Used AI GPU Prices and Useful Life: A100 to H100

Secondary-market prices for used A100, H100, and V100 GPUs reflect rapid generational turnover, with A100s trading at deep discounts (~30-50% of original MSRP) as hyperscalers decommission clusters for newer hardware, while H100s retain stronger residual value (~40-60%) amid ongoing demand for inference and mid-tier workloads.[1][1]

• A100 (Ampere, 2020): 40GB SXM4/PCIe used prices cluster around $3,500–$6,500 (eBay lows ~$2,350–$6,500; indicative Q3 2026 ranges $4,000–$6,000). 80GB variants trade at $5,000–$9,500 (some listings up to ~$13,000). A full 9,600-GPU hyperscaler cluster listed on eBay equated to ~$6,219 blended per GPU (including infrastructure).[2][3][4]
• H100 (Hopper, 2022): SXM 80GB used ~$14,250–$27,000 (Q2 2026 data showed $15k–$19k realized; later ranges $22k–$27k indicative). PCIe variants lower at ~$10k–$16,500. Residual value ~42–58% after 24–30 months vs. original ~$28k–$33k list.[1][5][6]
• V100 (Volta, 2017): 32GB used ~$700–$910 (eBay sold median ~$860; indicative liquidation $1k–$2k). Far below original ~$10k–$11.5k MSRP, reflecting end-of-life status for most commercial AI.[7][8]

These prices signal a vibrant but tiered resale market: A100s flood secondary channels from hyperscaler refreshes (e.g., 9,600-unit eBay lot from a hyperscaler environment), H100s command premiums due to versatility, and V100s approach liquidation.[9][4] For entrants or competitors, buying used A100/H100 offers capital-efficient entry for inference or fine-tuning but requires factoring in higher power draw, potential reliability risks on older silicon, and competition from cloud rentals (A100 ~$0.34–$1.75/GPU-hr; H100 higher).[10]

Nvidia maintains an annual architecture cadence, with Hopper (H100/H200, ~2022) giving way to Blackwell (B200/GB200/GB300, ramping 2024–2025/early 2026) and then Vera Rubin (Rubin GPUs + Vera CPUs, first production shipments H2 2026, NVL72 systems generally available by fall/November 2026).[11][12][13]

• Hopper introduced FP8 Transformer Engine and strong NVLink scaling.
• Blackwell added native FP4, higher bandwidth (HBM3e up to 8 TB/s per GPU), and rack-scale NVL72 designs for massive efficiency gains.
• Rubin (shipping/ramping as of Nov 2026) features HBM4 (288 GB/GPU, ~19.2–22 TB/s bandwidth after spec refinements), NVLink 6, and co-designed Vera CPU; some bandwidth specs were quietly adjusted downward (~13% HBM4, ~17% NVLink per GPU) to reflect volume-production realities, but overall platform targets remain aggressive.[11][14][15]
• Next steps include Rubin Ultra (~2027) and Feynman (~2028).

This ~12-month cadence compresses product lifecycles and accelerates secondary-market supply as operators upgrade for perf/watt advantages. New entrants face a moving target: designing around one generation risks quick obsolescence unless focused on software optimization or niche workloads.

Performance-per-watt and cost-per-token gains compound dramatically across generations, driven by precision formats (FP8 → FP4), memory bandwidth, interconnects (NVLink scaling), and full-stack co-design (rack-scale NVL systems, software like TRT-LLM).[16][17]

• A100 to H100: ~1.8x FLOPS/W (FP16/BF16 tensor; e.g., ~0.78 to ~1.41 TFLOPS/W system-level in some comparisons); FP8 support enables further effective gains. Cost-per-token for inference drops substantially (H100 ~$0.27–$0.28/MTok on 70B FP8 vs. higher on A100 equivalents).[18][19]
• Hopper to Blackwell (B200/GB300): Up to 25x perf/watt on certain MoE models (SemiAnalysis InferenceX); 10–50x throughput per MW; 15–40%+ lower cost-per-token (e.g., B200 ~2.5x throughput at ~1.5x price → ~40% cost/token cut; some benchmarks show 35x lower token cost vs. Hopper). FP4 enables major inference efficiency.[16][17][20]
• Blackwell to Rubin: Up to 10x higher inference throughput per MW and ~10x lower token cost vs. Blackwell on select workloads (e.g., agentic models); benchmarks show 9–21x+ tokens per dollar vs. GB300 at matched speeds, with some previews citing higher multiples (up to 30x in Nvidia commentary).[21][22][23]

