Prices pulled from vendor pages. Not a quote.

You probably don’t need to rent a supercomputer.

Most people starting with AI should pay for a chat app, or a cheap always-on box if they want an agent. Data-center chips are for training huge models. This page tells you which door is yours, and what it costs this week.

Start here

Just use AI

About $0–$20 / month

ChatGPT, Claude, Gemini, Grok. You type, it answers. This is the right door for writing, learning, and trying things. No server. No chip rental.

If you want an agent that stays on

A small VPS for Hermes or OpenClaw

About $5–$20 / month

The program lives on a cheap always-on computer. The thinking happens in Grok or another chat API. You do not need a GPU for this. 2 CPUs and 4 GB RAM is a normal start.

If you need to run a model yourself

A GPU VPS (a 4090 / L40S class box)

About $0.40–$2 / hour

Fine-tunes, local models, tinkering. Still not a data-center training chip. Example from our 11:35am PT pull: RunPod A40 $0.44/hr, L40S $0.99/hr.

Last door

H100 / H200 / B200 by the hour

A few dollars per GPU-hour

This is a data-center AI chip. You rent it to train or serve very large models. It is a semi, not a sedan. If you had to ask what an H100 is, you are not in this door yet.

Chip	Vendor	$ / GPU-hr