AI Infrastructure's Timing Trap: Billions Poured Into Hardware That May Obsolesce Before It Pays Off

Tech giants plan $725B in 2026 capex, but AI's rapid evolution means data center hardware may lose commercial value before investors recoup costs. The biggest winners may not be the builders.

Amazon, Microsoft, Alphabet, and Meta plan up to $725 billion in combined capital expenditure for 2026, a staggering bet on AI infrastructure that faces a fundamental timing mismatch: data centers take years to build, but the hardware inside them is depreciating faster than anyone modeled. Amazon already shortened the estimated useful life of some servers and networking equipment from six years to five, a single accounting change that added $1.4 billion in depreciation and cut net income by $1 billion in 2025. That is not a rounding error — it is an admission that the investment thesis is aging in real time.

The core problem is asset-life divergence. As Volker Filsing of Arthur D. Little told Business Insider, investors are funding assets with fundamentally different economic lifespans. Buildings and power connections can serve multiple generations of technology. Computing hardware cannot. The question for investors is how often that hardware must be replaced and how much revenue it can generate before replacement — a calculus that worsens with every efficiency breakthrough in AI models. Goldman Sachs CEO David Solomon has publicly acknowledged that no one knows the final scale of the buildout — whether it is $5 trillion, $10 trillion, or more. The uncertainty is not about whether AI succeeds; Brent Tomlinson of Kroll notes the technology will certainly be widely adopted. The uncertainty is about whether today's infrastructure investments will generate the returns investors expect before becoming yesterday's bets. The 1990s fiber-optic parallel is now being invoked explicitly. State Street Investment Management flagged in an October report that fiber-optic networks became essential to the internet economy, but some firms that financed their construction faced financial ruin. Their analysis suggests AI infrastructure may follow the same pattern: as computing becomes cheaper, the firms that use AI — not the firms that build the pipes — will capture the larger share of economic value created. Market stress is already visible. In October, rising interest rate expectations hit the sector hard, sending shares of CoreWeave and Nebius sharply lower. Higher rates raise the hurdle for capital-intensive projects and compress the margin for error on payback timelines that were already tight. Investors are becoming more discriminating. Tomlinson observes a shift from treating AI as a single investment theme to evaluating each segment individually — infrastructure providers, model developers, application layers. The riskiest position, per Filsing, is highly leveraged hardware purchases dependent on a handful of customers, built on the assumption that current rental prices will hold. That assumption faces pressure from both efficiency gains in AI models and potential oversupply of data center capacity. The paradox of efficiency looms over the entire thesis. More efficient AI models could reduce per-query compute costs while simultaneously expanding total demand — a Jevons paradox scenario. For infrastructure investors, the math depends on whether paid usage growth outpaces price declines enough to cover power costs, financing charges, and accelerated equipment replacement cycles. That is a narrow path, and the toll for getting it wrong is measured in billions.