Too Much AI, Too Soon (www.uncoveralpha.com)

A senior investor warns that AI enthusiasm has outpaced economic reality: while they remain convinced AI is a long-term transformational force, they’ve trimmed tech/AI positions because short-term market valuations, projected capital needs and financing terms look risky. Key concerns are (1) a shortage of organic capital leading to “creative” financing (Nvidia’s rumored $100B support for OpenAI, xAI’s $20B SPV with $12.5B debt, Meta’s $29B data‑center financing, Oracle’s $38B debt), (2) GPUs depreciating far faster than many firms assume, and (3) valuations that underweight the chance of a slowdown. The author estimates AI labs will need hundreds of billions to trillions in CapEx (OpenAI planning ~26GW of data center buildout; the author cites ~ $60B per GW leading to >$1.5T), while current revenues and free cash flow (OpenAI ~$15–20B revenue, rising losses) can’t cover it. Technically, the piece stresses we’re shifting into an inference-intensive, energy‑constrained phase where “tokens per watt” rules economics. Nvidia’s shift to ~1‑year product cycles (Hopper→Blackwell driving 10–20x reductions in token cost) means GPUs obsolesce faster; amortization should be ~1–2 years, not 3–6 as many neoclouds/hyperscalers model. That mismatch inflates near‑term profitability and risks large future write‑downs, strained debt collateralized by chips, and Nvidia acting as backstop/lender of last resort. Implication: higher funding costs, consolidation or government intervention risk, and a possible re‑rating of AI valuations as true CapEx/OpEx economics emerge.