Jassy Defends $200B AI Spend, Challenges Nvidia with Custom Chips

Investors uneasy about sky-high valuations and talk of an AI bubble just heard from the leader of one of its biggest players. Amazon CEO Andy Jassy used his annual shareholder letter to justify another year of enormous spending while arguing that the company is wresting control of its AI future away from dominant suppliers.

The central tension is straightforward: virtually all AI training to date has run on Nvidia hardware, yet Jassy insists a shift is underway toward Amazon-designed silicon that delivers better price-performance for customers. According to the letter, capacity for the latest Trainium3 chips is already nearly sold out. The still-unreleased Trainium4, 18 months from availability, has also seen nearly all capacity reserved. Jassy pegged the internal chip business at a $20 billion annual revenue run rate. External sales at similar scale would push that figure near $50 billion. Nvidia, by comparison, recorded $130.5 billion in revenue for its fiscal year ended January 2026.

The letter is careful to praise the ongoing Nvidia partnership. It simultaneously promotes Amazon's Graviton processors, now used by 98% of the company's top 1,000 EC2 customers. Two large buyers even requested the entire 2026 supply of Graviton instances. Amazon declined. Those details sit beside Jassy's defense of $200 billion in planned 2026 capital expenditures, the majority earmarked for AWS data centers. "We're not investing approximately $200 billion in capex in 2026 on a hunch," the letter states. It cites a commitment from OpenAI to spend $100 billion on AWS, plus additional signed and pending agreements.

Jassy extended the self-reliance theme to satellite broadband. Amazon's Leo constellation, targeting a mid-2026 commercial launch, has already landed contracts with Delta Air Lines, AT&T, Vodafone, Australia's National Broadband Network and NASA. The letter floats turning Amazon's million-robot warehouse fleet into broader robotics offerings for industry and consumers.

On personnel, the document outlines a future of smaller teams and flatter structures. One Bedrock project required a complete rewrite of its inference engine. The letter claims six engineers, aided by Amazon's Kiro agentic coding service, completed in 76 days what once demanded 40 people working a full year. Amazon has cut roughly 30,000 corporate roles in recent rounds. Jassy said he is pleased with the faster decision-making that followed removal of management layers. Employees, he added, must grow comfortable with ambiguity. Progress rarely follows a straight line. His own path from sportscasting aspirations through multiple careers before joining Amazon in 1997 serves as example.

The letter acknowledges public debate over hype versus substance. For Amazon, Jassy maintains the demand signals are clear and the infrastructure investments necessary. Real-world delivery of the promised chips, satellite network performance, and efficiency gains will ultimately decide whether that assessment holds. Those results remain 12 to 36 months away.