B200 Cost Calculator
Indicative monthly and annual cost for an NVIDIA B200 cluster across 9 GPU clouds and hyperscalers. Adjust GPU count, hours, and term.
An NVIDIA B200 in 2026 costs roughly $5.50–$7.20 per GPU-hour on-demand, with neoclouds (CoreWeave, Crusoe, Lambda, Nebius) at the low end and hyperscalers (AWS, Azure) at the high end. Reserved 1-year contracts cut 35–45% off list. A 64-GPU B200 cluster at 730 hours/month lands between roughly $250K and $340K — about 2.3× the cost of equivalent H100 capacity, with ~2.5× the training throughput.
| Provider | $/GPU-hr | Monthly cost | Annual cost | Notes |
|---|---|---|---|---|
| ★ BestNebius | $4.50 | $105,120 | $1,261,440 | — |
| Together AI | $4.80 | $112,128 | $1,345,536 | Per-token also available |
| Crusoe | $5.00 | $116,800 | $1,401,600 | Stargate-anchor capacity |
| CoreWeave | $5.50 | $128,480 | $1,541,760 | GB200 NVL72 available |
| Lambda | $5.99 | $139,926 | $1,679,117 | — |
| Oracle (BM.GPU.B200.8) | $6.40 | $149,504 | $1,794,048 | Stargate program |
| Azure (ND GB200 v6) | $7.20 | $168,192 | $2,018,304 | OpenAI-aligned |
| Google Cloud (A4) | $7.50 | $175,200 | $2,102,400 | — |
| AWS (p6-b200) | $8.00 | $186,880 | $2,242,560 | EFAv3 |
Pricing reflects public on-demand and 1-year reserved rates as of Q1 2026. GB200 NVL72 rack pricing differs; this calculator models discrete B200 instances. Excludes storage, egress, and support.
Frequently asked questions
- What does a B200 cost per hour in 2026?
- Indicative on-demand pricing ranges from ~$4.50/hr (Nebius) to ~$8.00/hr (AWS p6-b200). Reserved 1-year terms cut 30–50% off. Neoclouds lead on price; hyperscalers charge a 30–50% premium.
- Is GB200 NVL72 priced separately from B200?
- Yes. GB200 NVL72 racks (72 B200s on a single NVLink domain) are typically quoted as a rack unit ($350K–$500K/mo on-demand). The calculator above models discrete B200 pricing for simpler workloads.
- When is B200 worth paying more for vs H200?
- When you need >141 GB HBM3e per GPU (H200's cap), training models >100B parameters, or 4× throughput on FP4-quantized inference. B200 wins decisively on Llama-3 405B, DeepSeek-V3, and frontier MoE training.
- How much does it cost to train a 70B model on B200?
- Llama-3 70B equivalent training on B200 is ~1.5M B200-hours (≈40% reduction vs H100). At reserved $3.00/hr that's ~$4.5M. See our AI Cluster Cost Calculator to model end-to-end TCO.
Company → Datacenter → GPU → Customer → Industry
Every entity on this site is cross-linked. Follow the graph from operators down to specific facilities, GPU clusters, customers, and sectoral context.