When Amazon.com acquired chip design firm Annapurna Labs in 2015, the AI boom was still years away. The bet was clear: designing chips specifically for cloud workloads, rather than relying on general-purpose hardware, would deliver higher performance at a lower cost.
A decade later, this wager has become one of Amazon's fastest-growing business segments. The company's chip division – encompassing Trainium for AI training and inference, Graviton for general cloud computing and increasingly agentic AI, and the Nitro system for security and networking – recently surpassed an annual revenue run rate of $25 billion, achieving triple-digit year-over-year growth.
"Given that we have leading price-performance chips in both AI through Trainium and CPU through Graviton, we are extremely well positioned for this AI inflection point," said Andy Jassy, Amazon's CEO.
The following timeline details Amazon's journey across multiple generations of custom chips.
Why Amazon Chose Custom Chip Development
Historically, the AI chip market was slow to reduce costs, and industry leaders rarely disrupted themselves. Amazon chose a different path. Instead of relying entirely on off-the-shelf processors, the company designed chips from scratch for specific workloads. This approach enabled deep vertical integration: hardware and software engineers worked closely from chip design to server deployment, using a "working backwards" method to create chips perfectly aligned with customer workloads.
The result is three complementary chip families fulfilling distinct roles: Trainium, built for AI training and inference, handling the intense computations required to train models and run them at scale; Graviton, which powers general cloud computing, including websites, applications, databases, and increasingly agentic AI workloads, with 98% of Amazon's top 1,000 EC2 customers using it and revenue commitments growing nearly three times quarter-over-quarter, with the latest Graviton5 offering up to 25% better performance than its predecessor and growing nearly twice as fast as Graviton4; and Nitro, which drives the networking, storage, and security behind AWS's cloud infrastructure.
As CEO Andy Jassy noted during the company's Q2 2026 earnings call: "Our chip business now has an annualized revenue run rate of over $25 billion. Given that we have leading price-performance chips in both AI, through Trainium, and CPU, through Graviton, we are extremely well positioned for this AI inflection point. Beyond the world's top two AI labs, Anthropic and OpenAI, who have signed multi-year, multi-gigawatt agreements with Trainium, an increasing number of AI startups are also adopting Trainium."
What Amazon Chips Mean for AI Customers
For businesses building with AI, chip economics are critical. Training a frontier AI model can require hundreds of thousands of chips running for weeks or months. Running that model at scale, handling millions of queries daily, demands efficient inference infrastructure. Even small improvements in price-performance translate into massive cost savings at scale.
This is why leading AI labs and tech companies are making major commitments to Amazon's chips: Anthropic has committed to using up to 5 gigawatts of current and next-generation Trainium chips to train and run its Claude model, operating on over one million Trainium2 chips, while also using tens of millions of Graviton cores for scalable performance and cost efficiency across broad generative AI workloads. OpenAI has committed to using 2 gigawatts of Trainium compute power via AWS infrastructure starting in 2027 to power its frontier models. Meta has signed an agreement to deploy tens of millions of Graviton cores for CPU-intensive workloads behind its agentic AI. Uber is using Graviton to match passengers with drivers in milliseconds and piloting Trainium3 to train AI models that make every trip smarter.
AI Chips Built for Unprecedented Customer Demand
Amazon Bedrock, the company's managed AI service for hundreds of thousands of customers, runs most of its inference on Trainium. A growing number of AI startups are also adopting Trainium, including unicorns like Neura Robotics, which chose Trainium for physical AI, and Odyssey, which uses Trainium to build world models simulating physics, delivering nearly twice the useful compute per dollar compared to alternatives. Other adopters include TwelveLabs, DeCart, Poolside, Karakuri, Matagenomi, NetoAI, Splash music, and larger companies like Uber and Pinterest.
On the Graviton side, over 130,000 customers now use Graviton-based servers. More than half of all new compute capacity added by AWS runs on Graviton chips. This demand reflects a shift in how AI infrastructure is built. As AI systems move from answering questions to taking actions, including real-time reasoning, code generation, and multi-step task orchestration, the required compute power places high demands on both AI accelerators and CPUs. Amazon is well-positioned in both areas.
But single chips are only part of the story. Amazon connects them into increasingly powerful systems: Trn3 UltraServers pack up to 144 Trainium3 chips into an integrated system, delivering up to 4.4 times the compute performance of Trainium2 UltraServers, allowing customers to complete model training in weeks instead of months. Project Rainier, the world's largest AI compute cluster, is built for training frontier models at scale and is currently running Anthropic's Claude model. In terms of energy efficiency, Trainium3 delivers over five times more tokens per megawatt than Trainium2, an efficiency that is critical for both cost and environmental impact at scale.
AI, Chips, and Quantum Computing: The Future of Amazon's Chip Business
Amazon's chip roadmap is accelerating. Trainium4 is already in development. Graviton continues to evolve for the age of agentic AI. Under the leadership of Peter DeSantis, who now oversees a unified organization spanning AI models, custom chips, and quantum computing, Amazon is fusing these technologies together to reinforce each other.
"One of the ways we get an advantage in building these models is by leveraging our deep investment in chips to provide performance and cost efficiency that will differentiate our model development," said Peter DeSantis, Amazon's senior vice president of AI models, chips, and quantum computing.
This strategic bet rests on a simple premise: the company that controls its own chips will dominate the pace of AI development. With a portfolio spanning training, inference, and general computing, and a roadmap that has already attracted strong customer interest years in advance, Amazon is building the infrastructure layer that will support the next generation of AI.
Comments