Deploy custom AI models, ComfyUI workflows, LoRAs, and safetensors as autoscaling API endpoints. MCP-compatible. 10-20x cheaper than alternatives on bare-metal GPUs.