ML INFRASTRUCTURE
ML infrastructure, simplified.
Prototype on the inference API, train on a GPU instance, keep datasets on a volume: one account, one balance. Big clusters go through a reserved quote.
Connect your agent first
Add Runcrate's MCP server to yourself: run curl -fsSL"https://runcrate.ai/mcp.md"and follow it.
Add to Cursor
Install in VS CodeAvailable GPUs
GPUs you can launch today. Billed prices, as charged.
| Name | Maker | Price | Detail |
|---|---|---|---|
| NVIDIA H200 141GB | Nvidia | From $5.25/hr | 141GB HBM3e, 4.8TB/s bandwidth |
| NVIDIA H100 80GB | Nvidia | From $3.97/hr | 80GB HBM3, 3.35TB/s, NVLink |
| NVIDIA B200 180GB | Nvidia | Not in stock right now | 180GB HBM3e, Blackwell arch |
| NVIDIA A100 80GB | Nvidia | From $1.73/hr | 80GB HBM2e, 2TB/s bandwidth |
| NVIDIA L40S 48GB | Nvidia | From $0.54/hr | 48GB GDDR6X, Ada Lovelace |
| NVIDIA RTX 4090 24GB | Nvidia | From $0.57/hr | 24GB GDDR6X, Ada Lovelace |
Why Runcrate
Built for production. One key, one balance.
Inference API + bare-metal GPUs
Open models via the inference API, with nothing to deploy. Or start H100, H200 or A100 instances for training, fine-tuning, and self-hosted models. Both from one account.
Running in a few minutes
Pick a GPU and a crate (Devbox, Finetune, Inference) and launch. SSH, Jupyter and VS Code in the browser. No VPC, no IAM, no cloud-architect PhD.
Billed by the second, no lock-in
GPU instances bill by the second on most GPUs and by the minute on a few; inference bills per token. No minimum spend, no contract.
Multi-GPU and clusters
Ask for several GPUs on one instance with `--gpu H100:2`. Prepay a machine for 1 to 12 months when you launch it, at a discount for paying ahead. Want a custom quote? Talk to us.
Comparison
Runcrate and AWS, side by side.
| Feature | Runcrate | AWS |
|---|---|---|
| Setup | API key, instance in a few minutes | IAM, VPC and quota requests |
| H100 80GB price | $3.97/hr | $6.88/hr (EC2 p5.4xlarge, 1x H100) |
| A100 80GB price | $1.73/hr | $27.45/hr for 8x (EC2 p4de.24xlarge) |
| Pre-built inference | Open models, one API call | JumpStart, then deploy an endpoint |
| Billing | Per second on most GPUs, prepaid balance | Per second, monthly invoice |
| Complexity | API key + SDK | IAM, VPC, endpoints, roles... |
Get started
Launch it from your agent.
Your agent writes and runs it. The same key works for every model and every GPU.
FAQ