What an orchestrator is

The compute supply layer of Livepeer: what orchestrators do, how they fit into the network, and what running one involves.

An orchestrator is the compute supply layer of the Livepeer network. It accepts jobs from gateways, runs them on GPU hardware, and returns the results. Orchestrators do the actual work — transcoding video and running AI inference — and get paid for it.

The clearest mental model: an orchestrator is a GPU-for-hire on a decentralized marketplace. You connect your GPU, declare what it can run and at what price, and the network sends you work. You earn ETH for each completed job and LPT for participating in the protocol each round.

Where it sits

The network has three layers above the protocol. Orchestrators are the compute layer.

Layer	Who	Responsibility
Application	Developers, streaming and AI products	Send requests to gateways; receive results
Gateway	Gateway operators	Aggregate demand, select orchestrators, dispatch jobs, handle payment
Compute (orchestrator)	You	Execute video and AI work on GPUs; receive payment tickets
Protocol	Arbitrum smart contracts	Staking, reward distribution, payment settlement, discovery

A key consequence: you don’t choose your work — gateways choose you. Selection runs in one direction. Gateways rank orchestrators by capability, price, latency, reliability, and stake, then send jobs to the winners. Everything you do as an operator is about being an attractive choice in that ranking. See How the network works for the full job flow.

What it does

Core responsibilities of the go-livepeer node running in orchestrator mode:

• Execute jobs — receive video segments or AI inference requests and route them to GPU workers.
• Advertise capabilities — broadcast which pipelines, models, and codecs it supports, and the price for each.
• Receive payments — collect probabilistic micropayment tickets per segment or request from gateways.
• Call rewards — trigger the protocol’s reward mechanism once per round to claim LPT inflation.
• Manage workers — coordinate the transcoder (video) and AI runner (inference) processes.

Two pipelines, one node

A single orchestrator process runs two independent pipelines. Many operators run both at once — a “dual-workload” node.

	Video pipeline	AI pipeline
Input	Live video segments from a gateway	An HTTP inference request (prompt, image, audio)
Worker	Transcoder (NVENC GPU)	AI runner (a Docker container per model)
Output	Encoded segments	Image, video clip, JSON, or audio
Payment	Wei per pixel per segment	Wei per pixel or per millisecond
Session	Long-lived (whole stream)	Short-lived (one request or batch)

What it is not

Orchestrators handle compute and payment receipt. They do not route jobs between nodes, ingest RTMP from applications, or manage end-user API keys and billing — that is the gateway role. If you want to aggregate demand and route work across many orchestrators, you’re looking for a gateway, which is out of scope for these docs.

Ways to run one

Setup	What it means	Best for
Solo operator	One node on one machine handles everything	Single-GPU operators getting started
O-T split	Orchestrator and transcoder run as separate processes/machines	Multi-GPU operators optimizing throughput

These docs assume you’re running your own node.

Next

How the network works

Rounds, the active set, and the life of a job.

Economics

The two revenue streams and how they’re split with delegators.

Run your first orchestrator

Put this into practice on mainnet.

Hardware reference

Which GPUs work, and how much VRAM each workload needs.