NVIDIA B200 production ramps — used H100 SXM softens to ~$22-25k

**For new cap-ex**: pick [H200 SXM](/hardware/nvidia-h200) over H100 SXM for nearly all new builds — 76% more memory + 43% more bandwidth at +20-25% price is almost always worth it. H100 SXM only when matching existing cluster. **For frontier training**: [B200](/hardware/nvidia-b200) at $40k retail breakeven vs H200 cluster on FP4-aggressive workloads. Run the math against your actual workload — if FP4 gain isn't ≥30%, H200 cluster wins on $/throughput. **For used market**: used H100 SXM at $22-25k is a buying opportunity if you have existing DGX H100 infrastructure. Capacity adds make sense; greenfield H100 SXM cluster does not. **For cloud rental**: H100 SXM rental dropped to $2.50-3.50/hr on Runpod/Lambda. Cap-ex breakeven now requires sustained 70%+ utilization for 12+ months. See [B200 vs H200 verdict](/compare/hardware/nvidia-b200-vs-nvidia-h200), [H100 SXM verdict](/hardware/nvidia-h100-sxm).