Used NVIDIA H100 80GB PCIe/SXM GPU

How Much Does NVIDIA H100 80GB GPU Cost in 2026?

Used H100 80GB units range from $15,000 to $30,000 in 2026, with PCIe Gen5 models typically $2,000-$4,000 cheaper than SXM5 variants. Refurbished cards with warranty sit at $22,000–$28,000, while high-hour mining reclamation units bottom out around $15,000. This represents 35–57% savings versus new H100s, which still command $35,000-$40,000 when available through authorized distributors. Critical caveat: cards with >18,000 operating hours or those exceeding 85°C junction temperature for >40% of runtime show measurable HBM3 degradation—expect 8-12% slower memory bandwidth on STREAM benchmarks compared to factory specs. SXM5 models from liquid-cooled DGX H100 systems typically command 15-20% premiums due to lower thermal stress, but verify backplane compatibility before purchase.

NVIDIA H100 80GB Specifications and Compatible Infrastructure

The H100 delivers 989 INT8 TOPS and 1,979 FP8 TOPS via fourth-gen Tensor Cores, fed by 80GB HBM3 at 3.35 TB/s bandwidth. PCIe variant draws 350W TDP over Gen5 x16 interface, while SXM5 models reach 700W with NVLink Switch System support (900 GB/s bidirectional). Compatible with HPE Cray EX systems (2022+), Dell PowerEdge XE9680 (2023+), Supermicro SYS-521GE-TNRT (2022+), and any server supporting PCIe 5.0 CEM spec with adequate airflow (>200 CFM at GPU zone). Critical exclusion: H100 PCIe cards require PCIe bifurcation support and won't POST in legacy Dell R740/R750 systems without BIOS 2.18.0+ (released March 2023). SXM5 models are mechanically incompatible with SXM4 sockets from A100 generation—baseboard upgrade mandatory despite similar tray dimensions.

What to Check Before Buying a Used NVIDIA H100 in 2026

1. Run nvidia-smi --query-gpu=memory.total,temperature.gpu,clocks.max.gr --format=csv and verify reported 81,559 MiB total memory matches 80GB spec within 1%. Cards showing <80,000 MiB have failed HBM3 stacks.

2. Execute GPU memory test (CUDA MemTest or similar) for minimum 12 hours under 95%+ utilization. Reject any card with >100 ECC errors or single-bit error rate exceeding 1 error per 10^15 bits—indicates HBM3 approaching end-of-life.

3. Inspect NVLink connectors (SXM5 only) for bent pins or discoloration. Each of four NVLink interfaces should show <0.5 dB signal loss at 50 GHz on TDR testing. Bridges from hyperscale decommissions often have pin damage from rough extraction.

4. Verify firmware version is 96.00.7E.00.01 or newer via nvidia-smi -q. Earlier builds have known issues with PCIe 5.0 FLIT mode causing system hangs on AMD EPYC 9004 platforms.

5. Check thermal interface material condition—cards operated in mining/AI training farms beyond 24 months often have dried TIM. Baseline GPU should idle at 35-40°C ambient; >50°C idle suggests thermal path degradation requiring $400-600 rework.

FAQ

What's the difference between H100 PCIe and H100 SXM5 for AI workloads?
SXM5 provides 2x power budget (700W vs 350W) and NVLink connectivity for multi-GPU scaling, delivering 30-40% faster training on large language models exceeding 70B parameters. PCIe variant suits single-GPU inference where NVLink isn't required.

Can I use H100 GPUs from Meta/Google datacenter liquidations?
Yes, but verify no custom firmware locks remain. Meta's RSC cluster used standard NVIDIA firmware, while some Google TPU hybrid racks included H100s with telemetry agents requiring removal. Request seller run nvidia-smi successfully before purchase.

How do I verify HBM3 memory isn't degraded on a used H100?
Run bandwidth tests (CUDA bandwidthTest or STREAM) and compare against 3.35 TB/s spec. Acceptable used units should achieve >3.25 TB/s. Below 3.1 TB/s indicates >7% degradation from temperature abuse or manufacturing defects.

H100 vs A100 80GB for inference in 2026—which offers better value?
H100 provides 3x faster FP8 inference via Transformer Engine but costs 2-2.5x more used. For INT8 batch inference on models <30B parameters, A100 at $6,000-$9,000 delivers better tokens-per-dollar unless you need sub-10ms latency.