GPU cloud pricing

// billing models differ — read this before comparing

• RunPod Pods — per-second, always-on container (Secure Cloud)
• RunPod Serverless — per-second, scale-to-zero workers
• Modal — per-second, scale-to-zero ($30/mo free credit)
• Fal.ai — per-hour list price for custom deployments; committed-use discounts advertised
• Baseten — per-minute, dedicated deployments, scale-to-zero
• Replicate — per-second, private model deployments
• Lambda Cloud — per-hour on-demand GPU instances; prices vary by instance GPU count
• CoreWeave Inference — per-hour single-GPU inference rate; inference platform customers only (contact account executive)
• Nebius AI Cloud — per-second VM compute, shown as on-demand GPU-hour; H100/H200/B200/B300/RTX PRO rates include prescribed vCPU and RAM, while L40S is the minimum all-in 1-GPU preset
• Hyperstack — per-minute on-demand GPU VM, shown per GPU-hour; fixed CPU, RAM, root disk, and ephemeral disk are included, while public IPs and shared storage are separate
• Novita AI — per-second on-demand GPU instance, settled hourly; listed one-GPU compute prices include the product's fixed vCPU and RAM, with storage above the free container-disk quota billed separately
• Verda (formerly DataCrunch) — prepaid pay-as-you-go GPU instances in 10-minute increments, with unused terminated time refunded in the next billing period; listed one-GPU price includes fixed CPU and RAM, while storage is separate
• Salad Container Engine — per-second managed container instances at Lowest (formerly Batch) priority; GPU rates include selected vCPU and RAM, allocation and image-download time are unbilled, and distributed nodes are interruptible
• Vultr Cloud GPU — hourly on-demand GPU instances using actual calendar-month hours; listed instance rates include fixed vCPU, RAM, local storage, and outbound bandwidth allocations, and stopped instances continue billing until destroyed
• DigitalOcean GPU Droplets — per-second on-demand GPU VMs with a 60-second or $0.01 minimum; listed one-GPU rates include fixed vCPU, RAM, boot/scratch storage, and outbound transfer allocations, and powered-off Droplets continue billing until destroyed
• Paperspace Machines — per-hour GPU VM compute rates; compute billing stops only when Off, while the default 50 GB disk is separately billed at $0.0074/hour (maximum $5/month); dynamic IPs disappear at power-off, static IPs and other add-ons continue billing until removed; ingress and egress bandwidth are free
• Thunder Compute — per-minute on-demand GPU VMs while running, shown as one-GPU base rates; four vCPUs and 8 GB RAM per vCPU are included in the base, additional vCPUs cost $0.04/hour, the first 100 GB persistent disk per GPU is included, and snapshots are separate
• Massed Compute — per-minute debit for active on-demand GPU VMs using published hourly VM totals; one-GPU rates include fixed vCPU, RAM, and listed storage allocations; no bandwidth charges or long-term contracts
• Jarvis Labs — per-minute on-demand managed Templates or GPU VMs, shown per GPU-hour; public one-GPU rows include listed vCPU and RAM, paused instances stop compute billing while storage continues at $0.10/GB-month, and in-region egress is free
• Koyeb — per-second pay-per-use serverless GPU instances, shown at exact public hourly list rates; one-GPU configurations include fixed vCPU, RAM, and local disk, autoscaling is available, and saving-plan discounts are excluded
• Northflank — per-second managed-cloud GPU workloads, shown at exact public GPU-hour component rates; CPU, memory, persistent disk, and egress are priced separately, GPU deployment requires at least $50 account credit, and model availability varies by region

// methodology

Prices are read from each vendor's public pricing page and converted to hourly equivalents (per-second × 3600, per-minute × 60), rounded to cents. "As low as" and committed-use rates are noted but never used in the comparison cells. We verify the whole table twice a week and stamp the date above; the raw JSON with source URLs is downloadable. If a number is wrong, the fix is one verification away — tell us.

Want cost per workload instead of cost per hour? Use the LLM hosting cost calculator.

Need to choose a deployment shape, not just a rate? Compare serverless GPU idle tails and scale-to-zero rules before treating a VM, serverless worker, and managed endpoint as interchangeable.