🎙 The Quote

“The analogy I often make is if you could buy a used car… One was driven in a 72 hour nonstop race across the country… the other one was driven to church on Sunday… I want nothing to do with the one that was raced in some kind of bubblegum rally across the country. So in the context of GPUs… the failure rates of GPUs used so intensively for training purposes are much higher than inference specific usage.”

— @paulkedrosky with @kantrowitz / Big Technology Podcast — AI Chips Are Losing Value Faster Than Almost Anything Else

The Core Insight

Kedrosky’s argument is simple and devastating: a GPU doing heavy training runs is like a car in a nonstop cross-country race. The hardware clock doesn’t just tick — it sprints. Failure rates, he argues, are materially higher for training workloads than for inference.

The Structural Read · FourWeekMBA Analysis

In the Map of AI framework, GPUs sit at Layer 1 — raw compute infrastructure. That layer is typically treated as a commodity input. Kedrosky’s point complicates that assumption: the provenance of a chip’s usage history now matters to its residual value.

If training-worn chips depreciate faster and fail more, then the secondary market for AI hardware is not one market — it’s two. A chip that ran inference workloads is structurally a different asset class than one that absorbed frontier training runs. That distinction has real implications for how data centers are valued, how leases are priced, and how AI cloud costs are modeled over time.

Why It Matters Now

The industry is scaling training runs at a pace that stress-tests hardware in ways previous compute eras never did. As more capital floods into AI infrastructure, buyers and operators need to ask a question that didn’t exist five years ago: what did this chip actually do before I got it?

The used-car analogy is memorable precisely because it reframes an abstract hardware question as a tactile, human one. Nobody buys the race car without a discount. The GPU market may need to learn the same lesson.

The Bottom Line

GPU depreciation isn’t just about Moore’s Law making last year’s chip obsolete — it’s about what that chip was put through. Training-heavy workloads may be quietly creating a two-tier secondary market for AI hardware, and the pricing models haven’t caught up yet.