Inference instances

Your own dedicated AI backend for the product you are building.

Run the models behind your product (chat, OCR, vision and embeddings) on a dedicated GPU with an OpenAI-compatible API. Open-weight models, a key per application, and you pay for GPU runtime, not tokens. Now in early access.

Join the early-access waiting list arrow_forward See what it costs for your load

A GPU you size, rent and call like an API

An inference instance is one dedicated GPU machine running one open-weight model behind an OpenAI-compatible endpoint: the AI backend of your product, without a per-token meter.

schedule

Pay for GPU time, not tokens

One hourly price, whatever the volume. For steady or bulk work that is a fraction of a per-token bill. For a few requests a day an API is cheaper, and the calculator below shows where the line is.

memory

Dedicated, not shared

The GPU runs your model for you alone. No other customer sends traffic to it.

lock_open

Open-weight models

Chat, vision, embedding and OCR models such as Qwen3, Gemma 4, gpt-oss and GLM-OCR, served with a standard OpenAI-compatible API.

tune

Sized from your load

Give your peak requests per minute and a typical request. We show the GPU memory it needs, the speed to expect, and whether a GPU carries the load.

public

Your choice of region and hosts

Pick a region and a hardware trust tier. Verified hosts are the default, and the wizard warns before you choose community hardware for sensitive documents.

speed

Limits you control

Give each application its own API key and cap it at a number of requests per minute. Let the instance pause when idle; the next request wakes it.

What you can build on it

Four building blocks of AI products, each sized by the unit it is naturally measured in.

chat_bubble

Assistants and chatbots

The assistant inside your product, a support bot or an internal tool, behind your own endpoint. Size it by chats per minute at peak.

document_scanner

OCR and document pipelines

Read invoices, forms and scans in your product with OCR models such as GLM-OCR and PaddleOCR-VL and get text and markdown back. Size it by pages per minute. Bulk pipelines are where GPU time beats per-page prices.

image

Computer vision

Let your product answer questions about images and screenshots with vision models such as Gemma 4 and Qwen3.8. Send up to 4 images per request.

scatter_plot

Search and retrieval

Qwen3 Embedding turns text into vectors for the search and retrieval in your product. Embed a large corpus in a batch run on one GPU, then let the instance pause.

Tokens or GPU time: pick by the shape of your load

Neither is always cheaper. The difference is how cost grows with volume.

Pay per token or per page

Best for light and spiky use

You pay for what you send and nothing while idle. Cost grows with every request, so bulk jobs get expensive quickly.

  • closeCheapest for a few thousand requests a month
  • closeNo capacity to size and nothing to run
  • closePer-token and per-page prices add up on bulk work

Pay for GPU time

Best for steady and bulk work

You pay an hourly price while the instance runs, whatever the volume. A busy hour costs the same as a quiet one, and you can pause the instance when idle.

  • checkCost per request falls as the load rises
  • checkA known cost per hour, with a price ceiling you set
  • checkIdle time costs money unless you pause

The calculator below uses your own numbers, including the load at which the API is cheaper.

What would it cost for your workload?

From idea to endpoint

1

Choose a model

Pick chat, vision, embedding or OCR. The picker shows what each model is for and how big it is.

2

Describe your load

Give a peak requests per minute and a typical request. We show the memory it needs and whether a GPU carries it.

3

Pick the hardware

Take the cheapest GPU that carries the load, or one with more headroom. Set a price ceiling and when the instance pauses.

4

Call your endpoint

Use the OpenAI-compatible API with a key per application. Watch requests per minute against your peak.

Your data, plainly

We do not store or log what you send. Here is what happens to a request, and what we do not claim.

  • We do not store or log prompts or completions. Our proxy and our control plane keep usage counts and timings, never contents.
  • Requests are relayed through our proxy on Cloudflare to your GPU. The proxy reads a request only to enforce size limits and to detect image content, and keeps nothing.
  • The GPU is dedicated to you. It is not shared with other customers.
  • GPUs come from GPU marketplaces and cloud providers. You choose the region and a hardware trust tier, and verified hosts are the default. Community hardware is less vetted: do not choose it for sensitive documents.
  • Models are open-weight and run on your instance, and we do not train on your data.
  • Availability is best effort. Rented GPUs can be interrupted. We restart the instance on a new host, and there is no SLA.

If you need dedicated nodes inside a specific country, see syndicAI Enterprise. The details are in the documentation.

Questions

Run your workload on your own GPU

Inference instances are in early access. Join the waiting list and we will let you in.

Join the early-access waiting list arrow_forward Read the docs