The shift to physical AI

I built a second brain in Obsidian a few months ago. The whole point was to give Claude the context I work with — every meeting I sit in, every book I'm reading, the people I work with, the projects I'm running, my open actions. If the model could see what I see, it would stop living in my tabs and start showing up in my actual life.

It mostly worked. Claude reads my meeting notes, pulls cross-references between projects I'd forgotten existed, and works from almost the same context I do.

But there's still a manual layer I can't escape. Every podcast I listen to, I have to remember to run the ingest script. Every book, every article, every conversation — captured only if I sit at my laptop and process it. The context the model has is a couple-days-stale snapshot of my brain, not the live feed. I'm still doing the data entry.

The next evolution is the model having its own eyes and ears. A device with you all day, capturing what you see and hear with your permission, so the model has the live substrate instead of the curated snapshot.

This is why consumer hardware is the next frontier. The products that win won't introduce themselves as "AI products." They'll just be better objects with AI threaded through them. The rumored OpenAI device — the Jony Ive collaboration that's been leaking for months — is the biggest bet of all. Whoever wins that form factor controls the layer where AI stops being a chatbot and starts being a partner.

Physical AI isn't a side bet against software AI — it's the layer software AI needs to actually be useful. The context problem has only two long-term solutions: massively better proprietary data, or embodied capture that builds the context as you live. Both roads end at hardware.

And consumer hardware is genuinely hard in a way software isn't. DRAM prices have spiked in the past year because AI compute is eating the supply, and every consumer device now lives inside that constraint. Industrial design is the difference between something that feels like an Apple object and something that feels like a Kickstarter prototype. Manufacturing in Asia isn't easy nor accessible; it's vendor relationships built over years.

None of this is taught in school, and in industry the knowledge is siloed. Take semiconductor chips. TSMC makes the world's most advanced ones in a handful of Taiwan fabs — process knowledge so specialized that Intel and Samsung have spent a decade trying to replicate it and can't.

Most of my AI today is software. But Software AI is saturating fast. Every PM and engineer is shipping the same wrapper around the same model. The next decade's real action is moving to hardware, and the talent pool for it is thin. I see everyone around me optimizing their career for SaaS. The few people who paid attention to hardware are about to look prescient.