AI Infrastructure Capex Boom Faces Monetization Crisis

The AI Infrastructure Spending thesis is substantially supported by empirical capex data, but valuations across semiconductor and REIT beneficiaries have already significantly priced in robust 2025-2027 spending. The 5-year outlook is bullish on absolute capex volumes and revenue, but bearish on incremental valuation upside for most direct beneficiaries. The critical inflection point occurs around 2027-2028, when the market must transition from capex-driven growth to profit-driven validation. Goldman Sachs now forecasts 2026 capital spending of $527 billion, up from consensus estimates that were too low for two consecutive years (roughly 20% implied growth vs. actual 50%+). Goldman projects cumulative capex of $1.15 trillion from 2025-2027, more than double the $477 billion spent 2022-2024. However, maintaining historical returns on capital would require $500 billion average annual capex to generate over $1 trillion in annual profits—more than double 2026 consensus estimates of $450 billion. On the capex side, the thesis remains intact through 2027. Hyperscaler capex for the "big five" is forecast to exceed $600 billion in 2026, a 36% increase over 2025, with roughly 75% ($450 billion) directly tied to AI infrastructure. Google, Amazon, Microsoft, and Meta invested nearly $200 billion of the $290 billion total data center capex in 2024, with growth expected to exceed 40% in 2025. Beyond hyperscalers, data center infrastructure market segments are expected to sustain double-digit growth through 2030, reaching $1 trillion by 2030. However, critical headwinds emerge: (1) **Valuation Risk**: NVIDIA trades at a P/E of 44.3x with mixed short-term sentiment after a 20% drawdown from all-time highs amid valuation concerns. NVIDIA's forward P/E exceeds 27x on 2027 earnings estimates, implying assumption of exponential growth. (2) **Capex to Profit Gap**: Hyperscalers increasingly rely on debt markets, transforming historically cash-funded models into leveraged businesses with capex exceeding free cash flow. Companies borrowed $75 billion in recent months for AI data centers—more than double annual average issuance over the past decade. (3) **Operating Earnings Pressure**: Investors have rotated away from AI infrastructure companies where operating earnings growth is under pressure and capex is debt-funded. (4) **Structural Constraints**: U.S. data centers face a capacity shortfall exceeding 11 GW, with cumulative gap expected to exceed 40 GW by 2028. From May 2024 to March 2025, more than $64 billion in data center projects were delayed or blocked due to environmental and energy concerns. Direct beneficiary valuations are stretched. Equinix trades at 21x forward AFFO, while data centers trade at 23x AFFO vs. 21x for broader REIT universe, with analyst enthusiasm tempered by valuation concerns. AMD and NVIDIA both face competitive and execution headwinds: AMD outperformed NVIDIA in 2025 (82% vs. 34%), though NVIDIA maintains 963% three-year gains vs. AMD's 183%. Goldman Sachs expects the next phases of AI trade to involve AI platform stocks and productivity beneficiaries, not infrastructure companies. The semiconductor supercycle is credible. Global semiconductor industry expected to reach $975 billion in 2026, up 26%, with annual sales reaching $2 trillion likely by 2036. TSMC's CoWoS capacity expected to expand 60%+ from end-2025 to end-2026. Yet financial experts warn of significant gap between infrastructure spending and actual software revenue, with bubble risk peaking around 2027 when equipment sales hit record levels; failure of Agentic AI to deliver productivity gains could trigger sharp contraction similar to dot-com.