Livepeer.

Somewhere, a graphics card you'll never see is painting your video.

Right now, a stream is being generated, transformed, or interpreted on a GPU that doesn't belong to Amazon, Google, or any name you'd recognize. It belongs to an operator Livepeer calls an "Orchestrator," and it is doing the unglamorous work of turning raw video into something watchable - or, increasingly, into something that didn't exist a second ago.

That is Livepeer in 2026: an open, permissionless marketplace where video work gets routed to whichever machine will do it cheapest and fastest. No console to log into. No regional outage page to refresh. Just a network of strangers' hardware, coordinated by a token, processing video at a scale that would make a finance team flinch if it ran on the cloud.

It is a strange thing to build. Most of the internet's video runs through a handful of centralized providers, and most people are fine with that until the bill arrives, or the platform changes its mind. Livepeer was built by two people who had watched exactly that happen - and decided they'd had enough.

Video is expensive, and the people who own the pipes know it.

Encoding and delivering video is one of the heaviest workloads on the internet. Live video is heavier still. For years, the only realistic option was to rent that capacity from large cloud providers at prices that quietly capped what a small video startup could ever attempt. If your business model depended on cheap video, you didn't have a business model.

There was a second, quieter problem - the kind founders only learn the hard way. Build on someone else's platform, and you build on someone else's permission. The APIs you depend on can be narrowed, repriced, or switched off entirely, usually on a Tuesday, usually with little warning.

So the question that became Livepeer was less idealistic than it sounds: what would video infrastructure look like if no single company could shut it off, and if the cost was set by an open market instead of a sales team?

Two repeat founders, one stubborn idea: rent the world's spare GPUs.

Petkanics and Tang were not first-timers. They'd built Hyperpublic, a local-data company that Groupon acquired, and then Wildcard, the browser that taught them the cost of building on closed ground. By 2017 they had a thesis and a scar to go with it.

The bet was that the same crypto-token mechanics powering Ethereum could coordinate something physical and useful - not financial speculation, but video. Stake a token, take on jobs, get paid in fees and freshly minted rewards. The hardware was already out there, idling in machines all over the world. The trick was paying for it honestly and routing work to it reliably.

Investors eventually agreed. Livepeer raised across four rounds: early seed money from Northzone and crypto-native funds, then larger checks that culminated in a $40M Series B in 2022 led by Tiger Global, with Digital Currency Group, Northzone, and billionaire Alan Howard joining. Total to date: $51M. Not enormous by infrastructure standards - which is rather the point. The network's capacity doesn't come from the company's balance sheet. It comes from everyone else's hardware.

It started as a cheaper way to transcode. Then it learned to imagine.

For most of its life, Livepeer did one thing well: transcoding - the necessary, invisible work of converting a video stream into the formats and resolutions different devices need. Developers plugged in through Livepeer Studio, the company's API layer, and paid a fraction of cloud prices because the work ran across a competitive marketplace of orchestrators.

Then real-time AI video arrived, and with it a workload even hungrier for GPUs than transcoding ever was. Generating or transforming live video on the fly takes serious compute - far more than text or images. Which happens to be exactly the kind of demand a distributed GPU network is built to soak up.

For a developer, that means three verbs instead of one. You can still transcode. But you can also pipe a live stream through a generative model and watch it come back changed in real time - the sort of thing that, on a metered cloud bill, would end most experiments before they began.

A network is only as real as the people using it.

Livepeer's network has scaled to thousands of concurrent streams across a globally distributed set of orchestrators. The users are developers and startups building video and AI-video applications - the kind of teams for whom cloud video pricing was a non-starter. Open-source social network Minds worked with Livepeer Studio to ship live streaming. Daydream went to market around creative real-time AI, using the network and contributing back to it.

In 2025, the AI Subnet's fees grew quarter over quarter - small numbers in absolute terms, but pointed in the direction that matters - driven by Daydream's beta, ComfyStream usage, and a wave of hackathon activations. The token, LPT, isn't just a speculative chip; it's the mechanism that decides who gets to do the work and who earns from securing it.

An open media layer for a web that keeps trying to close.

The stated mission has stayed remarkably stable: build a fully decentralized, highly scalable, token-incentivized live video network that can serve as the media layer of the web3 stack. Underneath the jargon is the same instinct that started the company - video infrastructure that's cheap enough to be useful and open enough that nobody can revoke your access on a Tuesday.

It's governed in the open, too. A small core team sits alongside a broad ecosystem of orchestrators, delegators, and contributors who vote on the protocol's direction. That's slower than a boardroom. It's also harder to capture, which for a company founded on the memory of a closed API is the entire point.

Real-time AI video is coming for a lot of GPUs. That's the opening.

If the next wave of media is generative and live - avatars, effects, streams that respond to a prompt as they play - then the bottleneck is compute, and a great deal of it. Centralized providers can supply that, at centralized prices, with centralized control. Livepeer's wager is that an open market of idle GPUs can meet the same demand for less, and without a single off-switch.

That wager isn't won yet. Decentralized networks are harder to make reliable, the AI-video market is young, and plenty of well-funded competitors want the same workload. But the thesis has aged well: video keeps getting heavier, and the case for not paying monopoly rent on it keeps getting stronger.

So return to that graphics card you'll never see. A few years ago it would have sat idle, or its owner would have been a customer of someone else's cloud. Today it's part of a network, taking jobs, getting paid, painting video for people who never had to ask anyone's permission to send it. The pipe, for once, belongs to no one in particular. Which was the idea all along.