According to Jassy, Amazon's chip division might be valued at $50 billion, and he suggests that there could be plans to sell them externally.

      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, earns over $20 billion annually, with growth rates in the triple digits each year. Jassy asserts that if this business were traded publicly, similar to Nvidia, it could be valued at around $50 billion annually. He also hints that Amazon might start selling these chips directly to third parties and defends the company’s $200 billion capital expenditure plan for 2026 as a response to confirmed customer demand, not speculation.

      “Not a guess”: the $200 billion investment

      Jassy began the financial portion of the letter by addressing the doubts surrounding Amazon’s capital investments. “We’re not investing approximately $200 billion in capex in 2026 on a guess,” he stated. “We will not be conservative in our approach. Our goal is to be the significant leader, and this will significantly enhance our future business, operating income, and free cash flow.” This assertion comes from a company that experienced a drop in free cash flow from $38 billion to $11 billion last year due to a $50.7 billion rise in capital expenditures, primarily allocated for AI infrastructure.

      His defense is based on existing customer commitments. For the anticipated CapEx in 2026, Jassy mentioned that a significant part is already backed by customers, citing OpenAI’s commitment of over $100 billion to AWS as an example. This commitment extends an existing $38 billion partnership established in November 2025 and involves OpenAI utilizing around two gigawatts of Trainium capacity through AWS. SoftBank, which owns a majority stake in OpenAI and is financing its infrastructure expansion via various means, including a $40 billion bridge loan, is effectively supporting the demand that Jassy references to justify his CapEx position.

      A $50 billion chip enterprise in plain view

      Amazon’s custom silicon initiative encompasses three product lines. Graviton is a custom CPU that Jassy claims provides over 40% better price-performance compared to similar x86 processors, which are dominated by Intel and AMD. This chip is now utilized by 98% of the top 1,000 EC2 customers, highlighting a shift in cloud computing economics that has developed over recent years. Demand is strong enough that two major AWS clients inquired if they could buy all available Graviton capacity for 2026, which Amazon declined.

      Trainium serves as the AI training and inference accelerator, representing Amazon’s most direct attempt to compete with Nvidia. Trainium2, which Jassy states delivers about 30% better price-performance than similar GPU options, is largely sold out. Trainium3, launched in early 2026 and offering an additional 30 to 40% improvement over Trainium2, is nearing full subscription, with companies like Uber transitioning their workloads to it. Trainium4, anticipated to be widely available in around 18 months and featuring compatibility with Nvidia’s NVLink Fusion interconnect technology, is already significantly reserved. Nitro, the custom network and security chip that forms the backbone of AWS’s virtualization layer, completes this trio. Jassy points out that these three lines collectively generate over $20 billion in annualized revenue, growing at triple-digit rates year-on-year. “If we were an independent chip company,” he states, “our chips would be producing over $50 billion in yearly sales.” Currently, this business is exclusively part of AWS; customers access Trainium and Graviton through EC2 instances instead of purchasing the chips directly.

      At large scales, Jassy contends that Trainium could “save us tens of billions of capex dollars each year, and deliver several hundred basis points of operating margin advantage compared to depending on external chips for inference.” This assertion is pivotal to the rationale behind the $200 billion CapEx strategy: custom silicon is not only a competitive edge but also a structural cost benefit that increases over time as the ratio of inference to training in AI tasks rises.

      The Nvidia connection and the “new shift”

      Jassy is deliberate in how he describes the competition with Nvidia. “We maintain a strong partnership with NVIDIA and will always have customers opting to use NVIDIA,” he writes, while claiming that “essentially all AI to date has been conducted on NVIDIA chips, but a new shift has begun.” He notes that clients are “seeking better price-performance.” Nvidia, which reported revenues of $68.1 billion in the fourth quarter of 2025 (a 73% year-on-year increase), started 2026 from a position of market dominance that Amazon’s custom silicon is gradually eroding within its AWS customer base, rather than in the broader market. The incorporation of NVLink Fusion in Trainium4 suggests that Amazon is creating a connection rather than a blockade: customers can pair Trainium accelerators with Nvidia GPUs in the same system, preserving options for businesses that have