Building a Residual-Value Curve for an Accelerator Fleet

Building a residual-value curve for a fleet of GPU accelerators means answering one question with real precision: what will this A100 or H100 be worth in eighteen months, and how do you know? Most finance teams borrow assumptions from server hardware, or worse, from vehicle leasing models, and both are wrong in ways that compound fast at scale. The right approach pulls from three sources: on-fleet utilization logs, secondary-market transaction data, and replacement-cycle assumptions tied to the vendor roadmap. Get the weighting wrong and a lessor either overprices residuals and eats the loss at term-end, or underprices them and loses deals to a competitor running a sharper curve. ## Why accelerator depreciation doesn't behave like anything else in the data center Start with the obvious complication. Compute accelerators depreciate on two axes at once, physical wear and architectural obsolescence, and the second one usually wins by a wide margin. A server chassis or a storage array degrades in a fairly linear, mechanical way; you can model its residual value off hours-run and failure rates and feel decent about it. GPUs age along a different curve entirely. An H100 sitting idle in a rack for three years still loses most of its economic value, because Nvidia's release cadence resets the market's willingness to pay for the prior generation almost overnight. Call it the Jevons paradox playing out at rack scale. Compute gets cheaper and more available per dollar every generation, so demand for the older stuff doesn't taper off gracefully. It falls off a shelf the moment the new chip ships in volume. I watched this happen with V100 capacity in the months after A100 general availability: pricing power evaporated within a couple of quarters, not years. A residual curve that treats a GPU like a delivery van, depreciating at a steady percentage annually, will misprice that cliff every time it matters. ## Utilization logs: the input everyone has and few use correctly Every operator running a fleet already has utilization telemetry, usually from DCGM, or from a scheduler like Slurm or Kubernetes with GPU device plugins turned on. The mistake is treating utilization percentage as a stand-in for wear. Thermal cycling matters more than raw hours logged; a GPU sitting at a stable 70% in a well-cooled rack ages differently than one hammered to 100% and idled repeatedly in a facility with marginal cooling. Pull the logs and look at variance, not the mean. The mean lies to you. The more useful signal buried in that telemetry is demand elasticity. If a fleet's A100s are booked at 90%-plus while newer H100 capacity sits at 60%, the market hasn't finished absorbing the older generation, and the curve should reflect a slower initial decay. Flip that ratio and the curve needs to steepen immediately. Marketplaces like Vast.ai and RunPod, which broker capacity across a pool of providers, publish enough pricing dispersion across generations that a finance team can triangulate this elasticity without building its own multi-year dataset from scratch. ## Secondary-market listings: noisy, thin, and still indispensable The used GPU market is nowhere near as liquid as the used-car market. Anyone building a curve needs to sit with that discomfort rather than paper over it with false precision. eBay listings for data-center parts, broker platforms, contacts inside the GPU trading channels that quietly move a real share of secondary V100 and A100 inventory, OEM refurbished programs: each shows a different price for what is nominally the same part. None of them, alone, is enough to build a curve on. What you can do is triangulate a band instead of a point estimate. Take the lowest and highest observed transaction prices for a given SKU at a given age, throw out the outliers that smell like distressed liquidation or bundled-deal noise, and treat what's left as your confidence interval, not a single tidy number. A finance team that tells its board "this GPU will be worth $X in three years" is overstating its own certainty by a wide margin. A team that reports a range and explains why the range is that wide is being honest about a market that simply doesn't have the transaction volume of, say, used freight trucks, where actuarial-grade residual tables have existed for decades. ## Replacement-cycle assumptions: modeling around the vendor roadmap, not against it Most models fail hardest here, because they treat Nvidia's product cadence as an exogenous shock rather than the single most predictable input in the whole exercise. Nvidia has said plainly, through its own investor calls, that it intends to ship new architectures faster than the old two-year gap, closer to an annual rhythm now. AMD's MI-series, and increasingly custom silicon out of Google and Amazon, add pressure on the margin too, even if they don't yet compete head-on for CUDA-locked workloads. Build the curve around known launch windows, not around some generic depreciation percentage borrowed from a different asset class entirely. A GPU bought six months before a new flagship ships should carry a steeper near-term curve than one bought six months after, even though both are identical parts with identical specs on day one. Obvious once you say it out loud. And yet a striking number of leasing agreements still price residuals off flat five-year straight-line schedules, because that's the accounting system's default and nobody bothered to override it. ## Putting the three inputs together without letting any one of them dominate The honest version of this model produces a stepped curve, with sharp discontinuities at known architecture launch dates, weighted toward whichever input has the most reliable data at that point in the asset's life. Early on, utilization logs carry the most weight, since the secondary market for a two-month-old H100 barely exists yet. By month eighteen, secondary comparables start to dominate, because enough transaction volume has piled up for the price band to actually mean something. Replacement-cycle assumptions matter most as a leading indicator and least as a lagging one. Once a new architecture has been in market two quarters, you don't need the roadmap anymore; utilization and secondary data will have already shown the demand shift. The roadmap earns its keep in the six months before launch, when nothing else in the model has caught up yet. Vehicle lessors have a century of actuarial history and a genuinely liquid resale market behind them. GPU lessors have maybe a decade of relevant data and a market still figuring out its own conventions, still arguing over what a "fair" secondary price even looks like. Defensible curves are buildable anyway. The ones that hold up under audit are built from actual telemetry, actual transaction data, and an honest read of Nvidia's own stated roadmap, drawn up fresh rather than inherited from a flat depreciation schedule that happened to be sitting in the finance team's template folder since the server-leasing days.