Capacity flows from every desk into one model.

Nothing to rack and nothing to buy. A head node runs the fleet, every enrolled desktop contributes the compute it isn't using, and the pooled model answers over a standard API on your own network.

Install the head

One always-on PC runs the head: it loads the model, coordinates the fleet, serves the dashboard, and exposes the API. It is the only machine you configure.

Enroll each desktop

Every Windows or Linux machine on the network is approved once by an administrator. From then on it contributes its spare capacity automatically and steps back when its own user needs it.

Pool one large model

U4x runs a single model larger than any one desktop could hold, by spreading the work across the enrolled fleet. The result behaves like one model, not a cluster you have to think about.

Serve a standard API

The pooled model answers on OpenAI- and Anthropic-compatible endpoints at a private address. Your editors, scripts, and agents point at it with a one-line base-URL change.

Reach remote desks securely

Machines outside the LAN connect out through the U4x relay, which carries only ciphertext. No desktop opens an inbound port, and you can run the relay yourself under your own domain.

Desktops pool · head node · outbound-only relay

This page describes what U4x does, not the internals of how it schedules and distributes the work. The scheduling and distribution design is the proprietary part and is covered under NDA with design partners.

Common questions

What teams ask first.

Do my prompts or documents ever leave the network?

No. Inference runs only on machines you own, inside your network boundary. Traffic between machines is end-to-end encrypted, and the relay that connects remote desks carries only ciphertext U4x cannot read.

Will it slow down the desktops people are using?

Enrollment uses idle capacity and yields the moment someone is active at a machine. A busy fleet draws more power, which is why we report that cost openly on the Cost page.

What do my tools need to change?

Point them at the private endpoint. U4x speaks the OpenAI Chat Completions and Anthropic Messages formats, so most editors, agents, and scripts need only a base-URL change.

How big a model can it run?

A model far larger than any single desktop could load, because the work is pooled across the fleet. See the Benchmarks page for the sizes we publish figures on.