H100 GPU Monthly Rental Cost in 2026: Real Pricing Guide
H100 SXM5 monthly rental in 2026: from $1,070 (Vast.ai spot, 720h) to $8,855 (AWS p5 on-demand). Full hourly-to-monthly math across 12 providers.
An H100 GPU rents for between $1,070 and $8,855 per month in 2026 — an 8× spread for the same hardware. The cheapest options are specialty GPU clouds (RunPod, Vast.ai, Lambda Labs, Hyperbolic); the most expensive are hyperscalers (AWS, GCP, Azure). This guide shows monthly math for 12 providers and explains when each one wins. For real-time hourly-to-monthly cost calculation, use our GPU Pricing Calculator.
For monthly H100 rental decisions, the choice rarely comes down to hourly price alone. It comes down to: do you need hyperscaler ecosystem (VPC, IAM, compliance certifications) or just raw H100 hours?
What does an H100 actually cost per month in 2026?
Monthly cost calculation (730 hours = 24h × 30.4 days, the standard cloud billing month):
| Provider | $/hour | $/month | Notes |
|---|---|---|---|
| Vast.ai (spot) | $1.80 | $1,314 | 24h median spot price |
| Hyperbolic | $1.49 | $1,088 | Spot-style; community reliability |
| RunPod Community (spot) | $1.65 | $1,205 | 50%+ uptime, eviction risk |
| RunPod Community (on-demand) | $2.39 | $1,745 | Community-tier on-demand |
| Vast.ai (on-demand) | $2.40 | $1,752 | 24h median |
| RunPod Secure | $2.99 | $2,183 | Best price-reliability ratio |
| Lambda Labs | $2.99 | $2,183 | Pure on-demand |
| CoreWeave | $3.30 | $2,409 | Enterprise; often contract-only |
| Paperspace | $5.95 | $4,344 | Consumer-friendly UI premium |
| GCP A3 (us-central1) | $11.06 | $8,074 | Per-GPU from 8-GPU node |
| AWS p5 (us-east-1) | $12.29 | $8,972 | Per-GPU from p5.48xlarge |
| Azure ND-H100-v5 | $12.96 | $9,461 | Per-GPU |
The 8.7× spread between Vast.ai spot ($1,314) and Azure on-demand ($9,461) is real. Hyperscalers charge a 4-7× premium for the same H100 silicon because they bundle enterprise networking, IAM, compliance, and regional redundancy that specialty clouds don't.
When is hyperscaler H100 pricing worth it?
Three scenarios where you should pay AWS/GCP/Azure's premium:
1. Data residency and compliance
If your training data is in S3 with regulatory constraints (HIPAA, FedRAMP, GDPR with strict EU residency), running compute on a different cloud incurs cross-cloud egress and potentially violates compliance. Hyperscaler GPU rental keeps everything inside your existing compliance envelope.
2. Enterprise VPC requirements
Production workloads inside private VPCs with strict networking policies often can't connect to specialty GPU clouds via public internet. Hyperscaler integration is sometimes the only path.
3. Existing committed-use discounts
If you have AWS Enterprise Discount Program (EDP) or GCP CUD commits, your effective H100 rate may already be discounted 30-50%. RunPod or Vast.ai may not actually be cheaper after accounting for committed savings.
For everything else — research, fine-tuning, batch inference, startup experiments — specialty GPU clouds win decisively on cost.
Spot vs on-demand for monthly H100 use
Spot pricing math for AWS p5 over a 30-day month:
On-demand: 730 × $12.29 = $8,972
Spot (50% off): 730 × $6.40 = $4,672
Savings: $4,300/month
But factor eviction risk:
- AWS Spot evicts every 1-3 days median
- Each eviction = checkpoint recovery time, ~5-10 min lost
- Net effective uptime: ~85-95% of advertised
- Adjusted real cost: $4,672 / 0.90 = $5,191/month
Even adjusted for eviction overhead, spot saves $3,800/month on AWS. The math works for any workload that checkpoints properly (training, batch inference). For real-time HTTP serving, spot is too volatile.
On specialty clouds (RunPod, Vast.ai), the on-demand price is already so low that spot savings are modest in absolute terms — maybe $500-1000/month additional savings.
Should I reserve for 1 year?
AWS p5 Reserved Instance pricing at standard 1-year:
On-demand: 730 × $12.29 = $8,972/month
Reserved (1-yr, partial upfront): 730 × $7.50 = $5,475/month
Reserved (1-yr, all upfront): effective $5,000/month
Reserved breaks even at:
Utilization > 70% sustained over the year
If you genuinely run an H100 24/7 for 12 months, reservation saves $40-50k/year per GPU. If you run it sporadically (research, batch jobs), reservation strands capacity and effectively costs more than on-demand. Most teams over-commit on reservations — calculate utilization before signing.
