NVIDIA B200 Pricing (June 2026)
Nathanael Chiang
The NVIDIA B200 is the flagship of the Blackwell generation: 192 GB of HBM3e, 8 TB/s of memory bandwidth, and native FP4 support that roughly doubles inference throughput over FP8. Every team running frontier-scale models wants one. It also has the most fragmented pricing in the market right now.
The breakdown below compares current B200 on-demand pricing. Supply is still catching up to demand, so the spread between neoclouds and hyperscalers is wide, and that spread is where the savings sit.
Takeaways
- Neoclouds list B200 on-demand at $3.70 to $6.00/hr; hyperscalers charge 3 to 4x more.
- AWS p6-b200 lands near $14.24/hr per GPU, the most expensive tier in the market.
- Spot pricing drops as low as $2.12 to $2.74/hr for checkpoint-friendly workloads.
- Beam starts at $3.93/hr, at the floor of the on-demand market, on a serverless runtime rather than a raw VM.
Current on-demand B200 prices (per GPU)
| Provider | SKU / Instance | On-Demand $/GPU-hr* | Notes |
|---|---|---|---|
| Beam | B200 SXM6 | $3.93 | Serverless runtime, per-second billing |
| Spheron | B200 SXM6 | $3.70 | Marketplace floor; spot ~$2.74 |
| Lambda | 1x B200 | $4.99 | 8x node at $5.29/GPU |
| Nebius | B200 SXM6 | $5.50 | Per-hour billing, no spot |
| RunPod | B200 SXM6 (Secure) | $5.89 | Per-minute billing |
| CoreWeave | N/A | ~$6.50 | Contract-oriented, estimated |
| Azure | ND B200 v6 | ~$14.00 | Full VM shape |
| AWS | p6-b200 (8x B200) | $14.24 | $113.93/hr node / 8 GPUs |
Normalized to a single B200, even when only multi-GPU nodes are offered.
Last update: June 18, 2026
Methodology
Why you can trust these numbers:
- On-demand only: no reservations or prepaid contracts.
- Same silicon: every figure refers to the NVIDIA B200 (192 GB) SKU.
- Public price lists: pulled in June 2026 from each provider's published pricing pages.
- U.S. regions, USD: prices elsewhere vary by 5 to 20%.
Why the B200 spread is so wide
The B200 rollout is still underway, which is why pricing has not settled the way H100 pricing eventually did. MSRP runs $30,000 to $40,000 per GPU in clusters of 8 or more, each card draws about 1,000W, and most hyperscaler capacity is committed to enterprise contracts. On-demand buyers compete for the slice neoclouds make available, and the neoclouds, with leaner overhead, undercut the hyperscalers by three or four times.
The 1-year reserved discount on B200 is also the steepest in the market, roughly 39% off posted on-demand, because providers compete hardest for committed Blackwell capacity.
Where Beam fits
Beam's B200 SXM6 starts at $3.93/hr, at the bottom of the on-demand range, level with the cheapest marketplace floors and about 3.6x cheaper than AWS p6. The difference is what the rate includes. Rather than a bare VM you manage yourself, Beam runs your workload on a serverless runtime with per-second billing, so you do not pay for idle GPU time between jobs. For burst inference and intermittent training, that scale-to-zero behavior often matters more than the hourly number.
Why this matters for developers
- Memory you can use: 192 GB HBM3e plus FP4 fits 70B+ models on a single GPU with room for large batch sizes.
- The hyperscaler premium is large: AWS and Azure B200 instances cost 3 to 4x the neocloud rate for the same silicon, before egress and networking.
- Spot is the biggest lever: if your workload checkpoints, spot B200 around $2.12/hr leads on cost per token.
We refresh these numbers monthly.
