GPU inference cost projection across cloud, rental, and capex scenarios.

Three model types store scenario inputs (instance rates, hardware costs, facility expenses) and compute normalized $/GPU-hour projections. A comparison report queries all scenarios and produces a ranked table with crossover analysis.

No live API calls — rates are manually entered from quotes, pricing pages, or account team conversations. All scenarios in a comparison must share a currency.

Model Types

• gpu-cloud — Hyperscaler GPU instances (AWS, Azure, GCP) under various capacity models: on-demand, reserved, FTP, Capacity Blocks, etc.
• gpu-rental — Third-party providers (CoreWeave, Lambda Labs, RunPod) with simpler per-GPU-hour pricing and optional commitment discounts.
• gpu-capex — On-premises/colo hardware. Amortizes a capital purchase into synthetic recurring cost, adds facility/staff/maintenance, and normalizes to $/GPU-hour. Includes sensitivity analysis across utilization and useful-life assumptions.

Usage

swamp extension pull @webframp/cost-projection

# Cloud scenario
swamp model create @webframp/cost-projection/gpu-cloud kimi-k3-hyperpod
swamp model method run kimi-k3-hyperpod record \
  --input name=kimi-k3-hyperpod \
  --input provider=aws \
  --input region=us-east-1 \
  --input instanceType=ml.p6-b300.48xlarge \
  --input gpuCount=8 \
  --input gpuModel="NVIDIA B300" \
  --input capacityModel=flexible-training-plan \
  --input instanceRatePerHour=98.50

# Rental scenario
swamp model create @webframp/cost-projection/gpu-rental kimi-k3-coreweave
swamp model method run kimi-k3-coreweave record \
  --input name=kimi-k3-coreweave \
  --input provider=coreweave \
  --input gpuModel="NVIDIA H100 SXM" \
  --input gpuCount=8 \
  --input ratePerGpuHour=2.49

# Capex scenario
swamp model create @webframp/cost-projection/gpu-capex dc-east-b300
swamp model method run dc-east-b300 record \
  --input name=dc-east-b300 \
  --input gpuModel="NVIDIA B300" \
  --input gpuCount=8 \
  --input gpuCostPerUnit=35000 \
  --input serverCost=45000 \
  --input networkingCost=15000 \
  --input totalHardwareCost=340000 \
  --input usefulLifeMonths=36 \
  --input coloCostPerKwMonth=150 \
  --input powerDrawKw=10.5

# Run sensitivity analysis
swamp model method run dc-east-b300 sensitivity

# Compare all scenarios
swamp report run webframp/cost-projection-comparison

Per-Token Break-Even (Optional)

To compare self-hosted cost against per-token API pricing, provide apiComparisonRatePerMToken ($/million tokens from the API provider) and estimatedTokensPerGpuHour (from vLLM benchmarks or model card throughput figures) when recording a scenario. The projection will compute the monthly token volume at which self-hosting breaks even.

Release Notes

2026.08.01.1

Fixed: Report name now uses collective prefix (@webframp/cost-projection-comparison). Previous name (webframp/cost-projection-comparison) failed publish validation.

2026.07.31.1

Initial release.

• Three model types: gpu-cloud, gpu-rental, gpu-capex
• All normalize to $/GPU-hour for cross-scenario comparison
• Capex model includes sensitivity analysis (utilization × useful-life matrix)
• Workspace-scoped comparison report with crossover analysis
• Single-currency assumption (all scenarios must share a currency)
• Optional per-token break-even calculation against API pricing