Running large AI models requires burstable compute resources that can scale up/down based on workload needs. Current solutions lack transparent performance/cost benchmarks across providers.
Create a SaaS that automatically tests model training/inference across cloud GPU providers (AWS, GCP, Azure, etc.) with standardized benchmarks. Output ranking by price-performance for specific workload types.
Research labs and AI startups would pay for this to optimize cloud spend. The service could charge per benchmark or via subscription.
MVP: Simple CLI tool that runs predefined benchmarks on user's cloud accounts with basic cost-performance reports.
Biggest risk: Major cloud providers might release their own benchmarking tools, undercutting the need.