machine-learning A field note by Vikrant Sharma
AI models are starting to ship as torrents
Distributed file sharing might be the only way to serve 70GB models without bankrupting yourself on egress fees.
Someone finally said it: AI models should ship as torrents. Not because it sounds cool. Because AWS charges you 9 cents per gigabyte to serve a file, and a single Llama 70B download costs you 6 dollars.
The math is brutal. If ten thousand people download your model, you owe Amazon sixty thousand dollars. If you are Hugging Face or Meta, fine. If you are a research lab or a startup, that is not sustainable. You either put the model behind an API and lose the open source angle, or you haemorrhage cash.
BitTorrent fixes this by turning every downloader into a seed. The first hundred downloads cost you money. After that, the swarm distributes the load. No central server melting under traffic. No egress bill that makes your finance team cry.
The article points out that model registries already act like centralised bottlenecks. Hugging Face hosts everything, which is convenient until it is not. A torrent swarm is just a magnet link and a tracker. You can host the tracker on a 5 dollar VPS. The model itself lives everywhere and nowhere.
This is not some piracy nostalgia play. It is infrastructure pragmatism. Models are static files. They do not need a CDN. They need peers. The tooling is already there. Someone just has to package it so pip install can fetch a magnet link instead of a 70GB HTTPS download.
The weird part is that torrents solve the authentication problem too. If the model hash is baked into the magnet link, you know the file is intact. No one can swap a backdoored checkpoint into the swarm without breaking the hash. That is better than trusting a single registry to not get breached.
I would bet that within two years, at least one major model ships with a torrent option by default. The economics demand it.