When a Law Ends, the Industry Doesn't Stop — It Specializes

For decades, waiting a couple of years for the next generation of general-purpose chips was a reasonable strategy — transistor density roughly doubled, and every workload got faster for free. That free lunch has slowed considerably, and the industry's response wasn't to give up on performance gains; it was to stop expecting one chip design to be good at everything.

Specialized silicon — chips built for one workload, like a specific class of AI computation — can deliver order-of-magnitude efficiency gains that general-purpose chips can no longer match, precisely because they give up flexibility to get there. That trade-off used to not be worth making. Now, for high-value workloads run at large enough scale, it often is.

The strategic implication is that chip strategy is no longer just "wait for the next process node." It's "figure out which workloads are big enough and stable enough to justify designing silicon around them" — a much more deliberate, much more expensive bet.