it's going to be mostly open models running on consumer grade GPUs except for a minority of use cases. Depending on what eval you use, the best 20-30b parameter models are at about 5-8 month lag behind the frontier closed source models and that's been consistent for a while now.

There's a longer talk I give on this but briefly we are in a time period where the development of the technology is quickly outstripping people's needs and they will be satisfied by models runnable on commodity sub-$10k machines.

That threshold has arguably been met for many users over the past 6 months and it will continue to be met for the majority of use-cases in the upcoming months.

So not only does Anthropic and OpenAI have no moat - unless they have other compelling products, the demand for their token-based subscription and metering products isn't sustainable because consumer preferences go elsewhere after any product quality reaches a sufficient baseline.

Luckily there's many many options. Faster, cheaper, stronger they're there but these are all zero-profit condition qualifiers. The labs current push, and this is across the board, is to have a compelling suite of applications where people will have a preference for them or there will be some side hustle. Look at Meta muse - they have an app for your phone that will hoover up your personal data in the name of convenience. 5 million people said yes in 3 weeks.