Glue and coprocessor architectures

Vitalik Buterin argues that modern computation is increasingly organized as a "glue and coprocessor architecture," in which a general but inefficient glue component handles business logic while specialized, efficient coprocessors perform the bulk of intensive work. In a recent Ethereum transaction updating an ENS hash, about 73% of EVM execution (about 85% including the computational part of the base cost) went to storage reads and writes, logs, and cryptography. The same split appears in Python with torch, where matrix multiplication dominates, and in STARK proving, where specialized modules handle hashes and signatures alongside a RISC-V prover. The article suggests the EVM should be optimized through better precompiles and storage layout rather than by changing the VM itself.