What a weight file actually is
A model's weights are a very large list of numbers, a few hundred gigabytes for a frontier model. Once they exist, copying them costs nothing. That is the whole economic puzzle in one sentence. The thing that took hundreds of millions of dollars to make can be duplicated for the price of bandwidth.
Closed labs (OpenAI, Anthropic, Google DeepMind) keep the file private and sell access by the token. Open-weight labs (Meta with Llama, Mistral, DeepSeek, the Qwen team at Alibaba) publish the file and make money somewhere else, or not yet at all. The word 'open source' gets used loosely here. Most of these releases share the weights but not the training data or the full recipe, so you can run the model and fine-tune it, but you could not rebuild it from scratch.
Why a company would give away a billion-dollar asset
Meta is the clearest case. It does not sell model access. It sells advertising on Facebook, Instagram and WhatsApp, and it wants AI features in those products to be cheap. If open weights push the whole industry's prices toward zero, Meta's costs fall and its competitors' revenue falls. That is a good trade for Meta and a bad one for OpenAI. Mark Zuckerberg has said roughly this out loud on earnings calls.
The second reason is talent and standards. Researchers want their work used. A model that runs on every university cluster and every startup laptop becomes the default that people build tools around, and the lab that publishes it sets the conventions everyone else has to follow. Linux did this to Unix in the 1990s. Android did it to the phone market.
The third reason is the one nobody says in a press release: it is a way to be relevant when you are behind. DeepSeek's releases in early 2025 forced a rerun of every 'how much does frontier AI cost' spreadsheet in the industry. Publishing was the cheapest way to prove the claim.
What the closed labs are really selling
If weights were the product, closed labs would be finished. They are not finished, and the reason is that the product is the running service. Uptime, latency, safety tuning, a support contract, a compliance certificate, and the guarantee that the model you tested in March still behaves the same way in September. Enterprises pay for that. Governments pay more for it.
There is also a capability gap, and it matters how long it lasts. In most public benchmarks the best open model trails the best closed model by somewhere between six and twelve months. For a lot of tasks that gap is irrelevant. For the hardest tasks, the ones a bank or a drug company will pay top dollar for, it is the entire business.
So the honest picture is a price ladder. Free-to-run open models at the bottom, mid-priced closed APIs in the middle, and a thin, expensive frontier at the top that is rebuilt every year. Most of the volume will end up at the bottom. Most of the revenue will stay near the top for a while.
What I would watch
Two numbers. First, the price per million tokens for a model at a fixed capability level. It has fallen by roughly ten times a year since 2023. If that continues, the middle rung of the ladder gets squeezed out, and the closed labs live entirely on the frontier and on enterprise contracts.
Second, the gap between the best open and best closed model on hard reasoning tasks. If it closes to a couple of months, the frontier stops being a moat and becomes a marketing expense. If it widens, the closed labs' bet pays off.
My read: both camps survive, and the interesting company is whichever one figures out how to make money from running open models well. That is a services business, not a research business, and services businesses are where the boring, durable profits usually live.