An interview with Amazon's AI chief Peter DeSantis on plans to use in-house chips, Trainium and Inferentia, to develop AI models more cheaply, and more
Amazon's new artificial intelligence czar, Peter DeSantis, is a larger-than-life figure within the company where he has worked since its early days
Context & Ripple Effects
Amazon had already positioned AI chips as a meaningful part of AWS’s generative-AI strategy in an earlier AWS leadership discussion of generative AI and chips. DeSantis’s remit puts the emphasis on whether that infrastructure can lower the cost of building Amazon’s own models, not merely support cloud capacity.
The strategy later gained a potential external dimension when Amazon entered talks to sell Trainium chips for third-party data centers. That progression makes this interview an early indicator of custom silicon moving from an internal efficiency program toward a possible product business.
First-order effects
- Amazon can orient model-development workloads around Trainium and Inferentia where they meet its needs, making chip choice a direct lever on the cost of developing AI models.
- DeSantis’s role consolidates accountability for linking Amazon’s AI-model ambitions with its in-house chip roadmap.
Second-order effects
- AWS’s custom-chip effort gains strategic weight: better internal economics would strengthen the case for offering customers alternatives to externally sourced AI hardware.
- Rival cloud platforms and chip suppliers face added pressure to demonstrate comparable price-performance or tighter integration with AI-model development workflows.
Third-order effects
- If custom chips prove useful both inside Amazon and in outside data centers, AI infrastructure could become more vertically integrated, with cloud operators competing through silicon, models, and deployment economics together.
- The key uncertainty is portability: a durable market shift depends on whether customers can adopt these chips without accepting costly workload or software trade-offs.
The trend: This is part of AI industrialization, in which major cloud providers seek to turn proprietary silicon into a lower-cost foundation for model development and inference.