Managed inference and serverless GPU
Call a model or run a function and pay per token, per call or per second, without renting a machine.
Tools / Buying compute
Which compute offer type fits my work?
Describe the work and how you want to access it. The chooser narrows ComputeFinder's 12 offer types and shows the provider and programme records mapped to the result.
Before choosing a machine, start a model memory brief from explicit weight-storage assumptions. It does not estimate total runtime memory or hardware fit.
A route into research, not a recommendation. This does not rank providers or check prices, suitability or live availability. It uses fixed editorial mappings to the offer types already explained on this site. Read the contract, location, security and exit questions before choosing a supplier.
Offer types to compare
These routes fit run a model for users. They follow the rules shown below. Their order is the order used across ComputeFinder, not a ranking.
GPU clouds and bare metal
Clouds built mainly around accelerators, from single instances to dedicated clusters and whole servers.
Hyperscale cloud instances
Virtual machines and bare metal from very large general-purpose clouds, with accelerators beside many other services.
Regional and sovereign clouds
Clouds that keep data and operations in a chosen country or bloc, under its law.
Entries to investigate
The matching offer types currently map to these existing ComputeFinder records. Entries are deduplicated and alphabetical. Programme entries show their recorded region, deadline evidence and checked date before you open the full eligibility and application record. Programme mappings are application routes, not confirmation of free or funded access. Charges and conditions may apply. This is a research list, not an availability check, recommendation or ranking.
Providers (55)
- Alibaba Cloud
Mapped through Hyperscale cloud instances - Amazon Web Services (EC2)
Mapped through Hyperscale cloud instances - Baseten
Mapped through Managed inference and serverless GPU - Cerebras
Mapped through Managed inference and serverless GPU - Cirrascale Cloud Services
Mapped through GPU clouds and bare metal - Civo
Mapped through GPU clouds and bare metal, Regional and sovereign clouds - Cloudflare Workers AI
Mapped through Managed inference and serverless GPU - Core42
Mapped through Regional and sovereign clouds - CoreWeave
Mapped through GPU clouds and bare metal - Crusoe Cloud
Mapped through GPU clouds and bare metal - Cudo Compute
Mapped through GPU clouds and bare metal - Denvr Dataworks
Mapped through GPU clouds and bare metal, Regional and sovereign clouds - DigitalOcean GPU Droplets
Mapped through GPU clouds and bare metal - E2E Networks
Mapped through GPU clouds and bare metal, Regional and sovereign clouds - Exoscale
Mapped through Regional and sovereign clouds - Fireworks AI
Mapped through Managed inference and serverless GPU - Firmus Technologies
Mapped through GPU clouds and bare metal - Gcore
Mapped through GPU clouds and bare metal - Genesis Cloud
Mapped through GPU clouds and bare metal - Google Cloud (Compute Engine)
Mapped through Hyperscale cloud instances - Groq (GroqCloud)
Mapped through Managed inference and serverless GPU - Hetzner
Mapped through GPU clouds and bare metal - Hot Aisle
Mapped through GPU clouds and bare metal - Huawei Cloud
Mapped through Hyperscale cloud instances - Hugging Face Inference Providers
Mapped through Managed inference and serverless GPU - HUMAIN
Mapped through Regional and sovereign clouds - Hyperstack (NexGen Cloud)
Mapped through GPU clouds and bare metal - IBM Cloud
Mapped through Hyperscale cloud instances - Jarvislabs
Mapped through GPU clouds and bare metal - Koyeb
Mapped through Managed inference and serverless GPU - Lambda
Mapped through GPU clouds and bare metal - Latitude.sh
Mapped through GPU clouds and bare metal - Massed Compute
Mapped through GPU clouds and bare metal - Microsoft Azure
Mapped through Hyperscale cloud instances - Mistral Compute
Mapped through GPU clouds and bare metal, Regional and sovereign clouds - Modal
Mapped through Managed inference and serverless GPU - NAVER Cloud Platform
Mapped through Hyperscale cloud instances, Regional and sovereign clouds - Nebius
Mapped through Managed inference and serverless GPU, GPU clouds and bare metal - Nscale
Mapped through GPU clouds and bare metal - Oracle Cloud Infrastructure
Mapped through GPU clouds and bare metal, Hyperscale cloud instances - OVHcloud Public Cloud GPU
Mapped through GPU clouds and bare metal - Replicate
Mapped through Managed inference and serverless GPU - RunPod
Mapped through Managed inference and serverless GPU - Sakura Internet
Mapped through Regional and sovereign clouds - SambaNova
Mapped through Managed inference and serverless GPU - Scaleway
Mapped through GPU clouds and bare metal, Regional and sovereign clouds - Sesterce
Mapped through GPU clouds and bare metal - STACKIT
Mapped through Regional and sovereign clouds - Taiga Cloud (Northern Data Group)
Mapped through GPU clouds and bare metal, Regional and sovereign clouds - TensorWave
Mapped through GPU clouds and bare metal - Together AI
Mapped through Managed inference and serverless GPU - Verda
Mapped through Managed inference and serverless GPU, GPU clouds and bare metal, Regional and sovereign clouds - Voltage Park
Mapped through GPU clouds and bare metal - Vultr
Mapped through GPU clouds and bare metal - Yotta (Shakti Cloud)
Mapped through Regional and sovereign clouds
Matched programme records (1)
- Hugging Face ZeroGPU Global Free tiers
No future deadline recorded.
Recorded access conditions: Shared GPU time for compatible Spaces, subject to usage quotas and queues. This is not dedicated or unlimited compute.
Checked 13 November 2026. Mapped through Managed inference and serverless GPU
Link to this result
https://computefinder.net/tools/offer-type-chooser?goal=inference&route=any
How the chooser works
Each workload maps to four of the site's offer types. Your access preference filters that set. Direct overlaps appear first. Workload routes that do not fit the access preference appear separately, so an empty overlap is visible rather than silently replaced. The provider and programme list is a deduplicated alphabetical view of the mappings already published on those matching offer pages. The optional programme filters use the exact recorded region, the presence or absence of deadline text and this directory's programme type. They narrow recorded evidence and classification, not eligibility, award method or availability. Price, capacity and live availability do not affect the result.
Compare all 12 ways to buy compute, use the side-by-side comparison, or take the buyer's checklist into a supplier call.