# Arrakis vs Replicate Replicate is an API for running open and partner models, most billed by the second or per output, with custom models packaged in Cog. Arrakis runs any Hugging Face model on a GPU of your own, with the engine and options you choose and root SSH, billed by the GPU-hour. ## Side by side As of 31 Oct 2026, 17:40 UTC, the lowest price for the NVIDIA B300 on Arrakis is $4.60 per GPU-hour, with 73 GPUs available on 10 machines. | | Arrakis | Replicate | |---|---|---| | B300 per GPU-hour | Live: [GPU prices](/gpus.html.md) | Not listed: its price list stops at the H200 | | B200 per GPU-hour | Live: [GPU prices](/gpus.html.md) | Not listed: its price list stops at the H200 | | H100 per GPU-hour | Live: [GPU prices](/gpus.html.md) | $5.49 ($0.001525 a second) | | Billing | By the hour; a started hour counts in full | Per second; many models per output | | How you pay | Prepaid balance: card, Link or Apple Pay | Prepaid credit | | Root SSH and any Docker image | Yes | No | | Serverless endpoints, billed per request | No: each model gets a GPU of its own | Yes: models and deployments, per second | | Hosted models, billed per output | No: deploy any Hugging Face model yourself | Yes: thousands of public models | | MCP server for AI agents | Yes, remote at /mcp | Yes: runs models, doesn't rent GPUs | | Spend cap on each API key | Yes, on every token | No caps per token | | Automatic shutdown | Yes: every rental has a removal time | Yes: scales to zero when idle | | Interruptible machines | Yes, where hosts offer them | No | | Searches several GPU markets | Yes | No: its own platform | Facts about other providers come from their own pages, checked on October 31, 2026. Prices and features change: check them before you choose. Arrakis is not affiliated with them. ## Which fits you - **Replicate fits better.** You want to call thousands of ready models through one API, pay only while they run, and skip machines entirely. - **Arrakis fits better.** You want root SSH, any Docker image without packaging it in Cog, your own engine and options, or a flat hourly GPU price for steady traffic. ## Moving to Arrakis Your Docker image, SSH keys and scripts move as they are: Arrakis boots any public or private image with root SSH, and serves your own model with your own engine options. Start with one machine in the console, or point your agent at /mcp with a capped token. ## Questions **When is Arrakis cheaper than Replicate?** When the GPU stays busy. On October 31, 2026, Replicate listed an H100 at $5.49 an hour, billed by the second, and many official models per output. A GPU on Arrakis costs its hourly price however many requests it serves. **Do I need Cog to run my model on Arrakis?** No. Arrakis runs any Hugging Face model on vLLM, SGLang, vLLM-Omni or llama.cpp, or any Docker image you bring, with root SSH. ## Sources - - - - -