Amazon Web Services (ML Services)

Contribution to AI

Renting a machine by the hour and handing it back is a dull idea with enormous consequences: after 2006 the cost of trying something at scale stopped being a capital purchase and became a line on a monthly bill. A graduate student or a two-person startup could hold a hundred machines for one night, which is the economic condition under which experimental machine learning became something other institutions than well-funded laboratories could do. The GPU instance arrived in November 2010, a few years before anyone knew how badly the field would want it, so when deep learning became compute-hungry the hardware was already available as a rental. The 2016 launches of Rekognition, Polly and Lex moved the boundary again by selling vision and speech as metered API calls, so a firm with no machine learning staff could buy perception the way it bought bandwidth; SageMaker in 2017 did the same for the awkward middle of the pipeline, turning model training and hosting into managed configuration. Inferentia and later Trainium made Amazon a chip designer, breaking the assumption that one vendor's accelerators were the only substrate available. The commodification cut both ways. Once face matching was a purchasable service, its errors became a public matter — the ACLU's 2018 test of Rekognition against mugshots, and the eventual moratorium on police sales, established auditing a commercial API as a legitimate form of research.

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