Amazon

Amazon's relevance to this value chain runs through one specific part of the company: AWS's custom silicon program, built around Annapurna Labs, the Israeli chip design team Amazon acquired in 2015. That acquisition seeded everything that followed — Graviton (Arm-based CPUs for general cloud computing), Nitro (the networking and security chips inside every AWS server), Inferentia (AI inference), and Trainium, the AI training accelerator line that most directly overlaps with this value chain's story.

The strategic logic is straightforward: Amazon runs the largest cloud infrastructure business in the world, and every chip it doesn't have to buy from Nvidia at Nvidia's margins is a chip it gets to build closer to its own cost. AWS has been explicit that its hardware "north star" is the best performance per total cost of ownership, not chasing the single fastest chip on the market — a philosophy that shows up in how deliberately AWS multi-sources components and avoids over-committing to any one architecture.

Trainium3, AWS's newest AI training chip, reached general availability around re:Invent 2025 — built on a 3nm process, packing 144GB of HBM3e memory per chip at 4.9 terabytes per second of bandwidth, with hardware built specifically for the mixture-of-experts routing patterns common in today's largest AI models. It's part of a rapidly maturing family that now includes four generations across Trainium and Inferentia combined.

Today, Trainium's packaging still runs on established, non-glass technology: Trainium2 and Trainium3 use TSMC's CoWoS-R platform, with an organic thin-film interposer — six layers of copper redistribution wiring on polymer — rather than a full silicon interposer, chosen specifically because it's cheaper and mechanically more forgiving at reticle-scale package sizes than a rigid silicon interposer would be. That choice is itself a preview of the exact problem glass substrates are trying to solve: as these packages get larger, organic materials increasingly can't keep up with warpage at scale, which is precisely why the AWS chip a few generations from now may need a different material altogether.

It's worth being precise about what is and isn't confirmed here, because AWS's position is meaningfully less concrete than AMD's. Amazon has not published its own glass substrate roadmap the way AMD has. What's publicly reported instead is narrower: AWS is named, alongside AMD, as one of the customers to whom Absolics (SKC's glass substrate subsidiary) has been supplying prototype samples for performance testing, with the relationship reportedly approaching a "pre-qualification" stage as of mid-2026.

That's real signal, but it's evaluation-stage signal, not commitment-stage signal. Korean trade press — including ETNews, Techbrew, and Newdaily — have named AWS this way consistently from mid-2025 through early 2026, which makes it a credible data point. But all of them describe AWS testing glass substrate samples, not AWS having decided to build a product on them.

Amazon sits in stage 5, chip designers & AI demand, alongside AMD — both are named customers in Absolics' reported qualification pipeline, making them the two most concretely sourced names on the demand side of this chain so far.

Amazon is also profiled in The AI Infrastructure Value Chain — stage 07, Cloud & hyperscalers.

The most consequential recent Trainium news isn't about glass at all — it's that Meta confirmed in 2026 it would deploy AWS's Trainium and Inferentia chips inside its own infrastructure, alongside its existing Nvidia GPUs. That matters here because it's the clearest evidence yet that AWS's custom silicon has crossed from an internal cost-saving project into chips other hyperscalers are willing to run in production. More external validation for Trainium means more volume, and more volume is exactly the kind of pressure that pulls a chip program toward adopting whatever packaging material lets it keep scaling — which is the same pressure pushing AMD and Intel toward glass.

AWS also runs some of the largest known AI training clusters in the world on Trainium hardware, including "Project Rainier," a cluster built around hundreds of thousands of Trainium2 chips supporting Anthropic's model training. Clusters at that scale are a useful proxy for how large a single Trainium package might eventually need to become — and package size is the exact variable that determines when a chip designer stops being able to rely on organic substrates and has to move to glass.