AI Compute as Collateral: The $500B Bet on Depreciating Assets

The Infrastructure Behind the AI Boom

AI compute has evolved from a verb describing calculation into a specific, financeable asset class now attracting unprecedented capital. In August 2026, Nvidia and six major Wall Street firms signed memorandums to mobilize more than $500 billion in financing for AI infrastructure, subject to final execution. According to reporting by Robert J. Szczerba in Forbes, this represents a fundamental shift in how the technology industry views and finances computing capacity.

The term "AI compute" encompasses far more than the graphics processing units that typically dominate headlines. The full stack includes GPUs, high-speed memory, networking equipment, power systems, cooling infrastructure, and critical software layers like Nvidia's CUDA toolkit. Semiconductors and related components account for roughly two-thirds of total AI data center costs, with the remainder split between buildings, power distribution, and cooling systems that can last decades.

This mismatch in asset lifespans creates the central tension in AI compute financing. Long-lived infrastructure wraps around rapidly depreciating chips, and the question of residual value determines whether the economics work.

Why it matters

The AI infrastructure buildout represents one of the largest capital deployments in technology history, but it's being financed using assumptions about hardware longevity that have no historical precedent. Unlike toll roads or commercial aircraft—assets with decades of secondary market data—AI compute clusters have no established resale benchmarks. A secondary market for used AI hardware only opened in July 2026, and Bernie Margulies, who sells insurance against hardware depreciation, told trade press that estimates of residual value range from 10% to 60% of original cost. That fifty-point spread represents the difference between a performing loan and a total loss.

The Utilization Problem

Measured utilization rates tell a troubling story. Cast AI's telemetry across tens of thousands of enterprise GPU clusters found average utilization at just 5% of provisioned capacity across a full day. When companies estimate their own usage, 53% report figures between 51% and 70%. The gap between instrumented reality and self-reported estimates suggests many organizations are financing hardware they barely use while it depreciates on their balance sheets.

One 2026 model indicates that owning compute infrastructure only becomes cost-effective above roughly 70% sustained utilization, while renting wins below 30%. For workloads with variable demand, API-based token pricing eliminates depreciation risk entirely, though it becomes expensive at scale.

Limited Backstop from Nvidia

In announcing the financing platforms, Nvidia CEO Jensen Huang stated the company "may provide a residual-value support mechanism for up to 25% of an opportunity, assessed carefully on a project-by-project basis," according to coverage by TechCrunch and Bloomberg. That language—"may," "up to," "assessed...project-by-project"—reveals how much risk the chip designer is unwilling to absorb. The company that controls GPU architecture roadmaps and determines when chips become obsolete has capped its own exposure at a quarter of project value, evaluated case by case.

Amazon's experience offers a cautionary precedent. When the company reduced the assumed useful life of some servers and networking equipment from six years to five, it added $1.4 billion to a single year's depreciation expense.

Two Workloads, Different Economics

AI compute serves two distinct functions with different depreciation profiles. Training builds models in bounded, expensive runs that demand peak performance. Inference runs trained models continuously, producing answers at scale with less demanding requirements per query. Chips that age out of training workloads often migrate to inference or batch processing, extending their useful life. CoreWeave told investors in August 2026 it had signed contracts for A100 capacity—a 2020 architecture—running through 2029.

The question facing lenders and buyers is whether that migration pattern provides enough residual value to support half a trillion dollars in financing. Margulies, drawing on mainframe market history, warns that "technology obsolescence is sudden, not gradual," and that when mainframe values collapsed, lessors "went bankrupt within months."

These details were first reported by Robert J. Szczerba in Forbes.