These gains (often 1.6–2.5x+ effective per generation, amplified by software) imply economic lives of 2–4 years for flagship use before power/performance economics favor replacement, though inference extends viability. Power-constrained data centers amplify this: newer gens deliver more tokens per watt (and thus per MW of grid allocation), but higher per-GPU TDP (A100 400W → H100 700W → B200 1,000W+) means fewer GPUs per rack/MW, offset by vastly higher output.[24]

Implication: Competitors or buyers must model TCO including power (PUE effects), utilization, and residual value. Used older GPUs win on capex for low-intensity inference; new gens dominate high-volume or frontier workloads. Long-term, power availability (not just silicon) becomes the binding constraint.

Hyperscalers and analysts report accelerating retirement of A100 clusters (e.g., large decommissioned supercomputers entering secondary markets) as Blackwell ramps, with H100s shifting to inference or secondary uses rather than immediate full retirement; older chips like V100 linger in research/legacy roles.[25][26][27]

• A100 fleets are being displaced for training but repurposed for inference/fine-tuning; CoreWeave extended A100 contracts into 2029, showing retained earning power.[25]
• H100 supply remains tight in secondary markets due to strong productivity (rental rates above depreciation assumptions per a16z analysis); some owners redeploy internally rather than sell.[26]
• Analysts note GPU useful life may exceed traditional 3–6 year depreciation due to inference demand; resale volumes rising (e.g., 980k units projected annualized Q2 2026).[1][25]
• V100s approach end-of-life liquidation for commercial AI but retain niche value.

For market entrants: Plan for multi-generation fleets—buy used A100/H100 for cost-sensitive inference while reserving capex for Blackwell/Rubin on power-dense, high-margin workloads. Monitor hyperscaler refresh cycles for secondary supply gluts. Software optimization (quantization, frameworks) extends older hardware economics more than raw specs suggest. Power and cooling infrastructure decisions made today will dictate upgrade flexibility for years.

---

Recent Findings Supplement (November 2026)

Secondary-market prices for used A100, H100, and V100 GPUs have stabilized or declined modestly in late 2026, with A100 80GB cards trading around $12,500 and H100 SXM around $14,250 realized, reflecting sustained demand for inference workloads amid Blackwell/Rubin ramps.[1]

• As of November 4, 2026, used A100 80GB cards list from $12,500 on eBay (stable since September), with new/refurbished higher at ~$22,000; weekly tracking shows consistency through October.[1]
• October 29, 2026 dealer data shows realized H100 SXM 80GB sales at $14,250 (dealer asks ~$16,576), H100 PCIe $10k–$14k, and H200 SXM modeled ~$22,500 (thin market); A100 values appear in broader $4,800–$18,900 ranges across sources, with some liquidation values lower (~$5,400 fair market for SXM).[2]
• V100 32GB used prices reached ~$910 on eBay as of November 4, 2026 (median asks ~$1,099), with fluctuations but availability in the secondary channel.[3]
• Broader October–November 2026 indices (e.g., American Compute Rack Report through August data) show used H100 retention at 75–85% after 18–24 months and A100 steeper declines to 30–50% of original list, with resale indices down to 75 (H1 2025 baseline = 100) in August 2026.[4]

This pricing dynamic means secondary buyers can acquire capable inference hardware at 40–60% of new equivalents, pressuring new-gen premiums while extending fleet utility.

