How Amazon turned custom chips into a $25 billion-per-year AWS advantage in a decade

Amazon has spent the better part of a decade building purpose-built chips for its cloud business, and the company says the effort has now crossed a major threshold: its custom silicon business is running at more than $25 billion in annual revenue, with triple-digit year-over-year growth. The result is a deeper stack for AWS that targets not just raw computing speed, but also cost and energy efficiency for AI workloads that are increasingly expensive to run at scale.

The story is told through a generation-by-generation chronology of Amazon’s chip roadmap. Trainium is Amazon’s chip family built for AI training and inference, the two compute-heavy stages of AI use. Training is the work of teaching AI models, while inference is running those models at scale for real-world applications. Amazon says Trainium3 delivers up to 40% better price-performance than Trainium2, framing the upgrade in terms of dollars spent per unit of useful output.

Alongside Trainium is Graviton, which Amazon positions as its general-purpose cloud computing platform. The company says Graviton serves 98% of the top 1,000 EC2 customers, with up to 40% better price-performance than its predecessor generation. (EC2 is Amazon Web Services’ elastic compute service, where customers rent virtual servers for applications, databases, and other workloads.) Amazon also links Graviton to newer “agentic AI” workloads, describing agents as systems that can plan and take actions as part of software operations.

Amazon says Graviton5, its latest generation mentioned in the release, delivers up to 25% better performance than Graviton4, and is growing nearly twice as fast as the earlier transition. The company also says revenue commitments for Graviton have increased nearly three times quarter over quarter, suggesting customer demand that is translating into larger forward purchase or deployment expectations.

For cloud infrastructure rather than compute alone, Amazon points to Nitro, which it describes as the networking, storage, and security layer behind AWS cloud infrastructure. Nitro is positioned as the connective tissue that helps the overall platform deliver performance and security while offloading certain functions away from general-purpose CPU cycles.

The release also ties the chips to specific customers and AI models. Anthropic, for example, has committed to using up to five gigawatts of current and future Trainium capacity to train and power its Claude models, which Amazon says are already running on more than one million Trainium2 chips. Anthropic is also described as using tens of millions of Graviton cores for scalable performance and cost efficiency across broad generative AI workloads.

Amazon adds that OpenAI has committed to consuming two gigawatts of Trainium capacity through AWS infrastructure to power its frontier models beginning in 2027. (Frontier models are large, state-of-the-art AI systems trained on extensive data and compute.) Meta is also described as signing an agreement to deploy tens of millions of Graviton cores for the CPU-intensive workloads behind its agentic AI efforts. Uber’s applications are described more operationally: Amazon says Uber uses Graviton to match riders with drivers in fractions of a second, and is piloting Trainium3 to train AI models that aim to make each ride smarter.

On the hardware scaling side, Amazon highlights integrated systems designed to move fast from research training to deployment. It describes Trn3 UltraServers as configurations that pack up to 144 Trainium3 chips into a single integrated system, delivering up to 4.4 times more compute performance than Trainium2 UltraServers, which Amazon says can cut model training time from months to weeks. It also points to Project Rainier as its world-scale AI computing cluster built for training frontier models at scale, and says it is running Anthropic’s Claude models.

Efficiency is a throughline in the company’s pitch. Amazon says Trainium3 delivers over five times higher output tokens per megawatt than Trainium2. “Output tokens” are units of generated text (or similar discrete elements) produced by an AI system, and “per megawatt” is a measure of how much output can be generated for each unit of electrical power. In AWS’s framing, that matters both for cost and for environmental impact as data centers scale.

Still, the release leaves several specifics open. While Amazon cites $25 billion-plus annual revenue run rate and growth rates, it does not break out how that revenue is distributed between Trainium, Graviton, Nitro, or other related services, nor does it provide a detailed view of margins. It also does not quantify how quickly the mentioned capacity commitments translate into deployed revenue versus future run-rate, and it does not disclose the full terms or timelines behind each gigawatt agreement beyond the start date referenced for OpenAI’s 2027 commitment.

For customers and the market, the near-term watch items are how quickly newer generations like Trainium3 and Graviton5 convert into measurable workload performance and price improvements, and whether the “capacity commitment” model spreads from flagship AI labs to a broader set of enterprise customers. Amazon’s emphasis on large-scale training infrastructure like Project Rainier also suggests a continuing shift in how AWS competes for frontier AI development, not just on software and services but on the underlying compute economics. Investors and cloud buyers alike will likely look for the next update on revenue contribution, customer mix, and the pace of deployment for the newest systems.