Amazon has set a very specific figure and scene regarding the current state of artificial intelligence infrastructure. In its annual shareholder letter, published on April 9th, CEO Andy Jassy stated that two major Amazon Web Services clients requested to purchase the entire capacity of Graviton instances planned for 2026. The company, he added, rejected this possibility because it also needed to serve the rest of its customers. This phrase doesn’t refer to a one-time chip sale but to the full cloud capacity based on AWS’s own processors, serving as a gauge for how strained demand has become.
The message extends beyond the anecdote. Amazon is leveraging this demand pressure to defend its significant investment in AI, data centers, and its proprietary silicon. Jassy contends that AWS could grow even faster, but admits capacity constraints still exist. In 2025, the division added 3.9 GW of new electrical capacity and expects to double its total capacity before the end of 2027. Meanwhile, AWS closed Q4 2025 with 24% year-over-year growth and an annualized revenue rate of approximately $142 billion.
Graviton: From Internal Bet to Scarce Resource
Launched in 2018, Graviton was Amazon’s first major processor designed for its own cloud. Since then, the family has reached its fifth generation. According to the company, over 90,000 customers are now using it, and it’s present in 98% of the top 1,000 EC2 customers, AWS’s on-demand compute service. Amazon announced Graviton5 in December 2025, describing it as its most powerful CPU to date, with up to 25% more performance than the previous generation in certain scenarios.
This detail helps explain why Jassy’s statement is significant. For months, market conversations have focused on Nvidia’s GPUs and the shortage of AI accelerators. However, Amazon’s letter makes it clear that the bottleneck affects not just training large models but also overall computing capacity supporting enterprise applications, databases, analytics, inference, and large-scale cloud services. The fact that two clients sought to reserve all the yearly Graviton instances demonstrates that AWS’s proprietary processors have shifted from being a supplementary asset to a critical strategic resource within their portfolio.
Furthermore, Amazon is working to extend this strategy beyond CPUs. In the same letter, Jassy notes that Trainium2 has been nearly exhausted, Trainium3 began shipping in early 2026 and is almost fully booked, and a significant portion of Trainium4 is already reserved despite still being roughly 18 months away from general availability. Combining Graviton, Trainium, and Nitro, Amazon’s chip business generates over $20 billion annually, with triple-digit year-over-year growth. The company even hints at the possibility of selling racks of these systems to third parties in the future.
The Race for Capacity and Power
The core of this story isn’t just CPU design but the speed at which physical infrastructure must be deployed. Amazon plans to invest around $200 billion in capex in 2026, affirming that this isn’t “just intuition,” but based on commitments already signed or highly advanced with customers. A significant portion of this expenditure, it claims, will be monetized between 2027 and 2028. The company also emphasizes that the assets it is building have long usable lives—over 30 years for data center buildings and 5 to 6 years for hardware like chips, servers, and networking equipment.
This explanation matters because the market has been asking for months whether AI spending is accelerating too rapidly. Amazon’s response is that demand is already there and that its proprietary silicon could change the economic landscape of cloud business. Jassy asserts that Trainium could save the company tens of billions of dollars annually in capital investment and improve operating margins by hundreds of basis points compared to relying entirely on third-party chips for inference. These are forecasts from the company itself, not final results, but they help illustrate why Amazon is willing to make such substantial investments.
More Spending Today to Protect Margins Tomorrow
The financial impact is already visible. In its 2025 annual report, Amazon reported that free cash flow declined from $38.2 billion to $11.2 billion, primarily due to a $50.7 billion increase in capex related to purchases of fixed assets, net of sales and incentives. The company links this deterioration mainly to investments in AI-related infrastructure. Essentially, this is the price of staying ahead of competitors in computing capacity, energy, and data center deployment.
In the short term, this pressures cash flow. In the medium and long term, Amazon believes that the combination of AWS, AI, and proprietary chips will ultimately expand revenue, margins, and cash generation. The thesis isn’t without risks: it requires continued demand growth, timely monetization, and flawless execution of a massive expansion. But Jassy’s scene—two large clients trying to secure all the annual Graviton capacity—captures why Amazon believes this risk is worth taking. In the cloud business, scarcity isn’t just measured in chips; it’s measured in time, electricity, and ready-to-use capacity.
Frequently Asked Questions
What is AWS Graviton and what is it used for?
AWS Graviton is Amazon’s family of processors designed for its cloud platform. It’s used in EC2 instances, primarily supporting enterprise applications, databases, analytics, web services, and increasingly, tasks related to AI and inference. Amazon launched it in 2018 and states it’s currently in its fifth generation.
What does it mean that two clients wanted all Graviton capacity for 2026?
It means two major AWS clients requested to reserve the entire annual supply of Graviton-based instances for that year. Amazon declined because it needed to distribute capacity among more clients, but the request highlights very high demand for its proprietary infrastructure.
How much will Amazon invest in AI and data centers in 2026?
Amazon expects around $200 billion in capital expenditure in 2026. The company asserts that much of this spending is backed by customer commitments and will be monetized mainly between 2027 and 2028.
How do Graviton and Trainium differ within AWS?
Graviton is Amazon’s general-purpose CPU line for the cloud. Trainium, on the other hand, is its family of chips specifically designed for AI workloads, especially training and inference. In the shareholder letter, Amazon presents both as key assets in reducing costs and decreasing dependence on external providers.