Nvidia maintained its annual cadence with Blackwell (Hopper successor) ramping in 2025 and Vera Rubin entering production shipments and NVL72 general availability by late 2026, with minor spec revisions but strong hyperscaler uptake.[5]

• Rubin GPU shipping specs (as of October 31, 2026) include 288 GB HBM4 (unchanged), revised down to 19.2 TB/s memory bandwidth and 3.0 TB/s NVLink scale-up (from preliminary 22 TB/s and 3.6 TB/s), with scale-out raised slightly; NVL72 systems now generally available at hyperscalers.[5]
• Jensen Huang highlighted platform revenue per gigawatt rising to ~$40 billion for Vera Rubin (vs. $25 billion for Grace Blackwell and $18 billion for Hopper), tied to productivity X-factors and AI infrastructure spend projections of $3–4 trillion by 2030.[6]
• First NVL72 systems became generally available in fall 2026 on the original H2 2026 schedule, with production hardware arriving and purchase orders from major customers.[7]

Competitors entering the secondary or cloud market must time purchases around these annual waves, as Rubin availability will likely pressure H100/H200 values further in 2027.

Blackwell and Rubin deliver substantial performance-per-watt and cost-per-token gains over Hopper, with Rubin showing up to 30x throughput per megawatt and 45x lower token costs versus GB300 NVL72 on agentic workloads, driven by higher memory bandwidth, NVFP4 support, and rack-scale optimizations.[8]

• SemiAnalysis benchmarks (updated through October 2026) show Vera Rubin NVL72 achieving ~2.5x FP4 inference and ~2.3x training compute per rack versus GB300 NVL72, with memory bandwidth ~2.4x higher (1,400 TB/s vs. 576 TB/s) and NVLink ~1.7x; power draw rises to 156–227 kW per rack.[9]
• Agentic inference results indicate up to 5.7x throughput per MW on specific models versus GB300, with peaks supporting 272M tokens/s per MW; cost-per-token advantages reach 45x lower on some workloads due to efficiency at scale.[10]
• Blackwell (e.g., B200/GB200) examples show ~10x+ throughput per MW and 15–35x lower cost per million tokens versus Hopper on MoE/inference models, with software optimizations (TensorRT-LLM) adding further 4–5x gains post-launch.[11]

These gains accelerate replacement economics in power-limited environments, where older chips lose relative value despite absolute capability.

Power-constrained data center economics imply shorter effective economic lives for less-efficient generations (potentially 3–5 years for frontier training, longer for inference), as each watt favors chips maximizing tokens per MW; however, real-world retention exceeds traditional 3-year assumptions due to secondary demand and long-tail workloads.[12]

• Revenue-per-GW metrics and benchmarks tie directly to power envelopes: Rubin’s efficiency supports higher output without proportional power increases, while hyperscalers face 30–45% shortfalls in planned 2027 capacity due to grid/electrician constraints.[13]
• Older GPUs retain value when they remain the lowest-cost option for specific jobs; e.g., CoreWeave A100 contracts extend to 2029 (ninth year of the chip), with prior-gen pricing holding or rising amid shortages.[12]
• Depreciation trackers show H100 at 60–75% retained value by year 3 and floors around 25–35% beyond year 5, better than book assumptions but pressured by next-gen efficiency.[14]

Firms competing here should model TCO around actual MW utilization rather than nominal depreciation, prioritizing inference repurposing.

Hyperscalers and operators are actively retiring or rotating older fleets (especially H100 for Blackwell space and air-cooled incompatible racks), with rapid secondary-market absorption, repurposing for inference, and new platforms like Compute Exchange facilitating sales; V100-era retirements confirm multi-year lives.[15]

• 2026 sees the largest concentrated AI GPU retirement wave, with H100 clusters displaced by higher-density Blackwell systems and air-cooled facilities unable to host liquid-cooled successors; neoclouds (CoreWeave, Lambda) actively buy retired hyperscaler inventory.[16]
• Secondary volume has grown sharply (projected ~980k units annualized in Q2 2026 data), with H100 days-to-sell shortening and institutional intermediaries rising; Azure’s prior V100 retirement (pre-2026) and ongoing T4 activity illustrate 7–9 year useful lives in practice.[17]
• July 2026 launch of Compute Exchange secondary marketplace for H100/A100 reflects enterprise/cloud demand for non-frontier hardware; H100 supply remains tight in physical secondary channels despite new-gen ramps.[18]

New entrants or refurbishers can capture value by targeting inference buyers and timing sales ahead of major GA events, but must navigate export controls and compliance on large lots.