Aravolta: GPU Depreciation Varies 30-45% Based on Actual Workload, Making Industry Averages Obsolete for Lenders
Analysis published November 13, 2025
The GPU financing market is pricing risk off fantasy numbers. According to a detailed analysis from Aravolta, telemetry data from mid-market GPU deployments reveals that identical hardware can age at wildly different rates depending on actual usage patterns, with depreciation curves varying by 30-45% across customers despite using the same GPU models. This gap between underwriting assumptions and reality is creating significant exposure for lenders who continue to use uniform depreciation schedules.
The Industry's Broken Consensus on GPU Lifespan
The current market operates on conflicting assumptions that reveal a fundamental lack of consensus. CoreWeave publicly underwrites GPU assets assuming roughly 6 years of useful life, while Nebius uses closer to 4 years. Some analysts warn that heavy usage could push the real number down to 3 years or less. The problem is that everyone is arguing over an average that does not exist in practice. As Aravolta's analysis demonstrates, the question should not be how long an H100 lasts in general, but rather how any specific GPU was actually operated over its lifetime.
The current financing model treats GPUs like trucks or drilling equipment with predictable straight-line depreciation. This approach fails catastrophically when applied to compute infrastructure because identical GPU hardware ages very differently depending on workload intensity, thermal management, and utilization patterns. A GPU running steady inference workloads at 60-70% utilization under moderate thermals can remain economically viable for 5-plus years, while an identical unit subjected to irregular training workloads with daily spikes to 95-100% utilization may be economically obsolete in 3 years or less.
Real-World Data Exposes a Two-Year Gap in Asset Life Expectations
Aravolta's work with a mid-market lender financing GPU deployments in the 0-50MW range revealed the scale of the problem. The original underwriting assumed 80% steady utilization, 5.5 years of useful economic life, and no meaningful variance across customers or workload types. The actual telemetry told a dramatically different story.
Workload intensity spikes to 95-100% were not occasional as assumed but happening daily, causing accelerated component stress. Thermal envelope violations that were not modeled at all were occurring frequently during training bursts, with many GPUs regularly running above recommended thermal limits. Maintenance cycles showed elevated needs under high-variance workloads, with fans, thermal paste, and other components requiring attention sooner than expected. Most critically, economic obsolescence arrived 18-30 months earlier for some workloads compared to others, meaning the GPUs became uneconomical long before they physically failed.
The result was striking. One cohort of GPUs expected to deliver 5.5 years of useful life was trending toward just 3.7 years under actual deployment conditions. This nearly two-year gap between expectation and reality has cascading effects on salvage value assumptions, loan covenant triggers, and the entire debt-versus-equity calculus. For lenders, it represents the difference between a performing loan and an under-collateralized exposure that materializes without warning.
The Physics of GPU Degradation
The analysis identifies thermal stress as the primary driver of accelerated aging. For every 10 degrees Celsius increase in operating temperature, electronic component life is roughly cut in half. Sustained high utilization at 98% load around the clock accelerates wear of components, with industry data indicating that even at 60-70% average utilization, top data-center GPUs may only last 1-3 years. Power spikes strain voltage regulators and capacitors, with high-end GPUs surging past 500 watts and causing electromigration damage over time.
Thermal cycling from bursty workloads creates additional stress that steady 24/7 operation does not. Repeated heating and cooling causes materials to expand and contract, stressing solder joints in ways that constant-temperature operation avoids. Inadequate maintenance compounds these issues, with GPU fans rated for roughly 5 years but failing sooner under constant high RPM operation. The analysis recommends cleaning, fan swaps, and thermal paste renewal twice yearly as essential maintenance.
What This Means for GPU Financing
Lenders financing GPUs without telemetry-based asset-level monitoring face multiple risks that traditional underwriting cannot capture. They risk losing principal without warning when borrowers overrun their GPUs with constant maximum utilization and high thermals, turning a fleet underwritten for 5-6 years into one with only 2-3 years of economic runway. They overprice deals for good operators who manage their GPUs conservatively, driving away quality borrowers to cheaper capital. Simultaneously, they underprice deals where GPUs are being run to failure, taking on hidden risk that surfaces only when write-downs become necessary.
The lack of granular data also forces lenders into conservative, inflexible terms when creative structures like sale-leasebacks with performance triggers or usage-based financing could better match risk and return. Without actual utilization data, these structures remain theoretical.
Aravolta argues that the next generation of compute financing requires real-time monitoring of GPU fleets at the hardware signal level to produce depreciation curves tailored to each workload and environment. The company is working with mid-market operators in the 1-100MW segment to support structures like sale-leasebacks, usage-based leases, and revolving GPU credit lines anchored by actual performance data rather than industry folklore.
The old world of GPU financing runs on averages. The new world, according to Aravolta, runs on telemetry. For an asset class where identical hardware can depreciate at rates varying by 30-45% depending on how it is actually operated, that distinction is not academic. It is the difference between pricing risk correctly and flying blind.