Amazon.com (AMZN.US) CEO Andy Jassy expressed strong optimism about the company's internal infrastructure stack in a letter to shareholders focusing on its custom silicon operations. Hyperscale data center operators are currently confronting a significant "compute resource shortage" crisis, as existing infrastructure struggles to meet the massive demand driven by artificial intelligence and related services. This has prompted giants like Amazon.com, Alphabet, and Meta Platforms, Inc. to increasingly turn to custom chips to address the computational gap. Amazon.com's Trainium AI chips and Graviton server CPUs have demonstrated maturity and reliability, which underpins CEO Andy Jassy's confidence.
Jassy communicated to shareholders that Amazon.com's custom chip business is expanding rapidly and is projected to achieve an Annual Recurring Revenue (ARR) of $50 billion. This figure substantially exceeds the scale of businesses at AMD and Intel. The CEO stated that possessing internally developed, market-favored AI chips unlocks numerous possibilities, most notably the potential to reduce customer costs and improve the economics for AWS. It is estimated that scaling Trainium chips could result in annual capital expenditure savings of tens of billions of dollars and provide an operating margin advantage of hundreds of basis points compared to relying on other chips for inference.
The $50 billion ARR projection estimates the scale of Amazon.com's custom chip business if it were spun off as an independent entity, operating similarly to NVIDIA as a compute service provider. The growth of Trainium and Graviton is primarily fueled by AWS cloud services, and Amazon.com has not yet made its custom chips available to users who simply need computing power. Jassy indicated that the economic benefits delivered by AWS and its ASIC chip solutions are superior to those offered by NVIDIA's alternatives. While he reiterated support for NVIDIA, he also pointed out that customers demand options that dominate in "price-performance," an area where he believes Trainium excels.
The Amazon.com CEO also made indirect references to Intel and its CPU market share, noting that following the launch of the ARM-based Graviton chips, AWS infrastructure is now almost entirely dominated by them. Jassy suggested a similar dynamic is unfolding with Trainium for both training and inference. This hints at an underlying strategic argument that the industry may not have fully appreciated: the concept of custom infrastructure is not meant to replace mainstream computing solutions, such as those from NVIDIA, but rather to fill a massive compute void that makes it impossible for hyperscale data centers like Amazon.com's to rely solely on GPU manufacturers.
Interestingly, Amazon.com has also hinted at the potential to sell its server racks to third-party clients, suggesting it could directly compete with companies like NVIDIA. The company plans to invest "hundreds of billions of dollars" in capital expenditure for this business, making the future trajectory of the Trainium + Graviton technology stack a key area to watch.
**Custom Chips Built for the Cloud Era** Traditional server chips were not designed for cloud computing. AWS Graviton is a family of processors specifically engineered for the cloud, aimed at helping businesses run applications faster and more efficiently. Since 2018, Amazon.com has launched five generations of Graviton chips, with each iteration building upon the last. The company identified an opportunity: by creating chips designed specifically for the cloud, it could offer customers solutions with higher performance and lower costs. This led to the birth of Graviton, Amazon.com's first custom cloud processor.
The latest Graviton5 chip boasts 192 cores, representing the highest processing capacity integrated into a single chip by Amazon.com to date. It delivers up to a 25% performance improvement over Graviton4 while maintaining leading energy efficiency.
**Five Generations of Innovation** * **Graviton (2018):** Amazon.com launched its first custom cloud processor, proving that the energy-efficient chip technology behind smartphones could handle serious cloud workloads. * **Graviton2 (2019):** The second generation marked a significant leap, offering 40% better price-performance than comparable processors, substantially reducing costs for companies running their applications. * **Graviton3 (2021):** The third generation consumed 60% less energy while delivering similar performance to comparable Amazon EC2 instances, helping customers boost performance while reducing environmental impact. * **Graviton4 (2023):** Graviton4 tripled the number of processing cores, enabling it to handle more demanding workloads like large databases and analytics faster and more efficiently. * **Graviton5 (2025):** The newest generation doubles the core count to 192, providing up to 25% better performance than Graviton4, capable of powering the most demanding applications from real-time gaming to AI.
**Powering Everyday Digital Experiences** When you request an Uber, you might not think about the processor matching you with a driver. But the company uses Graviton to power its trip service area system, which processes millions of prediction and location data points within milliseconds to match drivers and passengers in real-time. Graviton's performance ensures shorter wait times and reliable service, even during peak hours.
In Formula 1 racing, while drivers speed around the track at over 200 mph, Graviton processors work behind the scenes, transforming real-time racing data into the graphics and statistics viewers see on TV broadcasts. Companies like Pinterest and Airbnb have also migrated to Graviton, gaining performance improvements while lowering costs and carbon emissions.
Today, 98% of the top 1,000 Amazon EC2 customers use Graviton chips. Over half of all new AWS compute capacity runs on these processors. Businesses leverage Graviton for web applications, databases, big data analytics, game servers, video encoding, machine learning, and high-performance computing. AWS utilizes custom chips, intelligent routing systems, and specialized infrastructure to support "inference" workloads, making AI faster and more affordable.
From the apps on your phone to streaming services on your TV, Graviton is quietly powering more of your digital life—making it faster, more economical, and more sustainable.
Comments