AWS’ Peter DeSantis says next Trainium chip will be ‘very differentiated’

AWS is not playing catch-up in the AI chip race. It is trying to lap the field.

Peter DeSantis, the AWS senior vice president overseeing AI models, custom silicon, and quantum projects, said the next Trainium chip will be highly differentiated. The comment lands at a moment when the company’s existing Trainium3 accelerator is already sold out, and early reservations for its successor are piling up fast.

For context: AWS has deployed more than one million Trainium chips.

What Trainium3 and Trainium4 actually mean on paper

Trainium3 became generally available on December 2, 2025. Compared to its predecessor, Trainium2, it delivers up to 4.4 times higher compute performance and roughly four times greater memory bandwidth, alongside improved energy efficiency.

Trainium4, currently in development, is targeting even more ambitious numbers. AWS projects it will deliver six times the FP4 compute performance of Trainium3, with memory bandwidth again quadrupling. Broad availability is expected around 2027.

FP4 refers to a lower-precision floating-point format increasingly used in large-scale AI model training. Higher throughput at FP4 means faster training runs at lower cost, which is exactly what hyperscalers and enterprise AI teams are chasing right now.

AWS is selling silicon, not just renting it

AWS reported triple-digit growth in its custom silicon segment, with annualized sales expected to reach into the tens of billions of dollars.

Trainium3 capacity is largely sold out. Trainium4 is accumulating substantial early reservations despite not launching until 2027.

AWS is also contemplating selling Trainium chips directly to external parties, rather than only offering them through its own cloud infrastructure. The timing aligns with a broader trend: sovereign AI initiatives, where governments and large organizations want to build AI infrastructure on hardware they own or control.

The Nvidia question nobody can stop asking

Nvidia’s H100 and B200 GPUs remain the default choice for large-scale AI training. Nvidia’s CUDA software ecosystem, built over nearly two decades, is the main reason: millions of developers know it, most AI frameworks are optimized for it, and switching costs are real.

AWS is addressing the software side through its Neuron SDK and the Neuron Kernel Interface, tools designed to lower the barrier for developers moving workloads onto Trainium hardware.