Self hosting an open model: what you pay for and what the host charges

When you self host an open model, the people who publish the weights usually charge nothing for the download, and you pay for everything needed to run it: hardware time, storage, operations and review. This post uses only public price pages, fetched on 18 October 2026, and makes no claims about any provider's internal costs.

Who are the parties?

Three roles are easy to blur.

• The model publisher. Releases weights and a model card. The card typically names the license, intended use and limits. Publishing is not the same as running.
• The platform or cloud. Provides the machines, and bills by the hour or by usage.
• You. Operate the service, secure it, monitor it and answer for its output.

"Free model" describes the first role only.

What do public price pages show?

We fetched Hugging Face's pricing page on 18 October 2026. Prices change, so treat these as a dated snapshot.

• Dedicated inference endpoints are listed as "starting at $0.033/hour".
• GPU examples on the page include $0.50 per hour for one NVIDIA T4 on AWS and $4.50 per hour for one NVIDIA H100 on AWS.
• Spaces hardware is listed from free for CPU Basic up to $22.00 per hour for 8x Nvidia RTX PRO 6000, with examples such as $1.00 per hour for one Nvidia A10G small and $2.50 per hour for an Nvidia A100 large.
• Storage for public repositories starts at $12 per terabyte per month, and private repositories start at $18 per terabyte per month.
• A PRO account is listed at $9 per month, Team at $20 per user per month and Enterprise at $50 per user per month.

These figures describe what a platform charges you. They do not tell you what any host pays for its own hardware, and we do not guess.

What does that mean per month?

Simple arithmetic shows why always on hosting adds up. A machine at $4.50 an hour that runs all day costs $108 for 24 hours, and about $3,240 over 30 days. A machine at $0.50 an hour costs $12 for 24 hours, and about $360 over 30 days. Those are calculations from the published hourly rates, not quotes. Scale to zero or schedule your endpoints, if your platform supports it, to pay for use instead.

What costs are on you besides the machine?

1. Storage and transfer. Weights are large, and you may keep several versions.
2. Operations. Patching, scaling, health checks and on call time.
3. Evaluation. Testing the model against your own tasks before and after each change.
4. Security. Access control, logging, input and output handling, and keeping secrets out of prompts.
5. License review. Read the license text. The Hugging Face model card documentation says the license is part of the model card metadata, and a custom license can be given with a name and link. Conditions vary, so read each one.
6. Supply chain. The runtime, libraries and the weights themselves are artifacts you should inventory and scan.

When does self hosting make sense?

Usually for one of four reasons: data cannot leave your environment, you need a specific model or fine tune, latency or throughput needs are predictable, or you want control over change. If none applies, a hosted API may cost less in total once you count staff time.

A decision checklist

1. Write the monthly request volume and the latency you need.
2. Pick two candidate models and read their cards and licenses.
3. Price one realistic always on setup and one scheduled setup with the platform's published rates.
4. Add storage and an honest estimate of operations time.
5. Compare against a hosted alternative's published prices, fetched the same day.
6. Decide who owns security review of the model and the runtime.
7. Revisit in a quarter.

Where does Safeguard fit?

Safeguard supports bring your own model, including on premises and local models, as well as a Zero AI mode with no model calls. If you run an open model, you can connect it to Safeguard that way. What you pay for hosting that model is a matter between you and your provider.

Quick check

Sources

• Hugging Face pricing, fetched 18 October 2026
• Hugging Face docs: model cards

Last reviewed: 18 October 2026.