Runcrate is an AI cloud platform that provides dedicated GPU compute and inference APIs, letting teams deploy models by the token or run raw bare-metal instances billed per second. It offers access to 200+ open-source models across a multi-cloud network of 10K+ GPUs, with cold-start boot times under 60 seconds and 2–3× more tokens per GPU than aggregators. The platform supports training, fine-tuning, RAG pipelines, and AI agents through a unified credit-based system.
Funding
Funding not disclosed
Founders
Product
Problem
GPU access for AI workloads is often expensive, fragmented across multiple providers, and operationally complex, requiring significant DevOps overhead. Teams face long deployment times, rigid contract commitments, and opaque pricing that slows experimentation and production scaling.
Solution
Runcrate provides a unified AI cloud platform that combines dedicated GPU instances with a high-performance inference engine, all billed by the second. The platform's Arc-powered inference engine delivers 2–3× more tokens per GPU, while raw bare-metal instances offer full root access with 60-second cold-start deployment. Users can scale from 1 to 128+ nodes across a multi-cloud network of 10K+ GPUs spanning 8 global regions, with per-minute billing and no long-term commitments. The platform includes pre-configured environments for distributed training, fine-tuning, RAG pipelines, and AI agents, all managed through a single credit balance with an OpenAI-compatible endpoint.
Target Audience
Primary customers are AI engineering teams, machine learning researchers, and enterprises needing flexible GPU compute for model training, fine-tuning, inference, and AI agent deployment without infrastructure lock-in.
Features
- Arc-powered inference engine delivering 2–3× more tokens per GPU with 80 TPS throughput and 247ms P50 latency
- Raw bare-metal GPU instances (H100, H200, B200, B300, A100, L40S) with full root access, SSH, Docker, and custom images
- 60-second cold-start boot and per-minute billing with auto-scaling from 1 to 128 nodes
- 200+ open-source models via one API endpoint, including Claude, DeepSeek, Llama, Whisper, FLUX, and Sora
- Pre-configured CUDA/PyTorch environments for fine-tuning (LoRA, QLoRA, Axolotl) and distributed training (DeepSpeed, FSDP, Megatron-LM)
- Unified credit balance with auto-recharge, zero transfer fees, and public rate cards requiring no negotiations
- SDKs for Python and TypeScript, plus CLI and MCP support for autonomous AI agent compute management