The Coming Collapse of the AI Factory.
For now, “AI factories,” those sprawling hyperscale compute centers, look inevitable. The money, the hardware, the narratives all point to centralization. But the scaling laws are bending back on themselves.
Every generation since GPT-2, we’ve seen the same pattern: you can reproduce or exceed prior-state performance with a millionth the parameters and a fraction of the energy. It’s not Moore’s law of transistors anymore; it’s the Moore’s law of intelligence.
That means in two years, the “AI PC” will be the next iPhone: 500 tokens per second and 32-bit context, local, private, and cheap. When that happens, the entire premise of centralized AI will collapse under its own economic gravity. Eleven-dollar-an-hour compute can’t compete with a one-time hardware purchase.
When that shift happens, we’ll see a new social fabric emerge. People will spend an hour a day with their agents, doing business, creating, experimenting. Work will become a simple conversation.
The edge will eat the cloud.
🍩