Amazon CEO Andy Jassy affirms long-term reliance on Nvidia chips even as in-house AI silicon scales past $20B

Amazon’s Andy Jassy wants to make one thing very clear: Nvidia isn’t going anywhere. In his April 2026 letter to shareholders, the Amazon CEO stated that “virtually all AI thus far has been done on NVIDIA chips” and affirmed that AWS will remain the best place to run them.

The words carry weight because they address what has quietly become Nvidia’s most discussed risk factor: that its biggest customers are also its future competitors. Amazon, Google, Microsoft, and Meta have all poured resources into custom AI accelerators.

The numbers behind the commitment

In August 2026, Amazon agreed to buy an additional 2 million Nvidia GPUs for deployment across AWS data centers in the 2027-2028 timeframe. That order sits on top of a previous commitment for 1 million chips, bringing the total pipeline to roughly 3 million GPUs.

Amazon’s own AI chip business, spanning the Graviton, Trainium, and Nitro product lines, has crossed an annual revenue run rate exceeding $20 billion. That figure could balloon to $50 billion if Amazon ever chose to commercialize its chips externally.

Trainium2, Amazon’s current-generation AI training chip, offers roughly 30% better price-performance than comparable GPUs and is nearly sold out. Its successor, Trainium3, is launching in early 2026 with promises of 30-40% better performance over its predecessor. It’s also nearly fully subscribed before it even ships.

The dual strategy and why it matters

At scale, Amazon expects its Trainium line to generate “tens of billions of capex dollars per year” in savings compared to relying exclusively on third-party chips.

Jassy explicitly acknowledged that many customers will continue choosing Nvidia within AWS. Some workloads, particularly those involving cutting-edge model training and inference at the frontier, benefit from Nvidia’s mature CUDA software ecosystem and raw performance.

What this means for the competitive landscape

Amazon’s Trainium chips don’t need to beat Nvidia at the high end. They just need to be good enough for the majority of workloads at a meaningfully lower cost. If Trainium3 delivers on its 30-40% performance improvement claims, the economic case for migrating routine inference and fine-tuning workloads to Amazon’s custom silicon becomes compelling.

Google has its TPUs. Microsoft is developing Maia. Meta has its own MTIA chips. Every major hyperscaler is running a version of the same playbook: maintain the Nvidia relationship for competitive and customer-facing reasons while quietly building alternatives that improve internal economics.

Investors watching this space should pay close attention to the $20 billion run rate on Amazon’s custom chip business. If that number accelerates toward the $50 billion external commercialization scenario, it would represent a fundamental shift in how AI compute is provisioned and priced.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.