NVIDIA GPU Pricing Trends in 2026: A100 vs H100 vs V100 on the Used Market
Current market prices for NVIDIA A100, H100, and V100 GPUs. Why H100 availability is tight, where to source, and the refurb vs new pricing delta.
This guide is for: AI/ML infrastructure buyers navigating the secondary GPU market for training and inference workloads.
The used GPU market for AI workloads has been volatile for three years running. What started as pandemic-era supply chain issues evolved into AI demand shocks that have kept prices elevated even as new supply has increased. If you're building out AI infrastructure in 2026, understanding the secondary market dynamics is essential—especially for the H100, where new allocation remains tight and used prices have stayed surprisingly firm.
Here's where the market stands in May 2026, with real pricing data and sourcing guidance.
The Current Landscape
Three factors are driving GPU pricing right now:
AI demand continues to outpace supply. Despite NVIDIA ramping H100 and H200 production, demand from hyperscalers, enterprises, and sovereign AI initiatives keeps allocation tight. When new H100s are scarce, buyers turn to the secondary market, keeping used prices elevated.
The V100 is aging out for training. The V100 remains viable for inference and smaller training jobs, but organizations building new clusters are targeting A100 and H100. This has created a two-tier market: premium pricing for H100, firm but softer pricing for A100, and declining but still relevant pricing for V100.
Enterprise refresh cycles are accelerating. Companies that bought A100 clusters in 2021–2022 are now refreshing to H100 or H200. This is putting more A100s into the secondary market than we've seen in previous years—good news for buyers who don't need the absolute latest.
Current Market Prices (May 2026)
These are actual transaction ranges I'm seeing from brokers and auction data:
NVIDIA H100
| Configuration | New Price | Used Price | Availability | |---------------|-----------|------------|--------------| | H100 SXM5 (80GB) | $25,000–$32,000 | $18,000–$24,000 | Very tight | | H100 PCIe (80GB) | $22,000–$28,000 | $16,000–$21,000 | Tight |
The H100 used market is unusual—prices have stayed at 70–80% of new because supply is so constrained. Most used H100s coming to market are from failed startups or distressed sales, not enterprise refreshes. If you find H100s at sub-$16K, verify provenance carefully.
NVIDIA A100
| Configuration | New Price | Used Price | Availability | |---------------|-----------|------------|--------------| | A100 SXM4 (80GB) | $12,000–$15,000 | $6,500–$9,000 | Moderate | | A100 SXM4 (40GB) | $9,000–$11,000 | $4,500–$6,500 | Good | | A100 PCIe (80GB) | $10,000–$13,000 | $5,500–$8,000 | Moderate | | A100 PCIe (40GB) | $7,500–$9,500 | $4,000–$5,500 | Good |
The A100 market has loosened as enterprise refreshes hit the secondary market. The 40GB variants are particularly available—fine for many inference workloads and smaller training jobs. At 50–60% of new pricing, A100s represent solid value for buyers who don't need H100 performance.
NVIDIA V100
| Configuration | New Price | Used Price | Availability | |---------------|-----------|------------|--------------| | V100 SXM3 (32GB) | $5,000–$7,000 | $2,000–$3,500 | Good | | V100 PCIe (32GB) | $4,500–$6,500 | $1,800–$3,000 | Good | | V100 PCIe (16GB) | $3,500–$5,000 | $1,200–$2,200 | Very good |
The V100 is now a budget option for inference and development work. At $2,000–$3,000 for 32GB variants, it's cost-effective for workloads that don't need the latest architectures. Availability is good because enterprises are retiring V100 clusters at scale.
NVIDIA RTX 6000 Ada
| Configuration | New Price | Used Price | Availability | |---------------|-----------|------------|--------------| | RTX 6000 Ada (48GB) | $6,800–$8,500 | $4,500–$6,000 | Moderate |
The RTX 6000 Ada is an interesting alternative for certain workloads. It's not a data center GPU—no NVLink, no multi-GPU scaling like A100/H100—but for single-GPU training and development, it's compelling. The used market is thin because these are workstation cards with longer lifecycles.
Why H100 Availability Is Tight
Several dynamics are keeping H100 supply constrained:
Hyperscaler absorption. AWS, Azure, Google Cloud, and CoreWeave are buying every H100 NVIDIA can produce. What's left for enterprise direct purchase is limited, and those buyers aren't reselling.
