Intel / article

Opencompute

Source: factory.com
Author: Factory News
Date: Published Oct 14, 2026

Context

OpenCompute manages working environments for coding agents on customer-controlled infrastructure. It builds on Droid Computers, extending the earlier machine-level setup with fleet management. According to the announcement, it records the requested machine state durably, then repeatedly works toward that state, including retrying recoverable failures. Restarting the controller therefore does not erase unfinished provisioning.

Provider-specific adapters translate those requests into infrastructure operations; optional capabilities such as suspension depend on the provider. The October 14 announcement places OpenCompute in private preview. Keeping execution on customer infrastructure still leaves model and session storage dependent on separate configuration. Architecture details.

Key takeaways · AI-distilled

Factory's OpenCompute, in private preview, is a self-hosted orchestrator that provisions, suspends and terminates sandboxes on customer infrastructure, extending its earlier one-machine-at-a-time Remote Control model to whole fleets.
OpenCompute dials out: it opens an outbound TLS WebSocket to Factory and proves possession of a private key, which can live in a cloud KMS, by signing a challenge, so customers need no public endpoint or inbound firewall port.
Its Go control plane records desired and observed state in a local Postgres and a reconciler drives each computer toward it, so a dropped connection, throttled provider or control-plane restart does not lose in-flight provisioning.
Each backend is a standalone gRPC driver held to a required Core conformance profile plus optional Suspend, Snapshot and Fork profiles; unsupported operations are rejected before any durable state changes.
Factory says the build was agent-native, with Droids implementing against defined contracts and conformance tests under human direction: six weeks from prototype to first customer deployment, and an on-prem driver built during a flight, with hardening continuing after.

Terms in this piece · Glossary

inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.

Read the source factory.com