails I win” play because the discretionary spend hasn’t crippled balance sheets, but that bet could be challenged if one of three events occurs—(1) scaling laws asymptote and AGI remains elusive (he notes if AGI isn’t achieved within ~3 years the odds fall), (2) cloud providers signal sufficient capacity (build times are 18–24 months so current shortages may be temporary and supply could rise 2–3x in 18–36 months), or (3) loss-making AI model wrappers run out of funding. He highlights hard technical/economic constraints: frontier model training costs tens of billions, chips can become obsolete for training every ~18 months, and most current product value is inference/token driven rather than revenue-rich training. The note draws company-specific implications: Meta must embed AI to defend consumer attention; Google balances monetization and disruption; MSFT may be optimizing for enterprise inference demand while Oracle uses OpenAI ties to fast-track hyperscale relevance; OpenAI lacks consistent cash flow despite a reported $500B post-money and faces pressure to monetize (est. $20–25B annualized exit revenue cited); NVDA benefits from keeping GPU supply tight to defend its position. Overall the piece warns that excess infrastructure and unmet monetization, not just hype, could reshape capital flows, chip demand, and the pace of frontier-model development.