GCP Committed Use Discounts (CUDs) work similarly. Azure Reserved VM Instances (RIs) similar discount math but with more flexible cancellation.
What hidden costs eat the monthly H100 bill?
Five line items that compound the headline rate:
1. Storage
H100 instances typically include 1-2TB local NVMe + you'll attach 100GB-1TB persistent storage. EBS gp3 on AWS: ~$0.08/GB-month. A 1TB attached volume: $80/month. Snapshots and backups add another 20-50%.
2. Egress bandwidth
Cross-AZ: $0.01/GB. Cross-region: $0.02-0.04/GB. Internet egress: $0.09/GB. For inference workloads streaming outputs to users, egress can add hundreds per month. Self-hosted Llama serving 100M output tokens/month generates ~50GB egress at $0.09 = $4.50 — trivial. Audio/video gen workloads can hit $1000+ egress easily.
3. Networking between nodes
For multi-GPU training requiring NVLink or InfiniBand between nodes, hyperscalers charge premium networking surcharges. AWS EFA-enabled instances are 10-15% more. Specialty clouds usually include high-bandwidth networking in headline rate.
4. Idle time
The most expensive H100 is the one running with no traffic. A typical research workflow uses GPU 8-10 hours/day, idle 14-16 hours. Without auto-shutdown, you're paying full on-demand for idle time. Use scheduled shutdowns, queue-based job execution, or spot pricing to avoid waste.
5. Snapshot and image storage
Custom AMIs and container images for ML environments are 10-50GB. Snapshot retention costs add up. Plan $50-200/month for image storage on long-running projects.
For comprehensive cost forecasting that captures these, use our GPU Pricing Calculator. For combined GPU + inference + storage cost modeling, see Agent Dev Cost Calculator.
What's the cheapest reliable H100 monthly rental?
By workload type:
For sustained inference serving (24/7, real users)
Winner: RunPod Secure Cloud at $2,183/month. Datacenter-grade reliability, no eviction risk, 99.5%+ uptime. The 30% premium over Community Cloud is worth it for production HTTP serving.
For training runs (intermittent, checkpointable)
Winner: RunPod Community Spot at $1,205/month. Eviction tolerance is high for training because all modern frameworks checkpoint frequently. Vast.ai spot is slightly cheaper but less consistent.
For research / experimentation
Winner: Vast.ai community at ~$1,300-1,752/month. Cheapest reliable option for one-off experiments. Use spot for batch jobs, on-demand for interactive Jupyter work.
For enterprise workloads in VPC
Winner: AWS p5 Reserved at $5,475/month. Pay the premium for ecosystem integration, get 40% off list with 1-year commit. Negotiate enterprise pricing if spending >$25k/month total.
What's the real total monthly cost of running an H100 cluster?
A realistic 4× H100 cluster monthly bill for a fine-tuning workload on RunPod Secure:
GPU rental: 4 × $2,183 = $8,732
Storage (2TB persistent): $200
Egress (50GB/month): $5
Container registry / model storage: $50
Monitoring (Grafana Cloud): $50
Total: ~$9,000/month
Same workload on AWS p5 (single instance, all 8 GPUs):
p5.48xlarge on-demand: $98.32/hour × 730 = $71,773
... wait, that's the full 8-GPU instance
Per-GPU: $8,972
For 4 GPUs equivalent capacity: ~$35,890
Plus EBS, egress, support plan, etc: ~$38,000/month
The 4× cost spread is the reason most AI startups in 2026 use specialty GPU clouds, not hyperscalers.
What's coming for H100 pricing through 2026?
Three trends to watch:
- H200 supply normalization — as H200 capacity grows, H100 spot prices will drop another 20-30% by Q4 2026.
- B200 displacement — B200 is 1.5-2× faster than H100 at similar power. For sustained workloads, B200 will replace H100 as the default training/inference GPU.
- Provider consolidation — expect 1-2 specialty GPU clouds to be acquired by hyperscalers. Pricing dynamics may shift if RunPod or Lambda Labs gets bought.
For ongoing monthly cost tracking, our GPU Pricing Calculator refreshes pricing on the first of every month. For benchmarking inference throughput across providers to decide cost-per-token, see the Inference Benchmark.
The right H100 provider in 2026 depends entirely on workload reliability requirements. Pay for hyperscaler tax only if you genuinely need the ecosystem; otherwise specialty clouds save 60-80%.