Methodology & data sources

Cloud GPUs are usually sold in fixed 8-GPU instances. We take the published on-demand instance price and divide by the number of GPUs in that instance:

per_gpu_hour = instance_hourly_price / gpus_per_instance

# Example: AWS p5.48xlarge
#   $98.32 / hr  ÷  8 H100  =  $12.29 per GPU-hour

Your total is then simply rate × gpu_count × hours × months. This isolates compute cost — it deliberately excludes storage, egress, and support.

• On-demand — the publicly listed pay-as-you-go rate with no commitment.
• 1-yr committed — a representative committed-use / reserved / savings-plan rate where the provider publishes one. Where none is published, we fall back to on-demand and flag it with an asterisk.

• Storage, networking, and data-egress fees
• Negotiated enterprise / private-offer discounts
• Spot / preemptible and short-term capacity-block pricing
• Region-specific premiums beyond the listed reference region
• Performance differences (an H100 hour ≠ an A100 hour in throughput)

Performance-normalized cost ($/token, $/effective-FLOP) and full TCO are on the Pro roadmap.

Neither sells a public per-GPU-hour cloud product. Including a fabricated number would undermine the comparison, so we exclude them until a sourced public rate exists.

Each row below is the exact figure used by the calculator, with its source and last-verified date. A ⚠ marks a sourced estimate still pending a final hands-on check against live pricing.