What happened
A new mathematical framework shows that as AI workloads grow larger and more general-purpose, specialized hardware becomes less efficient than flexible programmable chips — reversing the historical assumption that specialization always wins. This explains why GPUs have dominated AI compute despite years of startups building cheaper, narrower alternatives.
Why it matters
For decades, computer architecture assumed specialized hardware would always beat general-purpose chips if you optimized for a specific task; this paper shows that assumption breaks at scale, which means companies betting billions on domain-specific AI accelerators may be building toward obsolescence while GPU makers can keep improving.