NVIDIA's $500B AI compute deal draws skepticism from investors

NVIDIA's $500 billion compute financing deal represents a pivotal moment in how Wall Street is treating AI infrastructure. The agreement, announced with major financial firms, has prompted a fundamental shift: compute capacity itself is being securitized and packaged as a tradable asset. This mirrors the evolution of commodities like crude oil or power spreads, allowing institutional investors—pension funds, retirement accounts, and hedge funds—to hold exposure to GPU-driven infrastructure without owning the hardware directly.

However, the structure depends critically on a single assumption: that AI labs will continue to demand ever-increasing compute resources to train progressively better models. Mills highlighted a contrarian view circulating among sophisticated investors: the risk of "diminishing model returns," where existing frontier models may already be powerful enough that marginal improvements require less compute than the market is currently pricing in. Google's recent pullback on AI capex spending and shift to free cash flow positive operations suggests that even the largest hyperscalers are reconsidering their compute investment assumptions. If this trend spreads, the collateral underpinning the securitized compute market—and the securities issued against it—could face significant pressure.

The tension between the bull and bear cases will likely determine whether compute becomes a durable asset class or a sophisticated financing structure built on unproven long-term demand. Mills noted that some investors already believe the $500 billion package is "not even enough" to meet next year's projected ten-gigawatt need, suggesting the market is pricing in sustained growth. Whether that growth materializes depends on whether the AI industry's capex cycles are driven by genuine demand for better models or by financial engineering sustaining an expansion whose underlying returns remain uncertain.