Hugging Face compute is the group of usage-billed services that run workloads on Hugging Face infrastructure: Spaces on upgraded hardware, ZeroGPU, Jobs, dedicated Inference Endpoints and Inference Providers. The Hub documentation says that “the billing for our compute services is usage-based, meaning you only pay for what you use”, that it is billed separately from PRO and Team or Enterprise subscriptions and from private storage, and that invoices are issued at the beginning of each month.[1] Several services have a free tier, and users and organizations pay to reach more powerful accelerators.[1]
Editions
Spaces hardware
A Space is an app repository. Static Spaces are free for everyone, and Gradio and Docker Spaces run on compute and require PRO for personal accounts or Team or Enterprise for organizations.[3] The default hardware, 10c7635620d95922987ca717ace11708, has 2 vCPU, 16 GB memory and 50 GB of non-persistent disk at no hourly cost.[2] Upgrades are priced per hour:
| Hardware | Memory | GPU memory | Hourly price |
|---|---|---|---|
| 580db460b0275ad898dcc3a4abade0fd (8 vCPU) | 32 GB | none | $0.03 |
| e3c0f0ec2b8c5fdca6cd933554a64690 | 15 GB | 16 GB | $0.40 |
| 1x Nvidia L4 | 30 GB | 24 GB | $0.80 |
| 1x Nvidia L40S | 62 GB | 48 GB | $1.80 |
| 957d0af25fea55d18e6da81b4fc1e7e2 | 46 GB | 24 GB | $1.50 |
| f1d281cbb1c15dc793bd84fb0c8f3d78 | 142 GB | 80 GB | $2.50 |
| d42a72aaa3f95c41a34e4b710604b530 | 1136 GB | 640 GB | $20.00 |
These rows are from the hardware tables, which also list T4 medium, 4x L4, 4x and 8x L40S, 2x and 4x A10G and 4x A100 flavors.[2] The documentation marks the H100 and 8x H100 flavors as removed in December 2025.[2] Protected visibility, where the source stays private but the running app is publicly reachable, is part of PRO or Team and Enterprise.[3] Dev Mode and custom domains for Spaces are listed for Team, Enterprise and Enterprise Plus.[9]
ZeroGPU
4465a95c93b4502e90a154ee8bb34b96 is shared infrastructure that dynamically allocates and releases NVIDIA RTX Pro 6000 Blackwell GPUs, with a large size (half a GPU, 48GB VRAM) and an xlarge size (a full GPU, 96GB).[4] It works only with the Gradio SDK.[4] Using existing ZeroGPU Spaces is free for all users, within a daily quota. Hosting is limited by account: free personal accounts in good standing (verified email, account older than 30 days) can host 2 ZeroGPU Spaces, PRO accounts up to 10, and organizations need Team or Enterprise to enable ZeroGPU Spaces for all members.[4]
Jobs
Jobs run a command in a Docker image on a chosen hardware flavor. They are available to any user or organization with a positive credit balance, and the monthly credits included with PRO, Team or Enterprise count towards that balance.[5] CPU flavors run from CPU Basic at $0.01 per hour to CPU Performance at $1.90 per hour, and GPU flavors from the T4 small at $0.40 per hour to 8x Nvidia H200 at $40.00 per hour.[5] Exposing ports adds $0.01 per hour per Job.[5] See f1720e9ff3665a4585ae6632c1eec58e for one example flavor.
Inference Endpoints
Inference Endpoints deploy a model on dedicated, autoscaling infrastructure. The documentation lists instances by provider: for example AWS intel-spr x1 (1 vCPU, 2 GB) at $0.033 per hour, Azure intel-xeon x1 at $0.060 and Google Cloud intel-spr x1 at $0.050 per hour, plus GPU, Inferentia and TPU instances.[6] The AWS NVIDIA A100 x1 is $2.5 per hour and the x8 is $20 per hour.[6] The pricing page shows partly different instance lists, for example H100, H200, B200 and RTX PRO 6000 on AWS, so a catalog or contract should record which page was used.[8] See 49668ceac1cc5947adddf5c076c14b8e and 505bf6e6a1d258e694255068e7da4748.
Inference Providers
Inference Providers give API access to models served by Hugging Face and third-party providers. Hugging Face “charges you the same rates as the provider, with no additional fees”, passing provider costs through.[7] The built-in hf-inference provider, formerly the serverless Inference API, charges compute time multiplied by the price of the underlying hardware, and as of July 2025 focuses mostly on CPU inference.[7] See e3f66e51da2653709665e6154098d5f5.
