How Amazon.com, Inc. (AMZN.US) Became One of the World's Top Chip Companies in Ten Years

When Amazon.com, Inc. (AMZN.US) acquired the chip design company Annapurna Labs in 2015, the AI boom was still years away. The bet was clear at the time: chips designed specifically for cloud workloads, rather than relying on general-purpose hardware, could deliver higher performance at a lower cost. Different Bets on AI Infrastructure A decade later, this bet has become one of the fastest-growing businesses in Amazon.com, Inc.s history. Amazon.com, Inc.'s chip businesscovering Trainium for AI training and inference and Graviton for general cloud computing (as well as the increasingly prevalent agentic AI)both based on the Nitro system for security and networkinghas recently surpassed an annual revenue run rate of $25 billion, achieving triple-digit year-over-year growth. Given our leading cost-performance chips in both AI (through Trainium) and CPUs (through Graviton), we are in an extremely favorable position in this AI inflection point. Andy Jassy, CEO of Amazon.com, Inc. The following timeline details the development of Amazon.com, Inc.'s generations of dedicated chips. Why Amazon.com, Inc. Chose to Develop Its Own Chips Historically, the AI chip market has been slow to reduce costs, and industry leaders rarely self-disrupt. However, Amazon.com, Inc. chose a different path. Rather than fully relying on off-the-shelf processors, Amazon.com, Inc. designed chips from the ground up for specific workloads. This approach enabled deep vertical integration: hardware and software engineers collaborated closely throughout the entire process, from chip design to server deployment, using a Working Backwards method to reverse engineer systems and create chips that perfectly matched the actual workloads of customers. The result is three major chip series that serve different but complementary roles: Trainium: Designed specifically for AI training and inferencei.e., training AI models and running these models at scale in high-intensity computing tasks. Graviton: Handling general cloud computingi.e., powering modern software's websites, applications, databases, and increasingly agentic AI workloads. Among our top 1,000 EC2 customers, 98% are using Graviton, and its revenue commitments have nearly tripled quarter-over-quarter. The latest generation, Graviton5, offers up to a 25% performance improvement over its predecessor, with a growth rate nearly twice that of Graviton4. Nitro: Powering the networking, storage, and security behind the AWS cloud infrastructure. As CEO Andy Jassy pointed out during the company's Q2 2026 earnings call, Our chip business's annual revenue run rate has now exceeded $25 billion. Given our cost-effective chips in both AI (through Trainium) and CPUs (through Graviton), we are in a very advantageous position in this AI inflection point. In addition to Anthropic and OpenAI, two of the worlds leading AI labs signing multi-year cooperation agreements involving gigawatts of power with Trainium, an increasing number of AI startups are also adopting Trainium. What Amazon.com, Inc. Chips Mean for AI Customers For enterprises building with AI, chip economics are critical. Training a cutting-edge AI model may require hundreds of thousands of chips running continuously for weeks or months. Running that model at scaleresponding to millions of queries dailyrequires efficient inference infrastructure. Even a slight improvement in cost-performance can significant savings at scale. This is why top global AI labs and tech companies are making significant commitments to Amazon.com, Inc.'s chips: Anthropic has committed to using up to 5 gigawatts (GW) of current and next-generation Trainium chips to train and power its Claude model, which runs on over one million Trainium2 chips. Anthropic is also utilizing tens of millions of Graviton cores to deliver scalable performance and cost efficiency across a wide range of generative AI workloads. OpenAI has committed to 2 gigawatts of Trainium compute to power its cutting-edge models via AWS infrastructure starting in 2027. Meta has signed an agreement to deploy tens of millions of Graviton cores to support the CPU-intensive workloads behind its agentic AI. Uber is using Graviton to match passengers with drivers in milliseconds and is piloting the use of Trainium3 to train AI models for smarter rides. AI Chips Built to Meet Unprecedented Large-Scale Customer Demands Amazon Bedrock is the companys managed AI service offering to hundreds of thousands of customers, with most of its inference work running on Trainium. An increasing number of AI startups are also adopting Trainium, including unicorns like Neura Robotics (which chose Trainium for physical AI) and Odyssey (which chose Trainium to build world models that simulate physical laws, with cost-effective computing power nearly twice that of other alternatives). Other startups include TwelveLabs, DeCart, Poolside, Karakuri, Matagenomi, NetoAI, and Splash music, as well as larger companies like Uber and Pinterest. On the Graviton front, over 130,000 customers are now using Graviton-based servers. More than half of the new capacity in AWS is running on Graviton chips. This demand reflects a shift in how AI infrastructure is built. As AI systems transition from answering questions to taking actioni.e., real-time inference, code generation, and multi-step task orchestrationthe required computing power places high demands on both AI accelerators and CPUs. Amazon.com, Inc. holds a favorable position in both areas. But a single chip is only part of the story. Amazon.com, Inc. is connecting them into increasingly powerful systems: Trn3 UltraServers: Packaging up to 144 Trainium3 chips into an integrated system, delivering up to 4.4 times the computing performance of Trainium2 UltraServers. This enables customers to complete model training in weeks instead of months. Project Rainier: The world's largest AI compute cluster, designed for the large-scale training of cutting-edge models. Project Rainier is currently running Anthropics Claude model. Energy Efficiency: Trainium3 produces more than five times the number of tokens per megawatt compared to Trainium2this efficiency is crucial in large-scale applications for both cost and environmental impact. The Future of AI, Chips, and Quantum Computing at Amazon.com, Inc. Amazon.com, Inc.s chip roadmap is continuing to accelerate. Trainium4 is in development. Graviton will continue to evolve in the era of agentic AI. Under the leadership of Peter DeSantis, who now oversees a new organization encompassing AI models, self-developed chips, and quantum computing, Amazon.com, Inc. is bringing these technologies together to reinforce one another. One of the ways we gain advantages in building these models is by leveraging our deep investments in chips to provide performance and cost benefits that will make our model development stand out. Peter DeSantis, Senior VP of AI Models, Chips, and Quantum Computing, Amazon.com, Inc. This strategic bet rests on a simple premise: companies that master self-developed chips will dictate the pace of AI development. With a chip portfolio covering training, inference, and general computing, and a roadmap that has drawn strong customer interest years in advanceAmazon.com, Inc. is building the foundational infrastructure for the next generation of AI operations. This article is reprinted from the "Semiconductor Industry Observer" WeChat public account, GMTEight editor: Xu Wenqiang.