LeCun: serving one $200/mo power user costs about $15,000
At UN Open Source Week, Yann LeCun argued open weights are the only path for nations that can't build frontier models. The thread agreed on the destination and tore into the road.
LeCun's economic punchline: today's subscriptions don't cover their own compute, and he wants a diversity of AI systems "for the same reason we need a high diversity of the press."
The thread's flip is hardware, not weights. You can't have open-source AI, one widely-read commenter argues, without open hardware specs and manufacturing. Another asks the unglamorous question: can most of the world afford the machine to run any of it?
"The only way to get to that point is open-source AI platforms"
In a keynote at the UN's Open Source Week, Meta's former chief AI scientist Yann LeCun made an economic case before a political one. Most nations can't build frontier models alone, but they can pool effort into shared, open platforms, and the current pricing model is a mirage: by his number, serving one power user who pays $200 a month costs roughly $15,000 to serve. He framed AI as infrastructure that will soon mediate all interaction with information, and pitched diversity of models as a civic good, "for the same reason we need a high diversity of the press." Restricting open weights on existential-risk grounds, he said, is like limiting the printing press. He pointed to a federated effort, Project Tapestry, joining European states, the UK, UAE, India, Japan, Korea and chipmakers IBM, NVIDIA, AMD and Intel.
The argument the thread had wasn't about weights — it was about silicon
Consensus on the goal, dissent on the mechanism. The most-upvoted structural objection moved the bottleneck up the stack:
"We aren't going to have Open Source AI without Open Source hardware specs and Open Source manufacturing." — a commenter with karma in the 80,000s
Before anyone debates weights, one reader asked who can run them at all:
"How much does it cost a computer with enough resources to run a good enough AI model for agentic workflows… Can 'most of the world' afford buying one?" — a commenter on the device cost
Some doubted "open" means anything here. A skeptic noted the term is doing PR work:
"All models delivered for free as 'open-models' are just freeware as no source is really provided." — a commenter on what open-source AI even is
The optimists reached for history. One drew the operating-system parallel:
"This feels like it rhymes with the story of operating systems… linux still won because control matters over the long term." — a commenter invoking the Linux precedent
The gap nobody closed: LeCun's $15,000 figure is an argument for why closed pricing collapses, not proof that open weights become cheap to run. Sovereignty over a model you can't afford to serve is still a bill someone has to pay.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
They rebuilt Kibu's site by copying Mintlify pixel by pixel
Ben Wallace's principle: find the best work, recreate 100% of it, then sprinkle in your own 3%. He cops to copying Mintlify's 2025 site for Kibu. The thread split on the word "steal."