pull down to refresh
It's good enough until your work gets stolen and you lose income. Prevention is always a hard case to make.
I've been relying on TEE-based providers for a small subset of my work where I seek automation for open source research but on confidential problems, and I have been using some local models for processing actual confidential data (mostly: docs and embeddings)
I don't really like the TEE solution if I'm honest because that still requires me to work online, while I'd prefer to work offline in a real sandbox (not like the leaky things OpenAI uses for security benches, lmao). But gemma4 or qwen3.8 or some distill small enough for me to run locally on 64GB uRAM is often not cutting it for agentic work and even if it does get a result that is slightly acceptable, it takes forever. So I kinda need me one of those B300 racks if I ever want to take on a wider scoped confidential job with this.
Probably will break down at the atomic level, where the continuum model stops being valid?
I'm starting to think about that, too. But as always, even if more and more people will do it, it'd still be a majority compared to the average person who does not want to pay money for it ("free" online version is good enough for most tasks, so why invest in overpriced hardware?). Thus, not sure about the investment thesis just yet.