H100 and H200 GPU Depreciation in 2026: When Buying Used Hardware Is Smarter Than Renting

The H100 SXM 80 GB launched at approximately $30,000-40,000 per GPU in late 2023. By mid-2024, secondary market prices had settled at $20,000-25,000. By early 2025, with H200 becoming broadly available and Blackwell on the horizon, prices dropped to $12,000-18,000. As of mid-2026, a used H100 SXM 80 GB trades at $6,000-15,000 depending on condition, warranty status, and volume. This represents a 65-80% decline in 18-24 months, a depreciation rate that rivals luxury automobiles.

The depreciation curve is not linear. The steepest decline (40-50%) occurred between Q4 2025 and Q2 2026, driven by three factors: B200 general availability, the simultaneous release of H200 at similar performance levels, and a secondary market flooded with GPUs from bankrupt AI startups and over-provisioned hyperscaler capacity. The annual depreciation rate for data center GPUs has historically been 30-50%, but the H100 cycle compressed this into 18 months due to the accelerated pace of NVIDIA's architecture releases.

The 2026 secondary GPU market is stratified by model and condition. H100 SXM 80 GB with active warranty: $10,000-15,000. H100 SXM 80 GB without warranty: $6,000-10,000. H100 PCIe 80 GB: $5,000-8,000 (lower demand due to no NVLink). H200 SXM 141 GB: $18,000-25,000 (limited secondary supply). H200 PCIe 141 GB: $12,000-18,000. A100 80 GB (now a legacy product): $3,000-6,000. Prices vary by volume: 100+ GPU lots command 15-25% discounts over single-unit pricing.

The buy-versus-rent decision depends on utilization rate, time horizon, and cost of capital. At current prices, buying a used H100 at $10,000 and using it for 24 months before selling at an estimated $3,000-5,000 residual ($3,000-4,000/year depreciation) compares favorably to renting at $3.50/hr ($30,660/year at 100% utilization). The break-even utilization for buying vs renting is approximately 40-50%. Above 50% utilization, buying wins. Below 40%, renting wins.

GPU ownership carries four categories of risk that the rent vs buy analysis must account for. Technology obsolescence risk: B200 offers 3-5x the performance of H100, meaning H100's competitive lifespan may be only 2-3 more years. Demand risk: if AI inference shifts to specialized ASICs or if model efficiency improvements reduce GPU demand, resale values could decline faster than projected. Operational risk: GPU failures, cooling requirements, and power costs are owner responsibilities.

Counterparty risk: secondary market buyers may default on agreements. Quantifying these risks: the technology obsolescence risk is the most significant, potentially reducing H100 value to $1,000-3,000 by 2028 (85-95% total depreciation from new). Demand risk is moderate; the AI GPU market continues growing at 40-60% annually, supporting demand for previous-generation hardware for training and inference. Operational risk is manageable for teams with data center experience.

The probability-weighted risk-adjusted return on buying used H100s at current prices is positive for 2-year time horizons but marginal for 3+ year horizons.

Smart GPU buyers plan their exit strategy before purchase. The most common exit paths are: resale on the secondary market (via GPUlist, Deltabyte, or direct marketplace), lease-back to cloud providers (CoreWeave and TensorDock offer GPU hosting programs where they manage and rent your hardware for a 60-70% revenue share), trade-in programs (NVIDIA and Dell offer trade-in credits of $2,000-5,000 per H100 toward B200 purchases), and internal redeployment to lower-priority workloads as the hardware ages.

The resale market for H100s is evolving rapidly. As of 2026, specialized GPU brokerages (GPUlist.io, Cloudomax) offer liquidity comparable to equipment leasing markets. A 128-GPU H100 cluster purchased used at $8,000/GPU ($1.024M total) can be placed with a lease-back operator at an estimated 4-6% monthly return ($40,000-60,000/month before management fees). This strategy converts the GPU purchase into a yield-generating asset, offsetting the depreciation risk.

The lease-back approach works best for H100s in enterprise-grade configurations (NVLink-connected, certified servers) and less well for PCIe cards or non-standard configurations.

The decision framework centers on three questions. First: what is your utilization rate? Below 40%, rent. Between 40-70%, buy used with a 2-year exit plan. Above 70%, buy new or buy used with a 3-year hold. Second: do you have operational capability to manage hardware? If you do not have data center experience or a colocation relationship, the operational risk premium makes renting more attractive even at 50-60% utilization.

Third: what is your time horizon? For teams that decide to buy: buy used H100s with active warranties from reputable sellers (verified through GPUlist or Dell Financial Services), colocate at facilities with power price locks (preferably renewable-heavy locations like Virginia, Oregon, or Nordic regions), and plan the exit at purchase time by registering with 2-3 secondary market brokers. The optimal buy window for H100 appears to be mid-2026, with prices at or near bottom as B200 production ramps and the market absorbs the startup liquidation wave.

H100s bought at $6,000-8,000 and held for 2 years with 60-70% utilization are projected to deliver 30-50% total cost savings versus equivalent cloud rental.