Does anyone with 100 AI agents running simultaneously and $10k cursor bill actually ship?

At that point does it make sense to just invest in a local AI setup?

Good question

Do open source models compete with sota right now?

I haven't been following but it probably depends on the task. I'd be curious to see someone compare local vector db + RAG / mcp for specific tasks compared to big foundation models via API to see if you can build around your data set or problem space for cheaper* or try and figure out how long you'd break even for the 80/20 of your actual use case. Maybe use the expensive stuff as a fallback even