What problem does it solve? Choosing where to rent GPU compute for an ML experiment requires manually comparing prices across dozens of clouds, marketplaces, free tiers, and academic grant programs, which is slow and error-prone. ## Core Features & Use Cases - Catalog-driven ranking: Compares 18 providers (AWS, GCP, Azure, Lambda, RunPod, Vast.ai, TensorDock, DataCrunch, and more) against a workload defined by gpu_type, gpu_count, and duration_h, sorted by total estimated USD. - Free and academic tiers: Separately surfaces Colab, Kaggle, HuggingFace ZeroGPU, GCP TRC, NSF ACCESS, and national HPC centers when they match the workload. - Filters and fallbacks: Supports max_usd_per_hour budget caps, spot-price preference, region matching, GPU alias resolution, and fuzzy suggestions when a GPU type is unknown. - Use Case: After estimating that an experiment needs 1x A100-80GB for 24 hours, run the skill to get a ranked table of the cheapest providers plus a COMPUTE_PROCUREMENT.md saved into the research project. ## Quick Start Ask the assistant to use the research.compute.shop skill to find the cheapest provider for 1 A100-80GB GPU running 24 hours with a 5 dollar per hour budget.
