Amazon turns 2M more NVIDIA GPUs into a supply-chain signal—while AWS Trainium argues the “Nvidia-only” story is incomplete

Amazon and NVIDIA expanded their partnership in a way that is directly measurable in unit terms. The companies said the deal adds another 2 million NVIDIA GPU chips for AWS data centers, spanning 2027 and 2028, and covering Blackwell Ultra, Rubin, and Rubin Ultra.

Two details matter for investors who track supply chain capacity rather than just headline “AI demand.” First, Amazon is not merely extending legacy Blackwell—it is explicitly naming Rubin generations, which increases confidence that the order book is forward-looking rather than a catch-up purchase of already-shipped platforms. Second, the inclusion of multiple generations in one incremental bundle raises the probability that suppliers will see sustained wafer/HBM/package loading rather than a short, single-product spike.

Nvidia’s AI growth story has often been discussed in terms of visibility and supply constraints. Reuters reported that Nvidia forecast 70% revenue growth next fiscal year (with the market implication being that the cycle still has runway), and that framing was circulating right around the same time as Amazon’s expanded 2 million GPU commitment for 2027–2028.

The key analytical move is to treat this as a constraint-driven supply-chain indicator. If hyperscalers are willing to lock incremental volumes into 2027–2028, then suppliers downstream of Nvidia are more likely to receive orders that must be scheduled around capacity limits (packaging, HBM sourcing, and rack integration). That tends to improve leverage for the constrained links, at least on timing, even if the eventual margin outcome depends on pricing and mix.

A unit increase of this magnitude is rarely “GPU-only” economically. For each incremental Nvidia accelerator deployed at AWS scale, the supply chain must also deliver enough high-bandwidth memory capacity and packaging throughput to meet rack-level system targets. That means a hyperscaler upsizing its Nvidia GPU allocation tends to increase demand pressure for the links that schedule on HBM availability and advanced packaging capacity (the parts of the stack that often behave like the real bottleneck, not the bare chip fab).

Important caveat: this chain-of-custody behavior is an inference from how AI infrastructure is built, not a direct disclosure of CoWoS or HBM unit quantities in the sources we opened for this story. The article below therefore treats the bottleneck channels as the most likely transmission mechanisms, while the only hard, quoted numbers remain the GPU-unit expansion and Nvidia’s disclosed growth framing.

Amazon is simultaneously scaling its own custom AI accelerator portfolio. AWS describes Trainium as a purpose-built AI chip and positions it as delivering “the best economics” for training and inference at scale. The operational implication is simple: when a hyperscaler invests in custom silicon, it can moderate the rate at which incremental tokens and workloads translate into incremental Nvidia-GPU purchases—even while total AI compute spend continues to rise.

This is why the order book should be read as “share-of-wallet confirmation” rather than a blanket replacement of custom silicon. Amazon’s disclosed deal increment is a strong Nvidia tailwind; AWS’s Trainium positioning is the reminder that Amazon will still optimize cost-per-token and availability by routing workloads across accelerators.

The reason this matters for the Nvidia order signal is timing. Amazon’s expanded Nvidia GPU commitment points to deployments in 2027–2028. That makes the question less “does Amazon have AI demand?” and more “can it keep funding data center expansion through the back half of the cycle?” Amazon’s FY2025 filings show strong operating cash generation, but free cash flow compressed as capital expenditures stayed heavy.