Amazon’s chip push grows strategic
Andy Jassy’s annual letter framed Amazon’s custom‑silicon effort (Graviton, Trainium, Nitro) as a fast‑growing business that could be worth tens of billions and hinted at the possibility of external chip sales. That signals hyperscalers are becoming vertically integrated hardware platforms, not just cloud vendors renting generic compute. For product teams, the shift means infrastructure selection will increasingly involve hardware roadmaps and vendor incentives, not only API features. (thenextweb.com)
Amazon just said its in-house chip business is already running at more than $20 billion a year, even though Amazon mostly sells those chips only indirectly through Amazon Web Services cloud instances. Andy Jassy put Graviton, Trainium, and Nitro in the same bucket and said that bucket is still growing at triple-digit year-over-year rates. (aboutamazon.com) That is unusual because cloud companies usually rent you computing time built on somebody else’s processors. Amazon is saying a meaningful slice of its cloud economics now comes from silicon it designed itself. (aboutamazon.com) Graviton is Amazon’s central processor family, the general-purpose brain that runs ordinary cloud workloads like web apps, databases, and back-end services. Amazon says Graviton-based instances deliver up to 40% better price performance than comparable x86 instances. (aws.amazon.com) Trainium is Amazon’s artificial intelligence training chip, built for the expensive job of teaching large models from huge piles of data. Amazon says Trainium2 delivers up to 4 times the performance of first-generation Trainium and offers 30% to 40% better price performance than certain graphics processing unit instances in Amazon Elastic Compute Cloud. (aws.amazon.com) Nitro is the less famous piece, but it sits under almost everything Amazon Web Services sells. Amazon says Nitro moves virtualization and input-output work onto dedicated hardware, which frees the main processor and reduces the attack surface. (aws.amazon.com) Put those three together and Amazon is no longer just renting servers. It is designing the processor, the artificial intelligence accelerator, and the control layer that sits between customer software and the machine. (aboutamazon.com) (aws.amazon.com) Jassy also gave a clue about where this could go next. He wrote that the current run rate is “somewhat understated” because Amazon is “currently only monetizing” its chips through Amazon Elastic Compute Cloud, which is close to saying there are other ways to sell them later. (aboutamazon.com) If Amazon ever sells those chips outside its own cloud, it would be moving closer to the model used by companies that build both the factory floor and the machines on it. A customer would not just be choosing a cloud provider’s software services, but a hardware stack tied to that provider’s roadmap. (aboutamazon.com) (aws.amazon.com) The money behind this is large enough to matter inside Amazon. Amazon reported $128.7 billion in Amazon Web Services sales and $45.6 billion in Amazon Web Services operating income for 2025, so a $20 billion-plus silicon run rate is no side project inside the cloud division. (ir.aboutamazon.com) Amazon is also using chips to lock in flagship artificial intelligence customers. Amazon says Anthropic is building Project Rainier on nearly 1 million Trainium2 chips, which turns custom silicon from a cost-saving tool into a way to win and hold giant model builders. (aws.amazon.com) That changes the cloud buying decision for everyone else too. When one vendor controls the chip, the server design, and the cloud contract, the pitch is no longer just “our application programming interface is nicer,” but “our hardware gets you lower cost, more capacity, and better priority on the next generation.” (aws.amazon.com) (aboutamazon.com)
