Observations on AI and the Capital Markets in 2025 (vinaysridhar.com)

A Reddit product lead who works with LLMs lays out a sober 2025 read on AI and the capital markets: hyperscalers (Microsoft, Amazon, Alphabet, Meta) are running capex-to-revenue ratios near 22%—well above historical 11–16%—and are spending ~ $320B on AI infrastructure this year. That math implies investors expect $1.4–2.1T in incremental annual revenue to justify current build-outs. Yet operational realities lag: MIT’s NANDA finds only ~5% of pilots scale, integration—not model quality—is the bottleneck, and enterprise AI revenue is most likely to show up first in digital ads (higher ROAS) and developer productivity (coding tools). Valuations are elevated but not at 2000 levels (Nasdaq forward P/E ~28X vs >70X then), IPO activity is muted while private markets and private credit balloon ($1.7T AUM), and many deals and SPVs are shifting capex off balance sheets, creating opacity. Key technical and market implications: GPUs are iterating faster (annual launches), but companies are stretching depreciation schedules (4–6 years), masking true costs—shortening those schedules would materially hit pre-tax profits and market caps. Expect a multi-year adoption curve (cloud took ~10 years), a likely short-term correction in 2026–27 as P&L impact lags infrastructure, and eventual concentration of value among a few winners. The piece warns investors to watch private-credit structures, depreciation policies, and early revenue signals from ads and coding as leading indicators.