Analysis
AWS and Nvidia said they will deploy 2 million additional Nvidia GPUs across Amazon's global data center infrastructure, TechCrunch reported, tripling a chip commitment that started at roughly 1 million units when AWS first detailed its Nvidia buildout. The companies attributed the expansion to surging demand from startups, enterprises, AI labs and even governments outpacing Amazon's original capacity projections.
- AWS (Amazon) -- expanding from ~1M to ~3M Nvidia GPUs committed across global infrastructure
- Nvidia -- supplying the expanded GPU volume as part of a deepened partnership spanning AI factories, CPUs, networking, open models and robotics
- Anthropic, OpenAI -- among the AI labs whose compute demand is cited as driving both this Nvidia order and Amazon's separate custom-chip commitments
The Hedge Underneath the Headline
The more interesting detail sits just below the Nvidia order itself: Amazon is simultaneously scaling its own custom AI chip business -- built around its in-house Trainium silicon -- which has crossed a $25 billion annualized revenue run rate, backed by $225 billion in total commitments from AI labs including Anthropic and OpenAI. Buying dramatically more Nvidia GPUs while also racing to scale a Nvidia-competing chip business isn't contradictory; it's Amazon hedging both directions at once, capturing near-term demand it can't otherwise serve with an unproven custom chip supply chain while continuing to build toward long-term independence from Nvidia pricing and allocation decisions.
Pulse has tracked Amazon's broader AI infrastructure buildout across prior coverage of both its custom silicon ambitions and its Nvidia dependency.
Why the Timing Lines Up
This expanded Nvidia order landed the same week Nvidia reported 106% year-over-year revenue growth and guided to continued acceleration -- Amazon's order is one concrete data point behind that guidance, evidence that at least one hyperscaler's demand for Nvidia chips is genuinely outpacing what it originally planned to buy, rather than demand manufactured through vendor financing.
The Counterweight
Tripling a chip order sounds unambiguously bullish, but it also means Amazon's own capital expenditure obligations just grew substantially, at a moment when its custom Trainium chip business is still relatively early in proving it can match Nvidia's performance and software ecosystem at comparable cost -- a bet that both paths (Nvidia dependency now, custom silicon independence later) pay off, rather than a signal Amazon has resolved which strategy wins long-term.
What to Watch
The ratio between Amazon's Nvidia GPU deployment and its Trainium custom-chip run rate over the next several quarters will show which side of this hedge Amazon is actually betting will dominate its infrastructure mix by the end of the decade. The deepened collaboration across AI factories, networking and open models beyond raw chip volume also suggests Amazon and Nvidia are trying to lock in a broader technical relationship, not just a purchase order, which raises the switching cost for Amazon if it later wants to lean harder into Trainium at Nvidia's expense. That's a meaningful strategic tension for AWS's own enterprise customers to watch, since it shapes how much pricing leverage Amazon retains over the GPU capacity those customers are renting month to month.
