[Traster]

To be honest I almost think the numbers are irrelevant. In 2024/25 there was a lot going on - will AI replace authors, film makers etc. Will it replace social media (anyone remember Sora?). A tonne of that stuff didn't work out. At the tail end of 2025 a real product market fit emerged. Coding agents. They work. They do a job that you can actually profit from.

So everything else is kind of academic. Of course they were losing money in 2025, they had a technology that was kind of cool - clearly eventually going to deliver something great, but they didn't actually have anything somebody should pay for. Now they have a thing that people will pay for. So who cares what they lost in 2025?

So what's important today is - how competitive are they with Anthropic in delivering that product. How do the economics of companies using AI agents for coding work. That's all. I don't think there's really an argument about them losing money on inference any more.

[efficax]

coding agents aren't enough to justify the amount of capital invested

[dofm]

My key realisation from playing with open weights models on my own laptop is that at least where text is concerned, the vast majority of what an average non-programmer consumer thinks AI does, my laptop can now do with the wifi disabled. And arguably where speech and audio is concerned, too.

There is, put simply, a huge, huge information gap about the uniqueness of these commercial services.

There's an open question about how open weights models will be funded when they can't be used in a war between these companies, but the reality is that the amount Apple is paying Google for the right to distill Gemini, for example, is strongly indicative of the total size of the consumer market. Because pretty soon everyone's phones will be doing what local models can do.

Global markets will ultimately learn that coding agents are, at a first approximation, the only source of revenue for this stuff over the medium term at least, and the value proposition for consumer AI in the long term (beyond being a feature of a phone) hasn't yet been invented, and any that might exist depend on micropayments architectures that don't exist.

[danny_codes]

The free fast-follow situation with open models seems to be a big "if" here. It's not particularly hard to set up OpenCode for your company and plug in one of the myriad inference providers running free models. All of these stacks are one release away from being a dramatically different value proposition, as proved by, well, Anthropic themselves.

I guess we'll see if people will pay a premium for Anthropic in ~6 months, 12 months, etc. If not, well, it's a race to commodity.