H100 vs B200: Real Cloud Rental Cost Comparison (April 2026)

As of April 2026, the cheapest H100 SXM5 on-demand is $1.79/hr at Shadeform, while the cheapest B200 is $5.29/hr at Lambda. That's a 3.0× price premium for B200. For 8-GPU cluster nodes: AWS H100 (p5.48xlarge) runs ~$32.77/hr (~$4.10/GPU), while AWS B200 (p6.48xlarge) runs ~$55.20/hr (~$6.90/GPU). GridStackHub tracks these prices daily across 58+ providers — prices are updated automatically.

For most workloads, no — H100 remains the better value at $1.79/hr. B200's ~2.27× compute and ~2.39× memory bandwidth advantage doesn't fully offset the 3.0× price gap except in specific scenarios: continuous high-throughput batch inference on 70B+ parameter models, very large distributed training jobs where iteration speed matters more than GPU-hours, or workloads where B200's 192GB VRAM (vs H100's 80GB) eliminates model sharding overhead. If your GPU is under 70% utilised, H100 wins on absolute cost. The crossover point is high and depends on workload characteristics — use the GPU cost calculator to model your specific case.

As of April 2026, providers carrying both H100 and B200 on-demand include: Lambda, CoreWeave, RunPod, Google Cloud, AWS, Azure. Lambda is the standout option for single-GPU access to both — H100 from ~$1.99/hr, B200 from ~$5.29/hr — with no waitlist on H100 and generally fast access to B200. Hyperscalers (AWS, Google, Azure) offer both but primarily as 8-GPU cluster nodes with higher effective per-GPU rates.

For memory-bandwidth-bound decoding (the most common LLM inference bottleneck), expect 2–2.5× more tokens/second on B200 vs H100 SXM5. This comes primarily from B200's 2.39× higher memory bandwidth (8.0 TB/s vs 3.35 TB/s), which determines how fast the GPU can load model weights per forward pass. For prefill (compute-bound), the 2.27× FP8 advantage applies more directly. Real-world gains depend on batch size and sequence length — larger batches see closer to the theoretical maximum. B200's 192GB VRAM also allows serving larger models or larger KV caches, which can improve throughput per GPU beyond the raw bandwidth numbers.

Multiple providers offer H100 SXM5 on-demand with no minimum commitment: Shadeform ($1.79/hr), Together AI ($1.99/hr), RunPod ($1.99/hr), Lambda ($1.99/hr), Crusoe Energy ($2.06/hr). The cheapest on-demand H100 is currently $1.79/hr at Shadeform. GPU aggregators like Shadeform route to whichever cloud has the lowest available rate at booking time. For spot pricing (preemptible), Vast.ai and RunPod often go below $1.50/hr but with interruption risk. GridStackHub's cheapest GPU comparison tracks real-time prices across all providers.