Long deployment cycles. H100s bought in 2023–2024 are just now hitting production workloads. Organizations aren't refreshing yet—they're still in the deployment and optimization phase.
H200 transition. Some buyers are skipping H100 entirely and waiting for H200 allocation. This would normally free up H100 supply, but demand is so strong that the effect has been minimal.
Export restrictions. Geopolitical restrictions on AI chip exports have created parallel markets. GPUs that can't be sold to certain countries stay in domestic circulation, reducing available supply in unrestricted markets.
Refurbished vs New: The Pricing Delta
The table below summarizes the refurbished-to-new pricing ratios:
| GPU | Used/New Price Ratio | Notes | |-----|---------------------|-------| | H100 SXM5 | 72–75% | Exceptionally high due to supply constraints | | H100 PCIe | 73–75% | Same dynamics as SXM | | A100 80GB | 55–65% | Normalizing as supply increases | | A100 40GB | 50–60% | Good value for inference workloads | | V100 32GB | 40–55% | Aging but viable for lighter workloads | | RTX 6000 Ada | 65–70% | Limited supply, workstation use case |
The H100 ratio is historically high. For context, A100s were trading at 80%+ of new in 2022—today they're at 55–65%. When H100 supply eventually loosens, expect that ratio to drop to 60–70%.
Where to Source Used GPUs
Enterprise liquidators: When companies fail or downsize, their GPU clusters hit the market through specialized liquidators. These can be good deals but require fast action and cash-ready purchasing.
Broker networks: Established brokers like Mark III, Stallard Technologies, and Park Place maintain GPU inventories. Prices are higher than direct liquidation but so is reliability.
Auction sites: eBay and specialized hardware auction sites have GPU listings, but buyer beware—verify seller ratings, request benchmark results, and confirm warranty terms.
Cloud provider surplus: Some cloud providers sell retired hardware. This is a growing channel as hyperscalers refresh their own infrastructure.
Direct from enterprises: Large enterprises sometimes sell equipment directly. This requires relationships and timing but can yield the best pricing.
Sourcing Gotchas
SXM vs PCIe: SXM modules require specific server platforms (DGX, HGX, or compatible OEM servers). Don't buy SXM GPUs unless you have the right chassis. PCIe cards are more flexible but have lower interconnect bandwidth.
Cooling requirements: H100s and A100s are power-hungry. Verify your data center can handle the thermal load—300W+ per GPU is standard.
Warranty limitations: Used GPUs typically come with 30–90 day warranties from brokers. This is much shorter than new hardware. Budget for potential failures.
Firmware and drivers: Used GPUs should work with current NVIDIA drivers, but verify compatibility with your software stack before purchasing.
Mining history: Ask specifically if GPUs were used for cryptocurrency mining. Mining cards may have excessive wear even if they pass initial tests.
FAQ
Q: Should I buy used H100s at 75% of new pricing? A: It depends on your timeline. If you need GPUs now and can't get new allocation, used H100s at $18K–$24K are your best option. If you can wait 3–6 months for new allocation at $25K–$32K, the manufacturer warranty and support may be worth the premium.
Q: Are used A100s a good value in 2026? A: Yes, particularly the 40GB variants at $4,000–$5,500. For inference workloads and many training scenarios, A100s remain highly capable. The 80GB versions are also reasonable at $6,500–$9,000 if you need the extra memory.
Q: How do I verify a used GPU is legitimate and functional? A: Request: (1) GPU-Z or nvidia-smi output showing specs, (2) burn-in test results, (3) physical photos of the card including serial numbers, (4) seller warranty terms, and (5) references from previous buyers. For large purchases, consider independent inspection.
Q: What's the expected lifespan of a used A100 or H100? A: Data center GPUs are built for 5+ year lifecycles. A used A100 with 2–3 years of prior service should have 3+ years of useful life remaining. H100s are too new for meaningful failure data, but expect similar durability.
Q: Can I mix new and used GPUs in the same cluster? A: Technically yes, but be careful. Mixing GPU generations (V100 with A100, for example) complicates workload scheduling. Mixing new and used A100s is more manageable but monitor for performance consistency.
Need current GPU pricing or availability? Request a quote and our team will source options from our broker network.
Related Parts
Caladan stocks used and refurbished parts referenced in this article — tested, inspected, and ready to ship.