B200 GPU Rental Cost Per Hour 2026: Guide

Quick Answer

B200 and GB200 GPU rental cost per hour 2026: cheapest verified is CoreWeave HGX B200 spot at $4.26/GPU-hr ($34.11/hr per 8-GPU node). On-demand B200 from RunPod Community at $5.98/GPU-hr; Lambda 8x $6.69/GPU-hr; CoreWeave HGX $8.60/GPU-hr on-demand; AWS p6-b200.48xlarge Capacity Blocks at $12.36/GPU-hr. GB200 NVL72 CoreWeave at $10.50/GPU-hr (4-GPU instance).

Last verified: Sep 16, 2026.

At a glance

• B200 cheapest: CoreWeave spot $4.26/GPU-hr ($34.11/hr 8x node)
• B200 on-demand: RunPod Community $5.98/GPU-hr | Lambda 8x $6.69/GPU-hr | CoreWeave HGX $8.60/GPU-hr | AWS p6 $12.36/GPU-hr
• B300: RunPod Community $6.94/GPU-hr | CoreWeave HGX spot $4.48/GPU-hr | AWS p6 $14.04/GPU-hr
• GB200 NVL72: CoreWeave 4-GPU instance $42.00/hr = $10.50/GPU-hr
• Full NVL72 rack (72 GPUs): $756/hr = $544,320/month
• B200 vs H100 break-even: B200 wins if throughput >2x H100
• B200 memory: 192 GB HBM3e; 8 TB/s bandwidth
• GB200 NVL72 memory: 13.5 TB HBM3e per rack; 1.4 exaflops FP4

Why B200 and GB200 are different markets in 2026

B200 is the Blackwell-generation GPU shipping in HGX 8x nodes; GB200 NVL72 is the rack-scale Superchip building block for frontier training — they serve different workloads.

B200 delivers 192 GB HBM3e memory and 8 TB/s bandwidth per GPU. In an 8x HGX node, B200 enables dense transformer training and inference workloads that fit within the memory of a single node. GB200 NVL72 bundles two Grace Blackwell Superchips (4 Blackwell GPUs plus 2 Grace CPUs each) into a 4-GPU instance — the building block for Vera Rubin rack-scale deployments. A full NVL72 rack delivers 13.5 TB HBM3e memory and 1.4 exaflops FP4 throughput, but at $544,320/month continuous operation (cloudzero.com, September 2026).

B200 makes sense for production training and inference workloads under 8-GPU scale. GB200 NVL72 makes sense for frontier training jobs (200B+ parameter models, large-scale distributed training) where the rack-scale fabric eliminates the interconnect bottleneck that slows HGX training across nodes.

B200 specialist-cloud pricing

Specialist clouds offer the cheapest B200 rates because they don't enforce 8-GPU minimums.

Verified July 2026 (runpod.io/pricing, lambda.ai/pricing, cloudzero.com):

• RunPod Community B200: $5.98/GPU-hr (third-party hosts)
• RunPod Secure B200: $6.79/GPU-hr (RunPod datacenters)
• Lambda 8x B200 SXM6 (8-GPU instance): $6.69/GPU-hr
• Lambda 1x B200 SXM6: $6.99/GPU-hr (180 GB, 26 vCPUs, 360 GiB RAM, 2.75 TiB SSD)
• CoreWeave HGX B200 8x on-demand: $8.60/GPU-hr ($68.80/hr node)
• CoreWeave HGX B200 8x spot: $4.26/GPU-hr ($34.11/hr node)
• CoreWeave 1-Click Cluster B200 16-GPU reserved: $9.86/GPU-hr (2-week to 1-year)

RunPod's single-GPU B200 at $5.98-$6.79/GPU-hr is the cheapest verified rate among major providers — there's no 8-GPU minimum. Lambda's 1x B200 at $6.99/GPU-hr is a transparent published card. CoreWeave's HGX B200 requires 8 GPUs minimum but offers the deepest spot discount at $4.26/GPU-hr.

B200 hyperscaler pricing

AWS p6-b200 and GCP a4 are the hyperscaler Blackwell options — constrained supply keeps rates at the high end of the category.

Verified August 2026 (tech-insider.org, aws.amazon.com):

• AWS p6-b200.48xlarge (8x B200) Capacity Blocks: $98.84/hr effective ($12.36/GPU-hr); US East
• AWS p6-b200.48xlarge on-demand (rare): ~$112/hr ($14/GPU-hr); constrained supply
• GCP A4 (B200) on-demand: ~$16.11/GPU-hr (highest of tracked providers)
• Azure ND B200 v6: Rolling out; on-demand rates expected in $10-$14/GPU-hr range

AWS Capacity Blocks is the primary hyperscaler path to B200 access in 2026 — reserve a specific capacity for a specific future time window. Compared to specialist clouds, hyperscaler B200 is 30-50% more expensive on per-GPU-hour ($12.36/GPU-hr AWS vs $6.79/GPU-hr RunPod Secure). The premium pays for compliance certifications, multi-region availability, and integration with hyperscaler AI platforms (SageMaker, Vertex AI, Azure ML).

GB200 NVL72 pricing

GB200 NVL72 is the rack-scale Blackwell building block — pricing reflects the unique 4-GPU Superchip configuration.

Verified August 2026 (spheron.network):

• CoreWeave GB200 NVL72 (4-GPU instance) on-demand: $42.00/hr = $10.50/GPU-hr; no spot pricing
• Full NVL72 rack (18 x 4-GPU instances = 72 GPUs): $756/hr = $544,320/month continuous
• GB300 NVL72 (4-GPU instance): Contact sales only; no published rate

GB200 and GB300 NVL72 instances bundle two Grace Blackwell Superchips each (4 Blackwell GPUs plus 2 Grace CPUs), giving the rack-scale building block for Vera Rubin deployments. The Grace CPU provides high-bandwidth coherent memory access across the Superchips — what enables 13.5 TB HBM3e memory and 1.4 exaflops FP4 throughput per full NVL72 rack.

