Skip to main content
huggingface.health
All guides
Hugging Face basics

Hugging Face Inference Endpoints pricing, explained

Inference Endpoints are per-second, per-instance. Here's how the SKUs price out and when scale-to-zero actually pays for itself.

Last updated July 18, 2026

What you're actually paying for

Hugging Face Inference Endpoints pricing is refreshingly simple on the surface: you rent a dedicated instance (or a small autoscaling pool) running your model on AWS, Azure, or GCP, and Hugging Face bills per second the instance is up at a fixed rate per SKU. There is no per-token charge stacked on top, no per-request fee, no surprise egress markup from the HF side. It is pure infrastructure time — the same billing model that made EC2 dominant, wrapped in a managed control plane that handles image builds, weight caching, TLS, autoscaling, and a stable inference URL. That simplicity is also the trap. Because you are paying for wall-clock time rather than tokens, an underused endpoint can burn a four-figure monthly bill while serving a handful of requests per hour.

Rough per-hour SKU costs (2026)

  • CPU (2 vCPU, 4GB) — ~$0.06/hr, fine for embeddings, small classifiers, reranking heads
  • CPU (8 vCPU, 16GB) — ~$0.24/hr, decent for BGE embeddings and small distilled models
  • T4 (16GB VRAM) — ~$0.60/hr, good for 3–7B quantized LLMs and vision models
  • L4 (24GB) / A10G (24GB) — ~$1.30/hr, sweet spot for 7–13B full-precision or 13B–34B quantized
  • A100 40GB — ~$4.00/hr, needed for 70B quantized or long-context 34B
  • A100 80GB — ~$5.50/hr, 70B FP16 fits with headroom
  • H100 80GB — ~$8–10/hr, for 70B FP16 throughput or bleeding-edge Flash Attention 3 setups

How scale-to-zero actually works

Scale-to-zero shuts the instance down after N minutes of idle traffic (configurable, default 15). On the next incoming request the endpoint provisions a new instance, pulls the container image from a warm cache, hydrates weights from HF's model store, and boots your inference server. For a 7B safetensors model on an L4, cold start is typically 30–90 seconds; for a 70B model on an H100 it is 3–8 minutes. During the cold start the caller receives HTTP 503 with an `estimated_time` hint or a hanging request depending on your client's retry policy. That latency is the price you pay for the cost savings, and it is almost always the deciding factor for whether scale-to-zero is safe.

When scale-to-zero pays off (and when it burns you)

The break-even math is straightforward. If your endpoint would run more than roughly 20% of the day at baseline, always-on is cheaper than eating the cold-start UX pain. For a nights-and-weekends internal tool, scale-to-zero can cut the bill by 60–80% and nobody minds a 30-second first-load. For a customer-facing chat UI where a p95 cold start of 90 seconds would visibly break the experience, always-on is table stakes and you should size for peak concurrency, not average. A middle ground that works well: keep min-replicas at 1 with scale-to-zero off during business hours, and let a scheduled job flip to scale-to-zero overnight.

Worked example: a 7B chat assistant

Say you deploy Llama-3.1-8B on an L4 at ~$1.30/hr. Always-on for a full month works out to 24 × 30 × 1.30 = $936. With scale-to-zero and a 15-minute idle timeout, six productive hours per day of real traffic plus roughly two hours of idle-then-decayed warmth costs closer to 8 × 30 × 1.30 = $312 — plus the UX tax of a 30–60 second cold start for the first user after each quiet period. Add a second replica for redundancy and you double the always-on figure. Add an A100 for a 13B model and you are north of $2,800/mo before egress. This is why serious teams compare Hugging Face Inference Endpoints pricing against a router provider like Together or Fireworks before committing.

