GPU Resale Value Depreciation: Financial Modeling for AI Teams Planning Hardware Exit Strategies

For AI teams purchasing GPU hardware outright, depreciation is the single largest hidden cost in total cost of ownership. A $30,000 H100 GPU loses approximately 55-65% of its value within 24 months based on secondary market transaction data from 2024-2026. For a 256-GPU cluster, that represents $4.2-$5.0M in unrealized asset loss that does not appear on any cloud invoice but directly impacts balance sheet health.

The buy-versus-lease decision hinges entirely on depreciation assumptions. If a GPU retains 45% of its value after three years, buying beats leasing at most lease rate structures below 2.5% monthly. If depreciation accelerates to 70% loss by year two, leasing becomes the clear winner. The problem is that GPU depreciation is not a smooth exponential curve, it is a step function driven by NVIDIA architecture release cadence.

Hopper (H100) GPUs purchased in Q2 2024 at $28,000-$32,000 each now trade at approximately $18,000-$22,000 on the secondary market as of Q2 2026, representing roughly 33% depreciation over 24 months. This is significantly better than historical GPU depreciation rates, driven by sustained demand for Hopper-class compute even after Blackwell's launch.

Blackwell B200 GPUs tell a different story. Early units purchased at $35,000-$40,000 in Q4 2024 have already depreciated to $25,000-$30,000 by Q2 2026, a roughly 28% loss in 18 months. The B300 announcement accelerated this decline, with secondary B200 prices dropping another 8-10% in the quarter following the B300 launch.

GPU depreciation does not follow a smooth annual curve. It follows a step function tied to NVIDIA architecture announcements. When B300 was announced in March 2026, B200 secondary prices dropped 10-12% within two weeks. When H200 was announced in late 2024, H100 secondary prices fell 15% overnight. The release of a new architecture creates an immediate, discontinuous drop in prior-generation resale values.

This pattern is predictable: the announcement date is the primary depreciation event, accounting for roughly 60% of total value loss over a GPU's three-year lifecycle. The other 40% is gradual erosion from supply increases and demand normalization. Teams that time their hardware exit to occur before a new architecture announcement capture significantly better resale prices.

The buy-versus-lease decision reduces to comparing the monthly lease payment against the sum of monthly depreciation plus cost of capital. At current secondary market depreciation rates, buying an H100 delivers a lower effective monthly cost than leasing at rates above 1.8% of GPU value per month. For B200 and B300, the higher depreciation uncertainty pushes the break-even lease rate higher, to approximately 2.3% per month.

However, this analysis excludes operational considerations. Leasing provides balance sheet flexibility, technology refresh rights, and protection against residual value risk. For venture-funded AI startups with uncertain runway, the insurance premium embedded in lease pricing (typically 0.4-0.6% extra per month) is often worth paying.

Historical data from the H100 lifecycle reveals the optimal sell window: 10-14 months after purchase, before the next architecture announcement but after the initial steep depreciation from unpacking and burn-in has passed. H100 GPUs sold at month 12 captured approximately 80% of purchase price, while those sold at month 18 captured 70%, and month 24 captured 63%.

For B200 owners in 2026, the optimal exit window is Q3 2026, ahead of the expected Rubin architecture announcement in early 2027. Waiting until after Rubin launches could trigger a 12-18% step-function price drop, similar to what happened to H100 when B200 was announced. B300 buyers should plan for a 12-month hold cycle unless NVIDIA extends the architecture generation, which is unlikely given the historical 18-24 month cadence.

The GPU secondary market has matured significantly since 2024. Specialized brokers like ServerHub, GPUlist, and ClusterBid's resale marketplace now provide transparent pricing and verified hardware. Transaction volumes for H100-class GPUs exceeded 50,000 units in Q1 2026 alone, creating sufficient liquidity for teams to exit 256+ GPU clusters without moving the market.

Market liquidity varies by GPU generation. H100 has deep secondary liquidity with typical bid-ask spreads of 4-6%. H200 has moderate liquidity with spreads of 8-12%. B200 and B300 have very thin secondary markets with spreads exceeding 15%, meaning sellers face significant price discovery risk. This liquidity premium should be factored into buy-versus-lease models for newer architectures.

For teams with 12-18 month AI workloads, renting GPU capacity through a marketplace like ClusterBid eliminates depreciation risk entirely and provides technology refresh optionality. The effective cost premium of renting versus buying over 12 months is approximately 15-25%, which is essentially an insurance premium against the step-function depreciation risk of a new architecture announcement.

For teams committed to a 36-month ownership horizon, H100 remains the safest buy due to its deep secondary market liquidity and proven demand profile. B300 is a higher-risk, higher-reward asset: superior compute economics if held for the full lifecycle, but with significant valuation uncertainty in years two and three. We recommend a hybrid approach: buy 60% of required capacity and rent 40%, using the rental portion as a floating hedge against depreciation exposure.