Spend versus return cuts both ways: one way to improve the return on an AI buildout is to own the chip. Amazon's 2026 DEF 14A (filed 2026-04-09) describes "purpose-built silicon chips that advance innovations in power efficiency, like the AWS Graviton chip, AWS Trainium chip, and AWS" Inferentia chip. EdgarBeast surfaced the disclosure.
The prior year's proxy is more explicit about the why. Amazon's 2025 DEF 14A (filed 2025-04-10) states that for "running complex AI workloads like large language models, AWS has developed purpose-built silicon like the AWS Trainium chip and AWS Inferentia chip." The chips are named, and the workload — large language models — is named with them.
The economic logic is the capex story in reverse. Every accelerator a hyperscaler buys from a merchant vendor is margin handed to that vendor. Custom silicon — Trainium for training, Inferentia for inference — is Amazon's disclosed attempt to keep more of that margin in-house and to control power efficiency, which is itself a major data-center cost.
The proxies disclose the existence and purpose of the chips; they are not financial statements and do not quantify the savings or the volume. So read this as strategy-of-record, not a revenue line. The 2026 proxy on sec.gov and 2025 proxy are primary; EdgarBeast is the evidence index.