For workloads under 8 GPUs, GB200 NVL72 is overkill — HGX B200 delivers comparable per-GPU performance at lower cost. The NVL72 advantage appears at 16+ GPUs where the rack-scale fabric eliminates the interconnect bottleneck that slows HGX training across nodes.

B300 Blackwell Ultra pricing

B300 is the newest Blackwell-generation GPU with higher memory bandwidth and tensor throughput — supply is constrained and rates reflect it.

Verified July 2026 (runpod.io/pricing, usagepricing.com):

• RunPod Community B300: $6.94/GPU-hr
• RunPod Secure B300: $7.89/GPU-hr
• CoreWeave HGX B300 8x on-demand: Contact sales; spot $35.84/hr = $4.48/GPU-hr
• CoreWeave GB300 NVL72 (4-GPU): Contact sales; no published rate
• AWS p6-b300.48xlarge Capacity Blocks: $112.32/hr effective ($14.04/GPU-hr)
• Modal B300: $7.10/GPU-hr ($0.001972/sec)

B300 supply is constrained in 2026 because the chip is newer than B200 and most Blackwell production capacity is allocated to B200 contracts. Expect B300 rates to drop 10-15% by Q1 2027 as production scales.

B200 vs H100 break-even math

The B200 premium is worth it when B200 throughput advantage is >2x H100.

Verified July 2026 per-GPU ratios:

• RunPod Secure: B200 $6.79/hr / H100 SXM $3.29/hr = 2.06x premium
• Lambda 8x: B200 $6.69/hr / H100 SXM $3.99/hr = 1.68x premium
• CoreWeave HGX on-demand: B200 $8.60/hr / H100 $6.16/hr = 1.40x premium
• CoreWeave HGX spot: B200 $4.26/hr / H100 $2.46/hr = 1.73x premium
• AWS Capacity Blocks: B200 $12.36/hr / H100 $5.191/hr = 2.38x premium

For dense transformer training where B200 throughput is >2x H100 (typical for 70B+ parameter models with FP4 or FP8 precision), B200 wins on dollar cost. For workloads where B200 throughput is <2x H100 (smaller models, fine-tuning, inference), H100 wins. The break-even point is workload-specific and requires benchmarking on the actual training script.

Side-by-side B200/GB200/B300 pricing

How to choose B200/GB200 capacity in 2026

The decision hinges on workload scale, throughput requirement, and Blackwell-specific features.

1. Single-GPU fine-tuning and inference: RunPod B200 Community $5.98/GPU-hr or Secure $6.79/GPU-hr. Lambda 1x B200 SXM6 $6.99/GPU-hr. Single-GPU access without 8-GPU minimum.
2. 8-GPU dense training workloads: CoreWeave HGX B200 spot at $4.26/GPU-hr ($34.11/hr node) for interruptible workloads, or CoreWeave HGX B200 on-demand at $8.60/GPU-hr for non-interruptible training.
3. Vera Rubin rack-scale frontier training: CoreWeave GB200 NVL72 at $10.50/GPU-hr per 4-GPU instance ($42.00/hr). Reserved contracts required for sustained 16+ GPU workloads.
4. Blackwell Ultra (B300) frontier workloads: RunPod B300 Secure $7.89/GPU-hr. Modal B300 $7.10/GPU-hr. AWS p6-b300 Capacity Blocks $14.04/GPU-hr for production deployments.
5. Production Blackwell with compliance: AWS p6-b200 Capacity Blocks at $12.36/GPU-hr. The premium pays for compliance certifications, multi-region availability, and SageMaker integration.

What enterprise buyers should do next

Three actions for organizations evaluating B200/GB200 capacity in 2026.

1. Benchmark your workload on H100 vs B200 before committing to Blackwell capacity. The B200 premium is worth it only if your workload's throughput advantage is >2x H100. Use RunPod B200 single-GPU at $5.98/GPU-hr to benchmark before committing to a reserved 8-GPU contract.
2. Engage sales for reserved contracts on multi-month Blackwell workloads. CoreWeave reserved contracts target $2.46/GPU-hr on H100 with up to 60% off — equivalent B200 reserved targets ~$3.40-$5.20/GPU-hr. Lambda 1-Click Cluster B200 reserved at $8.87-$9.86/GPU-hr for 16-256 GPU reservations.
3. Plan Vera Rubin capacity now. NVIDIA Q2 FY27 earnings (August 26, 2026) confirmed Vera Rubin full production at CoreWeave, Google, Azure, OCI, and Nebius. Vera Rubin rack pricing expected by Q1 2027; reserve capacity now if frontier training is on the 2027 roadmap.

What to watch next

Three near-term datapoints. First, B200 supply normalization — current constrained supply keeps rates at the high end; expect 10-15% list rate reductions by Q1 2027 as production scales. Second, Vera Rubin pricing on hyperscalers and specialist clouds — Vera Rubin full production is confirmed at CoreWeave, Google, Azure, OCI, and Nebius; per-GPU-hr rates expected by Q1 2027. Third, GB300 NVL72 published pricing — current contact-sales-only posture reflects supply allocation; expect published rates by Q4 2026. NVIDIA Q2 FY27 earnings confirm Vera Rubin full production and hyperscaler capex of $800B 2026 / $1.3T 2027 — these numbers signal sustained Blackwell demand and gradual supply normalization through 2027 (cloudzero.com, September 2026).