TL;DR
Nvidia agreed to buy Hugging Face, the main platform for open AI models, for $12.93 billion on 3 September 2026. The Hugging Bay, a volunteer-built BitTorrent index of the same open models, went online in July, before the deal was even reported, built explicitly as protection against platform control, in its own words against “policy pressure, geo-blocking, bandwidth caps.” Its architecture is credible: license-scoped, checksum-verified, no single company controls it. Its current network isn’t up to the job yet: five listings, two peers, almost all of it from one uploader. It only helps if people seed important models before a platform crisis happens.
In plain English
Open-weight model: an AI model whose trained parameters (“weights”) you can download and run yourself, rather than only accessing through a company’s hosted service.
BitTorrent / magnet link / seeding: a way of sharing files where a “swarm” of individual computers hold pieces of the file and send them to each other, rather than one server holding the whole thing. A magnet link tells your torrent software what to look for; “seeding” means keeping your copy available for others to download from. A file with zero seeds cannot be downloaded, regardless of how complete its listing looks.
SHA-256 checksum: a short fingerprint calculated from a file’s contents. If two files produce the same checksum, they’re identical — it’s how Hugging Bay’s “Captain”-verified badge confirms a torrent matches the original release rather than a tampered copy.
Redistribution-permitting license: a software or model license (Apache-2.0, MIT, certain Creative Commons variants, and similar) that explicitly allows someone other than the original publisher to copy and share the file. Hugging Bay only indexes files under these — it isn’t a tool for pirating models whose licenses forbid resharing.
Nvidia agreed on 3 September 2026 to buy Hugging Face, the platform where most of the AI industry finds, downloads and shares open models, for $12.93 billion. The deal puts the industry’s dominant chipmaker in charge of the site millions of developers use to find, evaluate and deploy models. Two months earlier, in July, a volunteer developer put a different kind of infrastructure online: The Hugging Bay, a BitTorrent index of the same open models, distributed by a community swarm instead of any single company. Its own About page states the reasoning: “Centralized hubs are single points of failure — policy pressure, geo-blocking, bandwidth caps. A swarm is not.”
Hugging Face describes itself as “the platform where the machine learning community collaborates on models, datasets, and applications.” It hosts work from Meta, Alibaba, Mistral and thousands of independent researchers, alongside more than 500,000 datasets. Ownership of that platform means influence over which models get ranked, discovered and deployed easily, beyond simply which files exist.
Hugging Bay indexes BitTorrent magnet links only for models, weights and datasets under licenses that explicitly permit redistribution: Apache-2.0, MIT, BSD, the CC and ODC families, OpenRAIL variants, and community licenses like Llama’s or Gemma’s. It never hosts files itself; it parses torrent metadata, publishes magnet links, and lets volunteer “sailors” do the seeding. A “Captain”-verified badge means someone re-hashed a listing’s file against the original release’s SHA-256 checksum, per the project’s policy.
The project’s GitHub repository, run by a developer using the handle DrMaxis, was created 2 July 2026. Its torrent catalog was uploaded 23–24 July. Nvidia’s acquisition talks weren’t reported until 27 August, five to eight weeks later depending which of those dates is the starting point. Hugging Bay existed before the deal that people are now comparing it to.
Nvidia’s blog post confirming the deal, written by CEO Jensen Huang, puts the price at $12,930,300,000, the company’s second-largest acquisition on record after the $20 billion Groq purchase in December 2025. It states: “Hugging Face will remain an open platform for the entire AI ecosystem. Developers will choose the models they want, the frameworks they want, the clouds and inference service providers they want and the computing platforms they want. NVIDIA compute will not be required to build on or deploy through Hugging Face.” Hugging Face CEO Clément Delangue told CNBC he approached Huang over the summer, after an August hacking incident convinced him the platform “needed more, more resources, more scale, more visibility.” Observer reported Delangue had turned down a $500 million Nvidia investment offer in 2025.
Nvidia’s commitment covers compute only: no requirement to run Nvidia hardware. It says nothing about ranking, discoverability, telemetry, takedown policy, or which models get the easiest deployment path. On Nvidia’s earnings call eight days before the deal was announced, Huang told investors, on the record, that open and closed models “are simultaneously driving our sales” as both categories grow.
As of 15 September 2026, Hugging Bay’s fleet page lists five listings, 29.7 GB indexed, and two peers in total. Nearly all of that, 29.7 GB, was uploaded by one account, “syndicalt.” A second account, “the-bay-itself,” posted an automated 30 KB catalog file. The project’s GitHub repository hasn’t taken a code commit since 8 August.
My read: Hugging Bay works as a fallback if Hugging Face ever becomes unusable. It doesn’t replace what Hugging Face does day to day. Its architecture is credible: index-only, license-scoped, checksum-verified, no single company controlling access. Its network isn’t there yet — five listings and two peers can’t preserve much if a platform actually restricted access tomorrow. It becomes useful only if people seed the models worth protecting before a platform failure or policy change happens. No source says Hugging Face’s catalog will be censored or gated. What changed on 3 September is who controls discovery and deployment.