OpenInfer
Sign in Get access

OpenInfer Loom

Turn any
GPU
into a token business.

* If it has a GPU, CPU or NPU, it can join the fleet. Results on household appliances may vary.

Loom pairs GPUs from any cloud or rack, packs your models onto them, routes every request, and serves it all as an OpenAI- and Anthropic-compatible API you can sell.

3.9×

more sellable tokens on the same GPUs

Get access Pair your first GPU →

Agent first

Setup is a prompt.

Paste one line into your coding agent. It pairs your GPU, loads a model and hands you an API key.

Set up Loom for me: https://platform.openinfer.io/agent-setup.md
  1. ✓ installed weave CLI
  2. ✓ signed in · org acme
  3. ✓ paired gpu-01 · NVIDIA
  4. ✓ loaded chat-medium · routable
  5. ✓ minted key oi_••••3f9a
  6. endpoint ready → https://api.acme.ai/v1
Run it

Bring your own GPU

Pair a GPU in three commands.

Any machine with a GPU joins your fleet in minutes. No cluster to build first.

How pairing works →
  1. weave pair weave_pt_…
  2. weave runtime install
  3. weave up

your-gpu-box

node in your fleet

pairing…online

Heterogeneous

AWS, GCP, on-prem. NVIDIA, AMD. One pool.

You take GPU capacity wherever you can get it, and every cloud and vendor wants its own stack. Loom runs one stack on all of it, so no GPU sits stranded for being the wrong brand or in the wrong cloud.

one pool AWS cloud GCP cloud NVIDIA GPUs AMD GPUs on-prem your racks
● In the pool now:
  • NVIDIA GeForce RTX 5070
  • NVIDIA RTX PRO 6000 Blackwell Server Edition

Planning & placement

Say what to serve. It works out where.

Pick your models and how much each one matters. Loom lays them out across your GPUs and shows you the plan before anything moves.

What to serve

  • chat-large
  • vision
  • coder
  • chat-small

Plan preview

applied
  1. chat-large
    chat-small

    box 1

  2. chat-large
    coder

    box 2

  3. vision
    chat-small

    box 3

  4. vision
    coder
    chat-small

    box 4

Routing & rebalancing

Every request goes to the GPU that can take it.

No hot spots next to idle cards. Traffic spreads across every GPU serving a model, and when demand shifts, Loom proposes a better layout — you approve the move.

/v1 openai · anthropic aws-01 aws-01 · offline aws-02 aws-02 · offline gcp-01 gcp-01 · offline rack-07a rack-07a · offline rack-07b rack-07b · offline
planner · move chat-medium to rack-07b ApproveApproved ✓
Serve

API gateway

Speaks OpenAI. Speaks Anthropic.

Point the SDKs you already use — or Claude Code — at your endpoint. Change nothing else. 

$ curl https://api.acme.ai/v1/chat/completions \  -H "Authorization: Bearer $LOOM_API_KEY" \  -d '{    "model": "@acme/chat-large",    "stream": true,    "messages": [{"role": "user", "content": "Hi"}]  }'

Hello! Served from your own fleet.

Multi-tenant

One fleet. Many tenants. No crosstalk.

Every tenant gets its own keys, models and usage. One tenant's spike never becomes another tenant's outage.

  • acme oi_••3f9a 200 OK
  • globex oi_••81c2 200 OK
  • initech oi_••d047 200 OK
  • hooli oi_••5e6b 200 OK
Sell it

Cloud layer

Open a token storefront on your own GPUs.

Put a customer console on top of your fleet: sign-up, prepaid billing, metering, per-model pricing, your own API domain. Your customers buy tokens. Your GPUs serve them.

We run OpenInfer Cloud on it.

Acme.ai Console example

Token revenue · this month

$0.00

  • @acme/chat-large $0.40 / 1M out 0 tok
  • @acme/vision $0.30 / 1M out 0 tok
  • @acme/coder $0.22 / 1M out 0 tok
An example customer console with illustrative prices and token counts.

Everything in Loom

  1. 01 Setup is a prompt. Paste one line into your coding agent.
  2. 02 Pair a GPU in three commands. Any machine joins your fleet in minutes.
  3. 03 AWS, GCP, on-prem. NVIDIA, AMD. One pool. No GPU stranded for the wrong brand or cloud.
  4. 04 Say what to serve. It works out where. See the plan before anything moves.
  5. 05 Every request goes to the GPU that can take it. It proposes. You approve.
  6. 06 Speaks OpenAI. Speaks Anthropic. Change nothing else.
  7. 07 One fleet. Many tenants. No crosstalk. One tenant's spike never becomes another's outage.
  8. 08 Open a token storefront on your own GPUs. Your customers buy tokens. Your GPUs serve them.

Bring your GPUs. Leave with an API.

Get access Read the quickstart