RAW

RAW is a bare-metal cloud platform built specifically for AI inference, training, and agent workloads. Founded in 2025 by Michael Jakob — who previously built Eulerpool, a financial data platform — RAW grew out of a real $7,000 monthly cloud bill and the frustration of AWS complexity. The platform offers dedicated NVIDIA GPU and AMD EPYC CPU servers across five global regions, provisioned via a single REST API in as little as 3 seconds.

What It Is

RAW is a server-creation and fleet-scaling API designed to replace AWS, DigitalOcean, and GPU clouds for AI teams. Instead of navigating a 200-service console, users POST to a single endpoint to get a running server with full root SSH, CUDA, NVMe storage, and a public IP. The platform targets LLM inference, fine-tuning, AI agent hosting, and vector database deployments. RAW positions itself as "the anti-AWS console" — one API, flat pricing, and zero egress fees.

How the API Works

The entire product surface is a REST API with 24 endpoints. The workflow is intentionally minimal:

• Sign up with a single POST to get a Bearer token — no charge
• Add a card via /billing/setup (Stripe-backed, $0 until deploy)
• Deploy with POST /deploy passing type and region — server is live in ~13–30 seconds for CPU, 1–48 hours for GPU (dedicated hardware reservation)
• Scale by repeating the same call; destroy with DELETE /servers/:name

A CLI (npm install -g rawhq) and a web dashboard wrap the same endpoints. The dashboard provides 14 management tabs per server: metrics, networking, firewalls, volumes, backups, snapshots, rescale, rebuild, power, rescue, terminal, and an AI prompt helper for Claude/ChatGPT.

Infrastructure and Regions

RAW runs on dedicated hardware — no hypervisors, no shared tenancy — in Tier-3 datacenters across five regions:

• 🇩🇪 Frankfurt, Germany (eu)
• 🇮🇪 Dublin, Ireland (eu-fi)
• 🇺🇸 Ashburn, Virginia (us)
• 🇺🇸 Hillsboro, Oregon (us-west)
• 🇸🇬 Singapore (sg)

CPU servers use AMD EPYC and Intel Xeon processors with NVMe SSDs. GPU servers use NVIDIA A100, L40S, and RTX cards with full CUDA. GPU SKUs are currently EU-only. All servers include unlimited bandwidth with no egress metering.

AI Stack Support

RAW is designed to run the full AI stack on dedicated hardware:

• LLM inference: vLLM, Ollama, TGI with OpenAI-compatible endpoints; supports Llama, Mistral, DeepSeek
• Training and fine-tuning: Full CUDA access, up to 96 GB VRAM, NVMe for checkpoints; LoRA or full training
• AI agents: Persistent, always-on CPU servers with root access; scale workers via API from CI
• Vector databases: Qdrant, pgvector, Milvus on dedicated NVMe
• Pre-install apps: Docker, PostgreSQL, Redis, Node.js, Nginx, Caddy, Python, Go, Bun, MySQL, OpenClaw

Deployment Model and Company Background

RAW is bootstrapped, headquartered in Munich, Germany, and describes itself as "AI-native" — product development, infrastructure automation, and operations are run primarily by AI systems rather than a large human team. The founder, Michael Jakob, migrated Eulerpool's entire stack from managed cloud to bare metal before building RAW as a product. The platform holds SOC 2 Type II certification and offers GDPR-compliant EU GPU regions. Per-second billing is available, and the 99.9% uptime SLA applies across all regions.