The H100 Depreciation Curve
An H100 SXM purchased at MSRP of $30,000 in Q3 2023 has an estimated resale value of $15,000 to $18,000 in mid-2026, depending on usage hours and warranty status. This represents approximately 45% depreciation over three years, or roughly $4,500 to $5,000 per year. The curve is not linear: the steepest drop happened between Q2 2024 and Q2 2025 when B200 shipments began, then flattened as the secondary market matured.
PCIe variants depreciate faster, trading at $11,000 to $13,000 on the secondary market, due to lower demand from clusters that prioritize NVLink bandwidth. A used H100 PCIe purchased today at $12,000 would depreciate an estimated $3,000 to $4,000 per year through 2028, assuming Blackwell Ultra and Rubin continue to push down Hopper pricing.
| Purchase Quarter | New MSRP | Current Value (2026) | Annual Depreciation |
|---|---|---|---|
| Q3 2023 | $30,000 | $15,000-$18,000 | ~$4,500/yr |
| Q1 2024 | $28,000 | $16,000-$19,000 | ~$4,000/yr |
| Q3 2024 | $26,000 | $18,000-$20,000 | ~$3,500/yr |
| Q1 2025 | $25,000 | $20,000-$22,000 | ~$2,500/yr |
| Current secondary | $15,000-$18,000 | $12,000-$15,000 | ~$3,000/yr |
The Rental Baseline: What Cloud Costs Really Look Like
H100 on-demand cloud pricing across CoreWeave, Lambda, and RunPod ranges from $1.49/hr to $2.65/hr depending on commitment term and provider. At the blended market rate of $2.00/hr for a 3-month reserved instance, a team running a single H100 for 8,760 hours per year (24/7) pays $17,520 annually in compute costs.
That $17,520 annual rental cost does not include storage, networking, or data egress. A typical training workload with 10 TB of checkpoints and dataset storage on NVMe-backed volumes adds $300 to $600 per month. Adding 10 TB/month in egress at $0.05/GB brings the total to roughly $23,000 per year for a fully-loaded single-GPU cloud setup.
The Utilization Break-Even Calculation
Buying a used H100 at $16,000 with a 3-year useful life gives an effective cost of $5,333 per year in depreciation alone. Adding power ($1,200/yr at $0.12/kWh for 700W TDP), facility colocation ($3,000/yr for half-rack), and maintenance ($500/yr) brings the annualized cost to approximately $10,000 per year.
The break-even utilization rate is the point where annual rental cost equals annual ownership cost. At $10,000 per year on-premise versus $23,000 per year fully-loaded cloud, the break-even occurs at roughly 43% utilization. Any workload running a GPU more than 43% of the time (3,770 hours per year) is cheaper on owned hardware.
Cluster-Scale Economics
An 8-GPU H100 server on the secondary market costs approximately $70,000 to $85,000 for an SXM system with NVLink, dual AMD EPYC or Intel Xeon CPUs, and 2 TB of system RAM. At that price, the annual cost per GPU in the cluster drops to roughly $3,500 per year including power and colocation, half the single-GPU economics.
At an 8-GPU scale with shared infrastructure (one rack, one power feed, one network switch), the per-GPU colocation cost drops by approximately 60%. The break-even utilization for an 8-GPU cluster falls to roughly 30%, making ownership compelling even for moderately utilized clusters.
Risk Factors That Shift the Math
The biggest risk for buyers is architectural obsolescence. The H100 lacks FP4 support, has less than half the HBM capacity of B300, and uses NVLink 4 instead of NVLink 5. If your training workloads shift to models that require FP4 or >141 GB VRAM per GPU, the H100 becomes unusable regardless of how cheap it was.
Warranty and reliability are the second risk. Used H100s typically ship with 6-12 months of remaining manufacturer warranty. Out-of-warranty repairs cost $3,000 to $6,000 for a replacement unit. A 5% annual failure rate adds approximately $250 per GPU per year in expected replacement costs.
Liquidity risk matters if your compute needs change. The H100 secondary market has grown deeper in 2026 but still carries a 4-8 week sell time at fair market value. A sudden shift in model architecture or workload type could strand your hardware investment.
Decision Framework
Buy used H100s if your workload exceeds 45% GPU utilization on a sustained basis, your model fits within 141 GB of HBM per GPU, and you have a 12+ month horizon for the specific workload. The math favors ownership most strongly for batch inference and fine-tuning jobs that run continuously.
Rent if your utilization is below 30%, your model sizes are growing and may exceed H100 HBM capacity, or you need the flexibility to switch GPU architectures within weeks. Renting also wins for pre-production experimentation where GPU type and cluster size change frequently.
Use a hybrid approach for the middle band: reserve 12-month cloud contracts for your baseline load and use spot instances for burst. ClusterBid can source both reserved and spot across multiple providers, giving you the blended rate without buying hardware.