Chip design is basically a massive search problem. Let's take a design with 500,000 standard cells and simplify the problem massively. Forget routing, timing, power, memory, architecture, basically everything else. Just imagine 500,000 cells and ask: how many different ways could I arrange them? That's 500,000! Roughly 10^2,632,341 possibilities. The exponent alone is over 2.6 million. Just for some sense of scale, the number of legal positions in Go is ~10^170, and the commonly cited estimate for its full game tree is ~10^360. Protein folding has estimates around ~10^300 possible conformations for a typical protein. Obviously these aren't the same problems, and I'm not saying chip design is "X times harder" than Go or protein folding. It's just a way to get a feel for how quickly the numbers get ridiculous. And real chip design is much messier than this toy example. You're making decisions about architecture, memory hierarchy, interconnect, precision, voltage, clocks, floorplanning, routing, timing, power, packaging, cost... and everything affects everything else. Improve timing and you might hurt power. Change the memory system and you've changed bandwidth, latency, area and cost. Of course nobody brute forces this. EDA tools use hierarchy, constraints, heuristics and decades of optimization work to narrow the problem down. But we still only explore a tiny fraction of what could be built, and exploring each promising direction takes real compute, engineering time and money. I think AI changes the economics of that search. Search more of the space. Kill bad ideas earlier. Run more experiments. Try architectures we wouldn't have had the time or money to try before. And if that loop gets dramatically cheaper and faster, I think we end up taping out more chips, not fewer. More architectures become worth trying. More workloads become worth building custom silicon for. There's a lot of silicon we never even get the chance to explore today!!