Cost Analysis

Cost Comparison of Clinical AI Training Platforms

Clinical AI training costs vary by architecture, not just by GPU hours.

Cost breakdown per platform

Rapha Protocol (Compute-to-Data)

  • Training fee: Per-job USDC escrow. You set your budget per training run. No hourly GPU pricing. No data transfer costs.
  • Infrastructure: Zero. Rapha deploys and operates the edge appliance at the hospital. No cloud instances to provision. No GPUs to rent.
  • Legal/contracting: Institutional governance approval (varies). No per-job DUA negotiation. No BAA overhead (you never receive PHI).
  • Engineering: Zero FL coordination cost. No gradient leakage mitigation engineering. Submit your model, receive trained weights.

Cloud BAA (AWS/GCP/Azure)

  • Compute: GPU instances at ~$2-5/hour for medical imaging workloads. LLM fine-tuning: $10-30/hour.
  • Storage: PHI at rest in cloud storage. Ongoing cost scales with data volume.
  • Data transfer: Egress from hospital to cloud. Ingress for model artifacts. Ongoing cost per GB.
  • Legal: BAA, DPA negotiation, DPIA. One-time but substantial legal cost.
  • Security: Cloud security configuration, IAM, encryption key management. Engineering overhead ongoing.

Total cost of ownership for a typical radiology AI project (6 months)

Rapha Protocol: $25,000-$250,000 (10-50 training jobs at $2,500-$5,000 each)
Cloud BAA: $150,000-$500,000+ (GPUs + storage + legal + engineering)
Federated learning: $200,000-$750,000+ (FL platform licensing + engineering + coordination)