How Amazon Is Building Its Own AI Silicon

Amazon is best known for retail and cloud services, but one of its most important strategic bets is made of silicon. Over roughly a decade, Amazon Web Services has built an in-house chip business that now covers networking, general-purpose computing, and AI training and inference. For more on how the hyperscalers are racing to control their own hardware, see our AI chip and semiconductor coverage.

It began with an acquisition

Amazon’s chip story starts with Annapurna Labs, an Israeli startup founded in 2011. According to Wikipedia’s profile, Amazon acquired the company in January 2015, and it now serves as Amazon’s semiconductor division, responsible for the Nitro, Graviton, and Trainium product lines.

The order of those products tells you the strategy. Nitro offloaded networking and security work from servers. Graviton, an Arm-based CPU built for exclusive use by AWS, brought general-purpose compute in-house. Only after mastering those did Amazon turn to AI accelerators. AWS has since launched Inferentia for inference and Trainium for training, with Trainium2 designed to scale to over 100,000 chips in a single training cluster.

Why build instead of buy

The core logic mirrors what Google and Microsoft concluded. Training and serving large models on Nvidia GPUs means paying premium prices and depending on someone else’s supply. Amazon’s own engineers say that building their own devices lets them optimize across the entire stack, which shortens engineering time and the time to reach massive scale.

Cost is the other driver. Analysts covering Anthropic’s AWS deal say the Trainium chips give it a cost advantage over rivals such as OpenAI, which has leaned heavily on expensive Nvidia GPUs. One industry roundup estimates custom silicon’s total-cost-of-ownership advantage over conventional GPUs at up to 65 percent for inference at production scale. We looked at the same economics from the Azure side in our piece on why Microsoft is building its own AI chips.

Project Rainier: proof at scale

The clearest demonstration is Project Rainier, the cluster AWS built for Anthropic. According to Data Centre Magazine, it uses nearly 500,000 Trainium2 chips, represents a 70 percent increase over AWS’s previous AI infrastructure, and gives Anthropic more than five times the compute it used to train earlier model versions. The main site is an Indiana campus that Capacity describes as an $11 billion AWS investment.

This matters because it shows custom silicon running frontier-model training in production, not only in benchmarks. It also shows the partnership model: the chip was shaped by a demanding customer who gave direct feedback on speed, latency, and energy use.

Trainium3 and Trainium4

The roadmap keeps moving. Trainium3 is AWS’s first chip on a 3-nanometer process and was co-developed with Anthropic, according to one analysis of the Indiana campus; AWS has said it gives better performance, lower latency, and better power consumption per unit of compute. The same analysis expects Trainium4 to begin delivering in 2027 with roughly six times the FP4 performance of Trainium3, though that figure comes from a single secondary source, so treat it as an estimate.

The next generation is already being reserved. In April 2026, Amazon and Anthropic announced that Anthropic will commit more than $100 billion to AWS technologies over ten years and secure up to 5 gigawatts of capacity, covering Trainium2, Trainium3, Trainium4, and future generations. A chip-focused guide also reports that OpenAI agreed in February 2026 to use about 2 gigawatts of Trainium3 and Trainium4 capacity, which would make Trainium a platform for more than one frontier lab.

The infrastructure behind the chips

Silicon alone doesn’t make an AI cloud. Amazon is also designing the racks, networking, and power systems around its chips, which is the same shift we covered in why AI is changing the design of data centers. As rack power climbs and clusters grow to hundreds of thousands of accelerators, owning the chip, the server, and the facility design together lets AWS tune the whole system instead of fitting parts from different vendors.

What could go wrong

Custom chips come with real risks. Chip schedules can slip, and the same analysis that describes the Anthropic deal warns that Anthropic’s competitive position could suffer if Trainium3 misses its 2026 ramp or Trainium4 arrives late. Software maturity also matters: developers are used to Nvidia’s tools, so AWS has to keep making Trainium easy to adopt. And Nvidia isn’t standing still, so Amazon has to keep pace with a moving target.

The bigger picture

Amazon’s approach combines a long head start with a customer-led strategy. By buying Annapurna Labs early, building Nitro and Graviton first, and then co-designing Trainium with a major AI lab, AWS turned chip design into a core cloud advantage. Combined with Google’s TPUs and Microsoft’s Maia, it points to an industry where the largest cloud providers increasingly design the hardware their AI runs on. Follow our data center and AI hardware coverage for the latest on each of them.