Costs you'll forget to include

  • Egress: cloud-provider egress is billed separately when your app pulls responses across regions
  • Persistent storage: model weight caches for cold-start speedup cost extra per GB-month
  • Overprovisioning: min-replicas > 1 for HA doubles or triples baseline cost
  • Idle scale-out headroom: autoscaling policies with generous scale-up thresholds keep spare replicas warm
  • Observability: shipping logs and metrics to Datadog or Grafana Cloud is an out-of-band bill

How Endpoints pricing compares to Serverless and the router

Hugging Face Serverless Inference is free with rate limits (or Pro-boosted with a higher quota) and charges nothing per instance. The HF router (router.huggingface.co) is an OpenAI-compatible edge that fans out to hosted providers and passes through their per-token pricing with a small HF margin. Inference Endpoints only make financial sense once you have (a) enough steady traffic to keep an instance warm at meaningful utilization, or (b) a compliance/latency requirement that forces dedicated infrastructure. For prototyping or low-volume workloads, Serverless or the router almost always win on cost, and switching later is a URL and header swap.

Practical tips to keep the bill under control

  • Right-size on quantized weights first — a 13B model in AWQ int4 fits on an L4 and halves the SKU cost vs A10G
  • Turn on Prometheus metrics and alert on GPU utilization below 20% for 24 hours
  • Prefer autoscaling max-replicas as a hard ceiling; a runaway retry loop can 10x the bill overnight
  • Batch requests where possible — TGI and vLLM both continuous-batch, but only if concurrency is above 1
  • Reevaluate every quarter — HF adjusts SKU pricing and adds new instance types regularly

FAQ

Is Hugging Face Inference Endpoints pricing cheaper than OpenAI for a 7B model?
For steady, high-volume traffic on an open model, yes. A $250–1,000/month endpoint can replace an OpenAI bill that scales linearly with tokens. The crossover point is usually somewhere between 5 and 20 million tokens per day, depending on which SKU you land on and whether you can amortize the endpoint across multiple applications. For spiky, low-volume workloads, per-token APIs almost always win.
Do I need a Pro subscription to use Inference Endpoints?
No. Inference Endpoints is a separate paid product from Pro. Anyone with a valid payment method on their Hugging Face account can deploy an endpoint. Pro unlocks ZeroGPU on Spaces and higher Serverless quotas, but it has no bearing on Endpoints pricing or availability. Enterprise Hub customers get billing consolidation, SSO, and audit logs on top.
Can I bring my own cloud account for an Endpoint?
Not for the standard managed Endpoints product — you pick region and SKU and HF runs the underlying instance in their account. If you need BYOC (bring your own cloud) for compliance or committed-use-discount reasons, look at Inference Endpoints on your own infra via the Enterprise Hub. That is a sales-led offering and pricing is negotiated.
How do I minimize cold-start time?
Use safetensors weights (mmap-friendly), enable the HF weight cache on persistent storage, prefer smaller quantized models when acceptable, and pick a region physically close to the storage layer. Beyond those, the only reliable way to eliminate cold starts is to turn scale-to-zero off and pay for always-on capacity.
What happens if my endpoint's autoscaling maxes out during a spike?
Additional requests queue at the load balancer up to a timeout, then fail with 429 or 503. TGI and vLLM both handle continuous batching well, so a single replica can absorb surprising amounts of concurrency, but there is no automatic burst capacity beyond your configured max-replicas. Set max-replicas generously and alert on queue depth.
Are there discounts for annual commitments?
Standard Inference Endpoints pricing is pay-as-you-go with no long-term discount surface in the self-serve UI. Enterprise Hub customers can negotiate committed-use discounts, and HF occasionally runs private pricing for large accounts, but there is no public reserved-instance table like AWS EC2 offers.
Can I run multiple models on one endpoint?
No — each endpoint serves exactly one model repo. If you need to host several small models, you either deploy multiple endpoints or wrap them in a custom container yourself and route internally. The latter is more cost-efficient but sacrifices the managed autoscaling and one-model-per-URL simplicity that most teams pick Endpoints for.
How does Endpoints pricing compare to Replicate or Modal?
Replicate bills per-second of GPU time per-run and cold starts are typically shorter for their curated catalog, which favors spiky workloads. Modal is closer to Endpoints in shape but charges per-second on any container with more flexible custom code. For steady dedicated inference, HF Endpoints is usually the cheapest of the three; for burst inference on curated models, Replicate often wins.

Related on this site