AWS CEO Says AI Inference Is “Not Even 1%” of Its Eventual Scale

Amazon Web Services CEO Matt Garman put AI inference at less than 1% of its eventual scale, while saying AWS expects capacity shortages to continue for another couple of years.

“I don’t think it’s even 1%,” Garman said in an interview published Nov. 10. Inference is the computing used to run an AI model after it has been trained. The figure reflects Garman’s view of future demand, not a measured AWS market statistic. The interview did not define what the percentage measures or how he calculated it.

AWS generated $45.984 billion in revenue in the quarter ended Sept. 30, up 39.3% from a year earlier. Its annualized revenue run rate was $184 billion.

Bedrock, AWS’s inference service, also recorded higher spending and usage. Fortune 500 spending on the service was nearly seven times its year-earlier level, while Bedrock token volume was nearly 17 times higher. These figures describe Bedrock activity and do not define the denominator for Garman’s estimate of inference’s eventual scale.

Garman said each 50% reduction in inference prices had been followed by 10 times more usage. He did not specify a measurement period for that comparison.

AWS was adding one GPU every 14 seconds, Garman said. He expects capacity shortages to persist for another couple of years but gave no specific end date. Amazon CEO Andy Jassy made October comments that AWS had sold out its 2026 capacity and was largely sold out for 2027.

Amazon had about $578 billion in unrecognized performance obligations as of Sept. 30, primarily tied to AWS. The amount covers contracts with original terms longer than one year and is not recognized revenue. Revenue recognition depends on customer usage and Amazon’s performance under the contracts.