Blackwell is here — and for buyers of used H100 hardware, that's the best news of the year.
NVIDIA's GB200 NVL72 rack system represents the most significant leap in AI compute density the industry has ever shipped. But its arrival is setting off a chain reaction that savvy infrastructure buyers should understand: a wave of H100 and A100 decommissioning is sweeping through hyperscalers and enterprise data centers, and the secondary market is about to get very interesting.
What Is the GB200 NVL72?
The NVL72 is NVIDIA's full-rack Blackwell system — 72 B200 GPUs and 36 Grace CPUs interconnected by NVLink Switch fabric at 1.8TB/s of bandwidth per GPU. It delivers roughly five times the AI training performance and thirty times the inference throughput of an equivalent H100 system for certain large model workloads. The price tag reflects this: a full NVL72 rack runs north of $3 million.
Hyperscalers — AWS, Azure, Google Cloud, Oracle — have been standing up Blackwell clusters since late 2025. Large AI labs followed. Which means their H100 infrastructure, much of it only 18–30 months old, is beginning to be rotated out.
Why This Is Driving a Used H100 Supply Surge
Every generational GPU transition creates a secondary market opportunity. What makes the Hopper-to-Blackwell transition different is the sheer scale of H100 deployment that happened in 2023 and 2024. At the height of the AI infrastructure buildout, tens of thousands of H100 SXM5 and PCIe cards were installed globally — in hyperscaler clusters, in enterprise AI servers, and in co-location facilities. Now those systems are beginning their depreciation cycle at exactly the moment Blackwell demand is pulling operator attention forward.
The result: a meaningful increase in used H100 availability, with pricing that continues to move in favor of buyers.
What You Actually Get With a Used H100
The H100 is not an old GPU. It remains one of the most capable inference and fine-tuning accelerators available, and for the vast majority of enterprise AI workloads — running open-weight models, supporting RAG pipelines, handling real-time inference, fine-tuning domain-specific models — the H100 delivers exactly what you need at a fraction of what Blackwell costs.
- 80GB HBM2e per GPU: Sufficient for running LLaMA 3 70B in FP16, Mistral, Falcon, and a wide range of production inference workloads.
- 3.35TB/s memory bandwidth: Bottleneck-free for most transformer architectures in enterprise deployment.
- SXM5 form factor with NVLink 4.0: Scale-up multi-GPU configurations with 900GB/s GPU-to-GPU bandwidth — appropriate for 8-GPU inference nodes.
- DGX H100 compatibility: Full integration with NVIDIA's software stack, CUDA ecosystem, and NIM inference microservices.
For most mid-market enterprises, the H100 is not the second-best choice. It is the right choice — especially when the total cost of ownership is measured against what Blackwell commands.
The NVL72 vs. Used H100: A Realistic Cost Comparison
A single NVL72 rack at list price exceeds $3M. A comparable used H100 DGX system — eight H100 SXM5 GPUs, NVLink fabric, NVMe storage, and redundant networking — can be sourced in the $150,000–$250,000 range depending on configuration and condition. For organizations that need serious AI inference capacity but are not training frontier models from scratch, this cost differential is not a rounding error. It's a strategic decision.
Put plainly: for many enterprise use cases, a used H100 cluster delivers 80% of the capability at 10–15% of the cost. That ratio changes the calculus for every finance team in the conversation.
The Networking Wrinkle: Why Blackwell and H100 Don't Mix
One underappreciated aspect of the Blackwell transition is interconnect incompatibility. The NVL72 uses NVIDIA's NVLink Switch — a proprietary fabric that does not interoperate with InfiniBand, RoCE v2, or legacy NVSwitch-based H100 clusters. If your organization is considering mixing generations, you're looking at two distinct networking fabrics running in parallel. For most buyers, this is not practical. It is, however, an argument for building a clean, dedicated H100 cluster rather than attempting to integrate into a Blackwell environment you don't fully control.
This also means the InfiniBand and high-speed Ethernet gear that populated H100 data centers — Mellanox HDR/NDR switches, ConnectX-7 adapters, QSFP112 optics — is hitting the secondary market alongside the GPUs themselves. If you're building or expanding an on-prem AI network, this is your window.
Who Should Be Looking at the Used H100 Market Right Now
Not every organization needs Blackwell. The buyers who stand to benefit most from the current H100 availability surge are those who:
- Run inference workloads on open-weight models at a production scale that makes cloud API pricing painful
- Need data sovereignty — regulated industries where model inputs and outputs cannot leave your infrastructure
- Are building or expanding an on-premises AI capability and want to own the economics rather than rent indefinitely
- Have already concluded that cloud repatriation makes financial sense and need the GPU layer to execute it
If that sounds like your organization, the timing is working in your favor. H100 supply is increasing. Pricing is softening. And the systems coming off hyperscaler depreciation cycles are well-maintained, properly documented hardware — not end-of-life equipment.
What Resilient Tec Carries
We source and sell decommissioned H100 SXM5 and PCIe GPUs, DGX H100 systems, and the interconnects and networking gear that makes them run — including HDR and NDR InfiniBand switches, ConnectX-7 adapters, and high-density QSFP optics. If you're planning an on-premises AI buildout, reach out. The market window for quality used H100 hardware at current pricing won't stay open indefinitely.