Why Centralized AI Became the Default — and Why That Assumption May No Longer Hold

Over the last decade, we’ve mostly talked about AI in terms of scale: bigger models, more GPUs, larger data centers. Centralization slowly became the default assumption, often without being explicitly questioned.

This piece is an attempt to step back and look at the system as a whole — not from a hype perspective, but from an architectural one. Where do the real constraints come from? Which ones are still binding, and which ones are quietly dissolving as inference becomes dominant and hardware evolves?

I’m not arguing for a particular ideology or for the disappearance of centralized systems. I’m more interested in understanding what is actually becoming possible again, and what kinds of design choices we may soon have to make — whether we’re ready for them or not.

If you work close to the metal — infra, ML systems, hardware, distributed systems — I’d be genuinely curious how this resonates with your experience.

👉 Read the full piece here: If AI is centralized today, it is not a law of nature