Hugging Face Inference

Hugging Face sells two ways to run a model. Inference Endpoints give you dedicated hardware for one model at an hourly rate, with an A100 80 GB at $2.50 an hour on AWS. Inference Providers route requests to partner hosts with no markup from Hugging Face. On 31 October 2026 PRO cost $9 a month, Team $20 per user and Enterprise $50 per user.

Some links on this page go through /go/ and may earn us a fee; it never changes our order.

Facts at a glance

To put it beside GPU clouds and per-token APIs, use the interactive hosting compare and the AI hosting overview.

Two products with one name

People say "Hugging Face Inference" and mean different things, so separate them.

Inference Endpoints (dedicated). You choose a model from the Hub, pick a cloud, a region and an instance type, and Hugging Face deploys it on dedicated hardware with an autoscaling policy. The docs say endpoints are accessible to Hugging Face accounts with an active subscription and a credit card on file, and that you are charged for the compute while deployed endpoints are initialising and running.

Inference Providers (serverless router). A single API, integrated into the Python and JavaScript clients, that sends your request to one of several partner inference providers. The docs say it gives access to hundreds of models, and the pricing page says Hugging Face adds no markup: the price you pay matches the provider's. The docs list providers including Baseten, Cerebras, Cohere, DeepInfra, Fal AI, Featherless AI, Fireworks, Groq, Novita, Nscale, OVHcloud AI Endpoints, Public AI, Replicate, Scaleway, Together, WaveSpeedAI and Z.ai; which models each serves varies, so check the model page.

The first resembles renting a managed GPU. The second resembles a gateway such as those in OpenRouter alternatives.

Plans and credits

From huggingface.co/pricing on 31 October 2026:

The Providers docs say PRO, Team and Enterprise get monthly credits, that credits for Team and Enterprise organisations are shared among members, and that credits apply automatically before any pay-as-you-go usage is billed. Free users can use Providers by buying credits.

For a team of ten, Team is $200 a month and Enterprise $500, before any inference usage.

Inference Endpoints prices

Hourly rates from the Endpoints pricing page, checked on 31 October 2026. The page says prices are shown by the hour but cost is calculated by the minute.

CPU instances start far lower, for example an Intel Sapphire Rapids x1 at $0.033 an hour on AWS. The same page lists rows on Azure. The big gap between AWS and GCP for the same card, A100 80 GB at $2.50 against $3.60, is worth noticing.

The Spaces hardware table at the pricing page lists similar cards, such as an L4 at $0.80 an hour and an A100 large at $2.50 an hour, and a free ZeroGPU option for Spaces on a shared RTX Pro 6000 Blackwell with up to 96 GB.

What you pay in practice

The volumes are examples we chose.

Example one: a 120B-class model for a team, 220 hours a month. An AWS A100 80 GB at $2.50 for 220 hours is $550. Left on for 720 hours it is $1,800. Compare a Runpod Secure A100 at $1.79: $394 for 220 hours and about $1,289 for 720. Hugging Face costs about 40 percent more per hour, and in return manages deployment, autoscaling and monitoring for you.

Example two: a small model on one L4. An 8B-class model at 4-bit fits in 24 GB. $0.80 an hour for 720 hours is $576 a month. For comparison, Cloudflare Workers AI lists Llama 3.1 8B fp8 fast at $0.045 and $0.384 per million tokens, so 40 million input and 8 million output tokens cost about $4.87. A dedicated endpoint is therefore a poor choice at light volume.

Example three: autoscaling. The pricing page works its own example: a minimum of 1 replica and a maximum of 3, where a traffic spike every hour scales the endpoint from 1 to 3 replicas for 15 minutes. Cost is the hourly rate times hours at the minimum, plus the scale-up hours times the extra replicas. For an A100 80 GB at $2.50, 720 hours at one replica is $1,800, and in a 30-day month there are 720 spikes, each adding 2 replicas for a quarter-hour, or 360 replica-hours, about $900 more.

Example four: Providers pay-as-you-go. Prices match the underlying provider, so a call to a model on Groq costs what Groq charges. Groq lists gpt-oss-120b at $0.15 and $0.60 per million tokens. Hugging Face adds nothing on top, per its pricing page, and the $2 of PRO credits would pay for about 13 million input tokens at that rate.

