Who Makes AWS's Chips? Amazon's AI Silicon Supply Chain

AWS increasingly runs on Amazon's own Trainium and Graviton chips instead of buying everything from NVIDIA. Here's who actually fabricates that silicon, and who's buying capacity on it.

Make Sure I Know It

Aug 9, 2026 · 2 min read

Share

AWS is the largest cloud platform in the world, and a growing share of the compute behind it isn't bought from NVIDIA at all. Amazon designs its own Trainium AI chips and Graviton server processors, and that self-designed silicon now generates north of $15 billion a year in annualized revenue. Designing a chip and building one are different things, though, and Amazon doesn't own a single fab.

The short answer

Amazon's AI chip supply chain has three parts:

• Who buys the capacity: Anthropic is AWS's top customer for Trainium chips, running large-scale AI model training on them.
• Who builds the silicon: TSMC fabricates Amazon's custom Trainium and Graviton chips, including the advanced CoWoS packaging AI chips need.
• Who supplies the networking: Broadcom co-designs custom ASIC silicon and supplies the high-speed switching that connects it all inside AWS data centers.

The full interactive map is on Amazon's supply chain page.

Why Amazon builds its own chips

Buying every AI chip from NVIDIA means competing with every other hyperscaler for the same limited supply, at NVIDIA's margins. Amazon's answer has been to design Trainium for AI training and Graviton for general-purpose servers in-house, then have someone else build them. That self-designed silicon has become a real business line, not a side project, with Anthropic standing out as the top customer running large-scale model training on Trainium capacity inside AWS.

Who actually fabricates the chips

Designing a chip is only half the job. is the foundry that turns Amazon's Trainium and Graviton designs into physical silicon, and the packaging side matters as much as the fabrication: AWS's CoWoS advanced-packaging wafer volume at TSMC is projected to climb sharply through 2026 as Trainium demand grows. That's the same dependency nearly every AI chip designer shares, whether it's Amazon, NVIDIA, or : the world's leading-edge fabrication capacity runs through one company in Taiwan.