Amazon CEO says AI costs need to drop 90% for wide adoption

Amazon CEO Andy Jassy said in an interview with venture firm a16z that AI inference costs are still about ten times too high. For AI to spread widely in business software, costs need to fall to around one-tenth of today's levels.

Amazon plans to spend $220 billion on infrastructure this year, but still can't meet AI demand. Its computing capacity for 2026 is nearly fully booked, and most of 2027 is already reserved. Jassy believes the strong demand means more spending won't be wasted.

Jassy compared AI demand to a barbell: one end is AI labs using massive computing power, and the other is companies using AI for tasks like customer service and fraud detection. The middle market of everyday business software is still untapped and could become the biggest source of AI demand.

Nvidia now leads the AI chip market. Amazon launched its own Trainium chip, which is about 30% to 40% more cost-effective than similar processors. Trainium has secured over $225 billion in future revenue commitments, including deals with Anthropic and OpenAI.