Autoscaling and cold starts

The autoscaling docs describe a CPU-based and a GPU-based rule: a new replica is added when average utilisation across replicas reaches 80 percent, with GPU utilisation measured over a one-minute window, and a request-based trigger when more than 1.5 pending requests per replica build up over 20 seconds. Scale-up checks run every minute and scale-down every two minutes, with a stabilisation period of 300 seconds after scaling down.

Scale-to-zero is available and saves money while idle, at the cost of a cold start when a new request arrives. The docs warn that while the replica initialises, the HTTP server returns an error status, and that there is currently no request queue, so they recommend building your own queue client-side with error handling. Replicas download and load the model on start, so a model with a long initialisation may scale poorly. Also note that a scaled-to-zero endpoint still counts against quota until you pause it.

Data and security

The security page says:

• Hugging Face does not store payloads or tokens passed to an Inference Endpoint. It keeps logs for 30 days.
• Traffic uses TLS and SSL.
• AWS PrivateLink is recommended for organisations, giving a private connection that does not expose the endpoint to the internet.
• Public endpoints need no authentication, protected endpoints need a valid Hugging Face token, and a private option exists. Pick protected or private for anything not public.
• The Hub and Inference Endpoints are SOC 2 Type 2 certified, with role-based access control.
• A GDPR data processing agreement is available through an Enterprise Hub subscription.
• Private model repositories are not exposed publicly, and Hugging Face does not own your models or data. It scans repository contents for malware and pickle issues.

For Inference Providers, the data handling depends on the partner provider. Read each provider's terms. We did not read how Hugging Face's own terms combine with each partner's retention terms for routed requests.

Limits

Endpoints or Providers

The two products answer different questions, and the choice is usually quick once you know your volume and your model.

A path many teams follow: start on Providers to test models cheaply with the $2 PRO credit or pay-as-you-go, settle on one model, and move to a dedicated Endpoint when volume is high or the data needs tighter handling. Moving between them is mostly a change of client configuration, since both are reachable from the same Hugging Face client libraries.

Sizing an endpoint

Use the model's listed size and add room for the cache. From Ollama's library pages, gpt-oss 20B is listed at 14 GB, Llama 3.3 70B at 43 GB and gpt-oss 120B at 65 GB. Map that to the table above.

For the 120B class, the H200 at $5.00 gives roughly twice the A100's cache headroom, so it can serve more concurrent users or longer contexts. Whether that is worth double the price depends on your concurrency, which you should measure first. Our GPU requirements for LLMs page covers the arithmetic.

Cost habits for endpoints

1. Set a maximum replica count. It is your cap on a traffic spike.
2. Scale to zero for non-production. Dev and staging endpoints should sleep.
3. Pause, not just idle. The docs say scaled-to-zero endpoints still count against quota, so pause the ones you will not use for a while.
4. Watch the minute billing. Cost is calculated by the minute, so short tests are cheap and forgotten endpoints are not.
5. Compare with a pod each quarter. If an endpoint runs flat out around the clock, the same card on a GPU cloud may cost less. The trade is managed deployment against price.

Who should pick Hugging Face

Alternatives

• Runpod: the same A100 80 GB at $1.79 on Secure Cloud.
• Lambda Cloud: plain GPU instances, no egress fees.
• Modal: per-second billing, code-defined.
• Replicate: per-second public models.
• Cloudflare Workers AI: per-token open models.
• AWS Bedrock: managed models inside AWS.

Getting started

1. Open a model on the Hub and choose Deploy, then Inference Endpoints.
2. Pick a cloud, a region and an instance with enough GPU memory. Use the listed model size plus a quarter for cache; see GPU requirements for LLMs.
3. Choose protected access, set autoscaling bounds, and decide whether to allow scale to zero.
4. Test with a few requests and watch the minute-by-minute cost.
5. Pause or delete the endpoint when you finish.

Our take

Hugging Face is the easy route when your model lives on the Hub and you want it hosted without learning a GPU cloud. You pay for that ease: an A100 80 GB at $2.50 an hour on AWS is above Runpod's $1.79 on Secure Cloud. The no-markup Providers router is a fair deal for trying many hosts through one API. The security page is detailed, with SOC 2 Type 2 and PrivateLink, which matters for enterprise buyers. Our conclusions rest on Hugging Face's pricing and documentation pages. We have not deployed an endpoint for this page.

Related pages

• Hosting hub and AI hosting overview
• Private AI for teams
• vLLM