The End of Generalist Chips, the Return of Specialized Hardware
Source: a16z
After decades of x86 dominance, the economics of AI workloads are forcing a reversal: custom silicon for specific tasks (TPUs, NPUs, accelerators) now outperform general-purpose processors at scale and cost. This breaks the long monopoly of commodity chips and reshapes the entire stack—from cloud infrastructure vendors designing proprietary silicon, to device makers embedding task-specific accelerators, to startups building chips for narrow applications. The unified hardware market is fragmenting, and whoever controls the silicon-software integration gains a competitive advantage.