Andy Jassy’s annual letter to shareholders, published on April 9, 2026, indicates that Amazon’s custom chip business, involving Graviton, Trainium, and Nitro, exceeds $20 billion in annual revenue and grows at triple-digit rates each year. Jassy suggests that if this business were sold like Nvidia, it could be valued at approximately $50 billion annually. He also hints at the potential for Amazon to sell these chips directly to third parties and supports the company’s $200 billion capital expenditure plan for 2026, emphasizing that it is based on committed customer demand rather than speculation.
**The $200 Billion Investment**
Jassy addressed criticism of Amazon’s capital spending by stating, “We’re not investing approximately $200 billion in capex in 2026 on a hunch.” He emphasized that the investment aims to make Amazon a leader, predicting substantial growth in future business metrics. Last year, Amazon’s free cash flow decreased from $38 billion to $11 billion due to a $50.7 billion rise in capital spending, primarily allocated to AI infrastructure.
Jassy backs the expenditure with existing customer commitments, highlighting OpenAI’s over $100 billion commitment to AWS as an example. This agreement, building on a previous $38 billion partnership from November 2025, includes utilizing approximately two gigawatts of Trainium capacity through AWS infrastructure. SoftBank, majority stakeholder in OpenAI, effectively supports part of the demand that validates Jassy’s CapEx approach.
**The $50 Billion Chip Business**
Amazon’s custom silicon program includes three product lines. Graviton, a custom CPU, offers over 40% better price-performance than similar x86 processors and is used by 98% of the top 1,000 EC2 customers. Demand is high, with two major AWS clients expressing interest in purchasing all Graviton capacity for 2026, an offer Amazon declined.
Trainium, an AI accelerator, competes directly with Nvidia. Trainium2, offering about 30% better price-performance than similar GPU options, has an exemplary sales record. Trainium3, with a further 30 to 40% improvement, is almost fully booked, with companies like Uber using it. Trainium4, featuring compatibility with Nvidia’s NVLink Fusion, is already largely reserved. Nitro, a custom chip for AWS’s network and security, rounds out the product lineup. Jassy states these lines yield over $20 billion in annual revenue with triple-digit year-on-year growth. He asserts that as a chip company, the business would generate over $50 billion annually. Currently, Amazon sells these chips within AWS; customers access them via EC2 instances.
According to Jassy, Trainium will ultimately save billions in CapEx annually and improve operating margins by using Amazon’s chips. This claim strengthens the investment foundation of the $200 billion CapEx plan, positioning custom silicon as a differentiator and structural cost benefit that increases as AI workloads shift more toward inference.
**The Nvidia Dynamic**
Jassy carefully discusses the competitive aspect with Nvidia, maintaining, “We have a strong partnership with NVIDIA, will always have customers who choose to run NVIDIA,” but notes a shift as customers seek better price-performance. Nvidia, beginning 2026 with market dominance after a fourth-quarter revenue of $68.1 billion in 2025, is challenged by Amazon’s custom silicon, which targets customers within AWS rather than the wider market. Trainium4’s integration with NVLink Fusion allows customers to use both Trainium accelerators and Nvidia GPUs together, maintaining flexibility for those committed to Nvidia’s software.
A notable point from the letter is the mention of potentially selling chips directly to third parties. Currently, Amazon monetizes its custom silicon solely via EC2 compute services. Selling chips directly could shift Amazon’s competitive position in the silicon market alongside Nvidia and AMD, separating chip business economics from the supporting cloud revenue.
**Broader AI Strategy**
The letter frames the chip business within a broader AI infrastructure strategy. Amazon Bedrock, facilitating access to foundational models including its Nova family, processed more tokens in Q1 2026 than ever before, with inference volumes nearly doubling monthly in March. AWS’s AI revenue reached $15 billion in Q1, growing approximately 260 times faster than AWS at a similar stage.
Jassy also presents Amazon’s satellite internet service, Amazon Leo, as a rival to SpaceX’s Starlink, securing contracts with companies like Delta, JetBlue, AT&T, Vodafone, and NASA. The letter argues Amazon is investing in infrastructure on a scale not fully recognized by observers. While legal scrutiny over AI products such as a class action regarding Nova’s training data is noted, the letter doesn’t address it. With AI infrastructure establishing itself as the key capital allocation issue in 2025, Jassy’s letter argues Amazon identified and addressed this trend earlier and more effectively than the market appreciates.