AWS (AMZN) CEO says seeing shift in AI business from training to inference

Commentary of this kind from a hyperscaler chief executive slots into the long-running debate over where AI spend ultimately sits: the capital-intensive training phase, which concentrates demand in a small number of accelerator suppliers, or inference, which spreads workloads more broadly and shifts the economics toward utilisation, unit cost and recurring consumption revenue. For a cloud operator the distinction matters mechanically, since training demand tends to arrive in lumpy, contract-heavy blocks while inference scales with end-user adoption and is read as a proxy for whether the buildout is generating commercial return. Past episodes of this narrative rotation have tended to reprice the peer set along supply-chain lines, with accelerator and networking names treated as training proxies and cloud operators and software names as inference beneficiaries, rather than moving the whole complex uniformly. The follow-ons that have historically mattered are whether capacity commentary, backlog and capex guidance corroborate the claimed shift, and whether management frames inference growth as additive to training spend or as a substitution, since the two readings carry opposite implications for the hardware complex. As executive framing rather than a disclosed metric, the signal is directional and the burden of proof falls on the next reported segment figures.