Compute Exchange
Hardware market/Used GPUs

USED GPUs

Pre-owned hardware from verified data center operators

Used GPUs represent the largest segment of the Hardware Market. These are accelerators that have been deployed in production data centers and are now being cycled out as organizations upgrade. Pricing reflects deployment duration and condition — typically 40 to 60 percent below original MSRP.

Why USED GPUs

KEY ADVANTAGES

40 to 60 Percent Below MSRP

Significant discounts to new pricing, especially for previous-generation models as Blackwell supply expands and Hopper deployments roll off.

Proven in Production

Hardware that has run continuous workloads in real data centers. Most failures occur within the first 90 days; surviving units have demonstrated baseline reliability.

Immediate Availability

No 6 to 12 month allocation lead times. Used hardware is sourceable from existing supplier inventory in days to weeks.

Strong for Inference

Most production inference workloads do not require the latest silicon. Used GPUs deliver excellent throughput per dollar for serving LLMs and computer vision models.

Browse Models

USED GPUs

Select a GPU model to see indicative pricing, supply commentary, and FAQ specific to used hardware.

Indicative ranges based on aggregated supplier quotes and broker market data, Q3 2026. Final pricing depends on quantity, region, condition, accessory bundling, and warranty terms. Compute Exchange does not publish active counterparty listings.

Frequently Asked Questions

USED GPUs EXPLAINED

What does "used" mean on the Hardware Market?+
A used GPU is a unit that has been deployed in a production data center and is now being resold. Used hardware is sold as-is, with the supplier disclosing approximate runtime hours, original buyer if available, and known issues. There is typically no manufacturer warranty remaining; some sellers offer their own limited return windows.
How much do used data center GPUs cost compared to MSRP?+
Used GPUs typically trade 40 to 60 percent below original MSRP. The discount widens for older generations and narrows for current-generation parts where supply is constrained. Used H100 SXM5 trades around 25 to 30 percent below MSRP today, while used A100 40GB trades closer to 50 to 60 percent below MSRP.
What warranty applies to a used GPU?+
Most used GPUs are sold without active manufacturer warranty — NVIDIA enterprise warranties are typically tied to the original buyer and chassis. Suppliers may offer their own limited warranty, usually 30 to 90 days against dead-on-arrival or early failure. Compute Exchange surfaces warranty terms during the quote process so they can be compared across counterparties.
How do I assess the health of a used GPU before buying?+
Standard checks include the supplier's runtime hours and thermal logs, NVIDIA NVSMI diagnostic output, ECC error counts, and physical inspection of capacitors, baseboard connectors, and PCIe traces. Buyers running large orders typically negotiate burn-in testing or a return window keyed to NVSMI diagnostic results.
Are used GPUs reliable for production AI workloads?+
Yes, with disciplined buying. Most data center GPUs that survive the first 90 days of deployment continue to operate reliably for years. The main failure modes — thermal management, capacitor aging, and HBM stack stress — present early. Used hardware that has run 12 to 24 months of clean operation is statistically more reliable than fresh hardware.
Which used GPUs deliver the best value in 2026?+
Used A100 80GB at $8,000 to $10,000 is consistently the best value for production inference and fine-tuning of sub-70B models. Used H100 SXM5 at $22,000 to $27,000 offers strong economics for distributed training where Hopper-class compute and NVLink fabric matter. Used L40S is competitive for FP8 inference at lower power.
DISCLAIMER

Compute Exchange is a matching service for GPU buyers and sellers. We do not certify, inspect, refurbish, or take ownership of hardware. Condition verification, warranty terms, and transaction details are negotiated directly between counterparties.