4/ Our chip business is booming, changing the economics of AWS and becoming much larger than many think.. Nearly all AI to date has been run on NVIDIA chips, but a new shift has begun. We have a strong partnership with NVIDIA, and we continue to have customers who choose to run NVIDIA, making AWS the best place to run NVIDIA. However, customers are demanding better value for money. We’ve seen this movie before. In the CPU space, virtually all workloads ran on Intel chips until Graviton was invented in 2018. Graviton offers up to 40% better price performance than other x86 processors and is now widely used by 98% of the top 1,000 EC2 customers. The same story plays out with AI. The second version of the custom AI silicon (Trainium2) offers approximately 30% better price-performance than comparable GPUs and is mostly sold out. Trainium3 just started shipping in early 2026 and is nearly full with 30-40% better price performance than Trainium2. A good portion of Trainium4 is already pre-ordered, even though it’s still about 18 months away from general release. Additionally, Amazon Bedrock, AWS’s leading (and very fast-growing) inference service, runs most of its inference on Trainium. Demand for Trainium is skyrocketing.
