Elon could pull their compute a moment's notice (not out of character for him), and NVIDIA just lost its CEO (for possibly quite illegal trade)
They also failed to make any formal communication around the Cursor Claude extension exploit a couple weeks back.
I like Claude, but I see an immature company trying to expand much faster than they are ready for.
The thing is if models become 'good enough' that the main value is in running existing models, those will quickly proliferate and be available for commodity cost from many competing providers. Why pay Anthropic when I can just pay AWS or Azure directly to run whatever future version of Qwen does the job for me?
If the value actually remains in the frontier models, you can't stop spending on training new ones or you'll eventually be left behind by your competitors and they'll capture that value.
I haven't yet seen a good explanation of how companies like OpenAI and Anthropic will make the transition to a lower cost structure while maintaining the revenue that they currently get for being on the cutting edge of LLMs.
They have some stickiness to the tooling people are used to, but those can be overcome when there's a significant price difference for the same results.
Right, all the $200+ subscribers that I know are incredibly fickle.
The answer is: by lobbying the government to coordinate a slowdown "for safety." Obviously, Chinese labs won't and will therefore will pull ahead.
There is no scenario where they could realistically stop R&D. One solution would be to have the American taxpayer pay for R&D, sold as staying ahead in the AI arms race: so I'm expecting such lobbying attempts any month now. Hopefully the recently elected representatives follow through on their promises to represent their constituents.
I also worry about the compute costs. Maybe they can make a profit now but at some point the massive investments of everyone into compute must be payed back.
The two oldest monopolies: regulatory capture, and real estate
Peruse any ai dev forum and you’ll see strong preferences towards one or the other (Claude/ChatGPT) due to capabilities. A negligible amount of people are running their own inference.
So long as theres a felt gap between open models and frontier you will have people pay for the best of the best. Not doing so is an admission of actively handicapping yourself and not in a way that is easy to justify.
When the company is big enough, the decision doesn't rest on developers and execs are much more likely to go with established vendors -> no open models
When the company is small enough, the 200$ plans from either ant or openai are way way better value for money than anything you can get with open models. This may change in the post-IPO future.
It is as simple as that.