Advanced Micro Devices (AMD)

Contribution to AI

Second-sourcing was the original trade: manufacturing chips another firm had designed so that no buyer depended on a single supplier. Being the alternative is what the field eventually got out of it. The x86-64 extension, specified in 1999 and shipping in silicon from 2003, is the plainest case — instead of breaking compatibility as Intel's Itanium had, it grafted 64-bit addressing and sixteen general registers onto the existing instruction set, so the large address spaces that later data-heavy work assumes arrived on ordinary machines running unmodified software. The $5.4 billion purchase of ATI in 2006 put graphics silicon and processors under one roof, which is the origin of everything AMD later sold into machine learning. The Boltzmann Initiative of 2015 and the ROCm stack that followed in 2016 mattered less for raw throughput than for HIP, a translation layer that let CUDA source be converted rather than rewritten, and which remains the only serious route off a proprietary programming model that had quietly become the field's default. Frontier at Oak Ridge, the first computer measured past an exaflop in 2022, ran on EPYC processors and Instinct MI250X accelerators throughout. The MI300X of 2023 carried 192GB of HBM3, a deliberate bet that memory capacity, not arithmetic, is what decides how many devices a large model needs. The contribution is competitive pressure made concrete: prices, licences and portability across the accelerator market all reflect there being somewhere else to go.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).