Metrics
- Spaces hardware hour: “billing on Spaces is based on hardware usage and is computed by the minute”, counted for every minute the Space runs on the requested hardware regardless of use.[2]
- ZeroGPU quota: a daily quota in minutes. The xlarge size consumes 2 times the quota of large, so a 45 second effective task consumes 90 seconds.[4]
- Jobs hardware hour: hours billed by the minute while the Job is Starting or Running.[5]
- Inference Endpoint instance hour: the cost is the instance hourly rate times the minimum-replica hours plus scale-up hours times additional replicas.[6]
- Inference Providers pass-through usage: provider-priced usage passed through.[7]
- Compute credits: dollar credits that offset all of the above.[1]
Counting / floors
Credits
PRO users receive $2.00 of monthly credits and Team or Enterprise organizations $2.00 per seat. Free users have none and must buy credits. The credits are general-purpose compute credits that can be spent on Inference Providers, Inference Endpoints, upgraded Spaces hardware (including ZeroGPU beyond the quota) and Jobs.[7] They are credited every month and applied automatically before pay-as-you-go usage is billed.[1] For Team and Enterprise organizations, credits are shared among members, and administrators can set a spending limit and disable some providers.[7] Enabling automatic recharge is recommended to avoid disruption once credits run out.[1]
What is billed and when
Spaces are billed only while Starting or Running, so there is no cost during build, and a running Space that starts failing is suspended automatically, which stops billing.[2] Free-hardware Spaces are suspended after an extended idle period such as two days, while upgraded Spaces run indefinitely by default even without usage, unless a custom sleep time is set.[2] Changing the hardware to CPU Basic or pausing the Space interrupts billing, and paused time is not billed.[2] Replicas are available only on upgraded hardware and each is billed independently.[2]
Inference Endpoints are accessible to accounts with an active subscription and a credit card on file. An account is charged for compute while a successfully deployed Endpoint is initializing and running, and the cost is calculated by the minute even though prices are shown by the hour.[6] Instance types may need a quota request. Paused Endpoints do not count against used quota, but Endpoints scaled to zero do until paused. PRO users and Enterprise Hub organizations can request higher quotas.[6]
ZeroGPU quota
Included daily ZeroGPU quota is 2 minutes unauthenticated, 5 minutes for free accounts, 40 minutes for PRO accounts and Team members, and 60 minutes for Enterprise members, resets exactly 24 hours after the first GPU use, and drives queue priority.[4] PRO, Team and Enterprise users can continue past the quota by consuming pre-paid credits at $1 per 10 minutes of GPU time.[4] The plan comparison lists the same extension rate.[9]
Attribution to an organization
For Inference Providers, each user keeps their own token and passes an organization name in the X-HF-Bill-To header or the bill_to client parameter so that requests are billed to the organization. Included credits apply only when the organization is specified explicitly.[7] Jobs are billed to the user namespace by default, or to an organization through the namespace option, and on Enterprise plans and above to a resource group.[5] Resource groups can also cap monthly compute spend in total or per product.[10] Usage by member, model and provider can be retrieved by API on Enterprise plans.[7]
Virtualization & partitioning
Not applicable as a licensing rule. The hardware tables describe vCPU, memory and GPU memory of each flavor, and a Space or Endpoint is billed for the flavor requested, not for the physical host.
Cloud / BYOL
Inference Endpoints are sold per cloud provider and instance type, and the rate differs by provider for similar sizes.[6] Organizations can be billed through an AWS or Google Cloud marketplace subscription, which must be approved by Hugging Face before usage is billed through it and which disables adding credits.[1] For Inference Providers a custom provider key sends the billing to the provider, Hugging Face does not charge for the call, and the included credits do not apply.[7] See Cloud marketplace billing and Custom Provider Key.
Programs
Compute pay-as-you-go and credits is the mechanism behind all of the above. Community GPU grants lets Space owners apply for help covering GPU upgrade costs from the hardware settings, and the page states no criteria or amounts.[2]
Out of scope
This article does not cover Hugging Face Storage Buckets billing beyond the per-terabyte storage tiers, AutoTrain, the Infinity container, third-party provider rate cards, or the licence terms of the models run on this compute. Choosing between a dedicated Endpoint, Jobs and a routed provider call is an engineering decision, and tracking idle Spaces or scaled-to-zero Endpoints is practitioner commentary.