SECONDARY GPU MARKET SIZE AND TRADING VOLUME
The secondary market for data center GPUs has grown from an estimated $820 million in 2024 to $4.8 billion annualized in Q2 2026, reflecting the massive installed base of AI GPUs from the 2023-2025 investment cycle now rotating into secondary channels. Trading volume reached 245,000 GPU units in Q1 2026, up from 82,000 in Q1 2025. The primary drivers are threefold: hyperscaler and neocloud capacity refreshes (H100 to B200) releasing generation-old hardware, enterprise bankruptcies and restructurings among overleveraged GPU-backed startups, and the emergence of GPU-focused asset management firms that purchase, refurbish, and re-lease GPU servers.
The market has professionalized significantly. GPU asset lifecycle companies like GPU LeaseTrade, ServerHub Capital, and DigitalFleet now intermediate 42% of secondary transactions, up from 12% in 2024. These firms purchase used GPU clusters at 55-70% of original cost, refurbish them (clean, test, replace fans and thermal paste), and re-lease them to enterprise customers on 12-48 month terms at 18-25% annual ROI. The remaining 58% of trading occurs through direct OEM buybacks (Dell/HP/HPE takeback programs), auction platforms, and peer-to-peer trading among neocloud operators. The typical secondary GPU unit changes hands 2.3 times over its 5-6 year economic useful life, generating 1.4x total lifecycle revenue across ownership cycles.
| Metric | 2024 | 2025 | Q2 2026 (Annualized) | 2027 (Projected) |
|---|---|---|---|---|
| Secondary Market Volume (units) | 140,000 | 380,000 | 980,000 | 1,400,000 |
| Market Value ($B) | $0.82 | $2.40 | $4.80 | $5.60 |
| Avg Transactions per GPU | 1.1 | 1.6 | 2.3 | 2.8 |
| Institutional Intermediary Share | 12% | 28% | 42% | 55% |
| Avg Days to Sell Used H100 | 68 | 42 | 28 | 21 |
GPU DEPRECIATION CURVES AND SALVAGE VALUE ANALYSIS
Data center GPU depreciation has followed a steeper curve than traditional server hardware due to rapid generational improvement. H100 SXM 80 GB, with original list price of $28,000-32,000 in 2023-2024, trades at $15,000-19,000 in Q2 2026 (42% residual value at 24-30 months). A100 80 GB ($15,000-18,000 new in 2022-2023) trades at $5,000-7,000 (35% residual after 36-48 months). By comparison, traditional x86 servers retain 55-65% of value after 24 months, indicating GPU-specific technology risk. The H100 depreciation curve shows an inflection point at month 18: values fell 35% in the first 18 months (to ~$18,500) then an additional 23% between months 18-30, as B200 availability accelerated H100 obsolescence perception.
Salvage value projections for H100 at end-of-life (60 months, 2028-2029) range from $3,000-5,000 (10-16% of original value), based on A100's current trajectory. However, the AI inference market's insatiable appetite for compute means generation-old GPUs remain viable for inference workloads for longer than prior GPU generations. H100 delivers adequate throughput for Llama 3.1 70B inference at 38-42 tok/s even in 2026, and its FP8 Tensor Core support extends its useful life for lightweight inference. We project H100 salvage value at 18-22% in 2028, higher than original estimates, because the inference tail market is larger than anticipated. GPU asset-backed lenders now advance 55-65% LTV on new H100 purchases, down from 75-80% in 2024, reflecting more realistic residual value assumptions.
| GPU Model | Original List Price | Current Used Value (Q2 2026) | Residual % | Projected EOL Value (2029) | Depreciation Rate (Annual) |
|---|---|---|---|---|---|
| H100 SXM 80GB | $28,000-32,000 | $15,000-19,000 | 54-58% | $3,500-5,500 | 22-25% |
| H100 PCIe 80GB | $25,000-29,000 | $13,000-16,500 | 52-57% | $3,000-4,500 | 24-27% |
| H200 SXM 141GB | $34,000-38,000 | $24,000-28,000 | 68-74% | $7,000-10,000 | 18-21% |
| A100 SXM 80GB | $15,000-18,000 | $5,000-7,000 | 33-39% | $1,500-2,500 | 28-32% |
| A100 PCIe 40GB | $12,000-15,000 | $3,500-5,000 | 29-33% | $800-1,500 | 35-38% |
TRADING PLATFORMS AND LIQUIDITY DYNAMICS
GPU secondary trading has congealed around three platform types. The largest is the institutional OTC market ($2.7B annualized), where asset lifecycle firms and large neocloud operators trade bulk lots (100-5,000 GPUs) with standardized inspection and warranty terms. The standard OTC contract includes 14-day inspection, Dell/HP OEM certification, and 12-month warranty at 1-2% of purchase price monthly. The second channel is auction and marketplace platforms ($1.4B annualized), led by ServerMonkey, GPUauctions.com, and eBay Enterprise, where individual units to lots of 50 GPUs trade with 7-28 day settlement. The third is direct OEM certified pre-owned ($0.7B), where Dell Financial Services and HPE Asset Recovery sell certified refurbished GPU servers with full OEM warranty at 15-25% premium over OTC prices.
Liquidity has improved materially: the average days-to-sell for H100 has compressed from 68 days in January 2025 to 28 days in Q2 2026. The bid-ask spread on bulk H100 lots has narrowed from 12% to 6%, approaching the liquidity of established enterprise hardware markets like Cisco switches (3-4% spread). The improvement is driven by the entry of institutional market makers who quote firm bid/ask prices on GPU lots, and the standardization of GPU server configurations across the industry. The H100 has become a commodity: identical in specification across deployments, with standardized firmware and known performance characteristics, enabling price discovery and efficient matching.
REFURBISHED GPU PERFORMANCE AND RELIABILITY DATA
Performance testing on 2,000 refurbished H100 units from three refurbishers shows that properly refurbished H100 GPUs perform within 97-99% of new-unit specifications on sustained training workloads. The primary failure mode is HBM3 thermal degradation: after 18-24 months of continuous 700W operation, HBM3 temperature margins degrade by 4-7 degrees Celsius, increasing ECC error rates by 0.3-0.8 errors per GB-hour versus new units. Refurbishers address this with thermal pad replacement, heatsink cleaning, and firmware voltage optimization that restores temperature margins to within 2-3 degrees of new. Fan replacement is required on 22% of H100 units at the 24-month mark, costing $45-85 per fan in refurbishment.
The failure rate data: refurbished H100 units experience 1.8% DOA (dead on arrival) rate versus 0.3% for new, and 4.2% annual failure rate (AFR) in year 1 versus 1.8% for new. By year 3, refurbished AFR converges with new units at 5-7%. The economic tradeoff is favorable: refurbished H100 at $17,000 average versus new at $30,000 (43% savings) with 2.3x higher DOA risk and 3-4x higher year-1 warranty claims. For inference workloads with redundancy, the refurbishment economics strongly favor secondary GPU acquisition. For training clusters (tight coupling, job-level fault tolerance), new GPU acquisition is recommended despite the premium.