According to Jassy, Amazon's chip division might be valued at $50 billion, and he suggests that it could potentially sell chips to external customers.

      In summary: Andy Jassy’s annual letter to shareholders, released on April 9, 2026, indicates that Amazon's custom chip division, which includes Graviton, Trainium, and Nitro, generates over $20 billion in annual revenue, experiencing triple-digit growth rates each year. Jassy states that if this division were traded publicly like Nvidia, it would be valued at around $50 billion annually. He also suggests that Amazon might start selling these chips to third-party companies and defends the company’s $200 billion capital expenditure plan for 2026 as being based on firm customer demand rather than speculation.

      “Not on a hunch”: the $200 billion investment

      Jassy begins the financial argument in his letter by directly addressing the skepticism regarding Amazon's capital investments. “We’re not investing approximately $200 billion in capex in 2026 on a hunch,” he stated. “We’re not going to be conservative in our approach. We’re investing to become a significant leader, and our future business, operating income, and free cash flow will greatly benefit from this.” This assertion comes amid a backdrop where the company's free cash flow declined from $38 billion to $11 billion last year, driven by a $50.7 billion rise in capital spending, primarily directed towards AI infrastructure.

      The defense of this investment relies on existing customer commitments. Jassy noted that a significant portion of the expected CapEx for 2026 already has backing from customers, highlighting OpenAI’s commitment of over $100 billion to AWS as an example. This commitment expanded upon an existing $38 billion seven-year partnership established in November 2025 and includes OpenAI utilizing about two gigawatts of Trainium capacity through AWS infrastructure. SoftBank, which owns a majority stake in OpenAI and has been financing its infrastructure development through methods like a $40 billion bridge loan, essentially subsidizes part of the demand that Jassy references as justification for his CapEx approach.

      A $50 billion chip business in plain sight

      Amazon's custom silicon program includes three main product lines. Graviton is a custom CPU which Jassy claims provides more than 40% better price-performance than comparable x86 processors, the domain dominated by Intel and AMD. It is currently used by 98% of the top 1,000 EC2 customers, indicating a significant shift in cloud computing economics. Demand is robust enough that two large AWS customers inquired about purchasing all available Graviton capacity for 2026, a request that Amazon declined.

      Trainium serves as the AI training and inference accelerator and represents Amazon's most direct challenge to Nvidia. Trainium2, which Jassy claims delivers around 30% better price-performance compared to similar GPU alternatives, is largely sold out. Trainium3, which began shipping in early 2026 and offers an additional 30 to 40% improvement in price-performance over Trainium2, is nearly fully subscribed, with companies like Uber moving their workloads onto it. Trainium4, expected to be broadly available in about 18 months and featuring compatibility with Nvidia’s NVLink Fusion interconnect technology, has already seen significant pre-orders. Nitro, the custom network and security chip supporting AWS’s virtualization layer, completes the trio of chips. Jassy notes that together, these three lines generate more than $20 billion annually in revenue, growing at triple-digit percentages year-over-year. “If we were a standalone chip company,” he says, “our chips would generate over $50 billion in annual revenue.” Currently, this business operates solely within AWS; customers access Trainium and Graviton through EC2 instances rather than purchasing the chips directly.

      At scale, Jassy argues, Trainium will “save us tens of billions in capex annually and yield several hundred basis points of operating margin advantage compared to using third-party chips for inference.” This assertion is pivotal to the investment rationale behind the $200 billion CapEx program: custom silicon is not just a competitive edge but a long-term cost advantage that increases over time as the ratio of inference to training in AI workloads rises.

      The Nvidia relationship and the “new shift”

      Jassy carefully articulates the competitive landscape with Nvidia. “We maintain a strong partnership with NVIDIA and will always have customers who opt for NVIDIA,” he notes, while emphasizing that “virtually all AI to date has been conducted on NVIDIA chips, but a new shift has begun.” He states that customers desire better price-performance. Nvidia, which reported revenues of $68.1 billion in the fourth quarter of 2025, reflecting a 73% year-over-year increase, began 2026 from a position of market leadership that Amazon's custom silicon is gradually eroding from within the AWS user base rather than through the broader market. The integration of NVLink Fusion in Trainium4 implies that Amazon is creating a connection rather than a barrier: customers can pair Trainium accelerators with Nvidia GPUs in the same system, maintaining options for enterprises heavily