a16z: Why specialized chips are making a comeback in the AI era
What's the deal? a16z's Zain Asgar argues that hardware specialization is returning as AI workloads force the chip industry to rethink computing. The piece revisits a 30-year-old debate — general-purpose chips versus specialized ones — now reopened under new constraints.
The backstory: In 1990, Eugene Brooks of Lawrence Livermore warned that "nothing would withstand the attack of the killer micros" — general-purpose microprocessors sold by the millions into desktop PCs. Specialized architectures such as vector processors, massively parallel processors, and Lisp machines died off soon after, with companies like Multiflow, SymbolicsDealroom has a profile for this one. Try Dealroom →, and Thinking MachinesDealroom has a profile for this one. Try Dealroom → folding between 1989 and 1995.
Why did generalists win? They rode a virtuous cycle: generality drove volume, volume drove performance, and that performance reinforced generality. From 1986 to 2003, single-processor performance grew over 50% year over year, mostly from faster clock speeds. Low-volume specialists could not fund the same cadence — and the generalists absorbed their best ideas anyway, as CPUs did with vector processing.
Why now? AI workloads differ enough from traditional software that the industry is again rethinking chip design. GPUs now generate more revenue for chip companies than CPUs, alongside what Asgar calls a "massive explosion of new accelerators."
The open question: Asgar asks whether this time is different, or whether specialized chips will once again be absorbed into general-purpose designs, as Thinking Machines founder Danny Hillis predicted when he said "ultimately these things were just going to be boxes of microprocessors."
The signal: The argument lands amid a clear industry shift. Google, Amazon, Microsoft, and Meta are all advancing in-house AI accelerators, with inference a growing focus, as they look to cut cost and energy use and reduce dependence on scarce GPUs. The idea is not new — Google's TPU dates back years — but rising AI inference volume makes specialization more commercially compelling, even as alternatives still face Nvidia's software ecosystem and scale.
Read more: a16z.news
Image credit: Sangitiana
