Top economist warns that the AI math doesn’t make sense: ‘Profits are currently being funded by investors rather than earned from customers’

Apollo Chief Economist Torsten Slok warns that AI profit margins are wildly mismatched across the value chain: silicon & equipment leads at ~41% profit margins while models & applications show ~-59% operating margins, implying AI profits are “funded by investors rather than earned from customers.” BIS notes hyperscaler AI capex is outpacing earnings/free cash flow, with Bank of America estimating $121B of debt issuance by the five hyperscalers in 2025 (4x the prior five-year average), raising tail risk that a financing slowdown turns the capex boom into an investment bust. Separately, a cited example highlights Oracle’s negative cash flow (~-$23.7B, fiscal 2026) and large AI infrastructure commitments, underlining the risk that upstream AI spending may not convert into end-customer ROI quickly enough.

Analysis

The key market issue is not whether AI is “real,” but who is forced to finance it. When the marginal buyer of infrastructure depends on external capital rather than end-demand cash flow, the equity story becomes a balance-sheet story; that makes leveraged infrastructure plays the most fragile part of the stack. ORCL is the cleanest expression of that fragility: the market is implicitly assigning it utility-like durability while its economics still look like a financing vehicle for someone else’s growth.

The second-order winners are the companies that can monetize every incremental dollar of capex early and walk away with the cash, especially NVDA and, to a lesser extent, AMD. But if hyperscaler spending pauses, the first pain is not just at the chip layer; it propagates into networking, memory, storage, data-center REITs, and power procurement expectations. MSFT, GOOGL, AMZN, and META may actually see a near-term “discipline premium” if they slow spend, because the market will re-rate FCF and buybacks faster than it will reprice AI optionality.

The real catalyst path is 1-3 months: earnings guidance, capex commentary, and credit spreads for AI-heavy issuers. The thesis is falsified if hyperscaler capex keeps accelerating while AI-related revenue/margin disclosure begins to show credible payback, or if ORCL converts its backlog into durable free cash flow without widening leverage metrics. Over 6-18 months, the bubble risk shifts from equity multiples to funding availability; the first place to watch is whether debt markets stop rewarding “AI adjacency” with cheap capital.

Key Decisions for Investors

• Short ORCL on strength or buy 3-6 month puts; the risk/reward is asymmetric because the equity is most exposed to a funding freeze and multiple compression if AI capex growth slows.
• Pair trade: long NVDA / short ORCL for relative balance-sheet quality and pricing power; use it as a spread, not a market-direction bet, and cover if semiconductor order growth broadens beyond hyperscaler demand.
• Overweight MSFT and GOOGL versus ORCL on any AI-capex scare; if management pivots toward efficiency, these names can absorb slower spend with higher FCF and less leverage risk.
• Set a credit-spread alert on ORCL and AI-linked investment-grade issuers for the next 1-2 earnings cycles; widening spreads would be the cleanest falsifier/confirmation signal for the bubble-financing thesis.
• Do not chase CEG yet; treat it as a watch item until power-demand forecasts are revised, because deferred data-center buildouts would be a slower-moving but meaningful second-order headwind.