AWS Triples Down on Nvidia: 2 Million More GPUs, Trainium Gets NVLink Fusion, and Federal AI Factories at IL6

On the same day Nvidia reported $96.2 billion in quarterly revenue, Amazon and Nvidia quietly rewrote the terms of their relationship. The two companies announced an expanded strategic collaboration that will put another 2 million Nvidia GPUs into AWS data centers during 2027 and 2028 — effectively tripling a procurement deal that was signed just five months ago, when AWS committed to deploying more than 1 million GPUs starting in 2026.

The announcement, made during Nvidia’s fiscal Q2 2027 earnings call on August 26, is the clearest signal yet that demand for AI compute is still running ahead of every forecast the industry sets for it. Neither company disclosed financial terms, but given current GPU unit costs, the deal is comfortably worth tens of billions of dollars.

What’s actually in the deal

The headline number — 2 million additional GPUs — covers Blackwell Ultra, Rubin, and Rubin Ultra architectures deploying across AWS’s global infrastructure, including the AI factories AWS is building for frontier-scale training and inference. Rubin, Nvidia’s next-generation platform, began production shipments this quarter, and investors have been watching closely for evidence that demand carries across the generational transition. A commitment of this size from the world’s largest cloud provider is exactly that evidence.

But the more interesting parts of the announcement are the ones that go beyond buying chips:

• Vera CPUs come to AWS. Amazon will deploy Nvidia’s Vera CPU-based infrastructure, including some units integrated with Rubin and others standalone, giving agentic AI workloads a high-performance general-purpose compute tier alongside GPUs. Nvidia CFO Colette Kress said Vera shipments are already underway to lead partners including Oracle and SpaceXAI, and that she expects deployment by “every major hyperscaler, neocloud, AI lab, and system OEM.” Jensen Huang has previously sized the Vera opportunity as a “$200 billion TAM.”
• Trainium gets Nvidia’s interconnect and memory technology. This is the most strategically loaded item. At re:Invent 2025, AWS announced support for NVLink Fusion in next-generation Trainium chips. Now Amazon’s Annapurna Labs is extending that work to NVHBM, Nvidia’s custom high-bandwidth memory technology, in partnership with memory suppliers. The upshot: Trainium and Nvidia GPUs will integrate within a common rack-scale architecture, with Trainium gaining access to faster, more power-efficient memory. Amazon’s custom silicon and Nvidia’s platform are becoming collaborators rather than pure competitors.
• Federal AI factories at Impact Level 6. AWS and Nvidia plan to build AI factories for the U.S. government, with plans for 100,000 GPUs on AWS’s secure infrastructure for federal and national-security workloads classified at IL6 and above. This puts the pair at the center of the U.S. government’s AI buildout.
• Physical AI for Amazon’s robot fleet. Amazon plans to adopt Nvidia’s full physical AI stack — Omniverse for simulation and digital twins, Cosmos world models, Isaac for robot development, and Jetson edge compute — to power its warehouse robotics fleet.
• G7 instances with RTX PRO 4500. AWS becomes the first major cloud to offer instances powered by Nvidia’s RTX PRO 4500 Blackwell Server Edition GPUs, delivering 4.6x the AI inference performance of the previous G6 generation.
• Open models on Bedrock. Nvidia’s Nemotron family of open models will be served through Amazon Bedrock and SageMaker.

The context: Amazon’s silicon hedge keeps growing

What makes this deal fascinating is that it lands while Amazon is aggressively building its own Nvidia alternative. On its most recent earnings call, Amazon said its custom silicon business has crossed a $25 billion annualized revenue run rate, driven by $225 billion in total commitments from AI labs including Anthropic and OpenAI. AWS AI chief Peter DeSantis has confirmed the company is in talks to sell Trainium chips to other companies for their own data centers, and its Arm-based Graviton CPUs are already a credible challenger to Intel and AMD in general-purpose server compute.

So why triple an Nvidia order at the same time? Because demand is not choosing sides. Frontier labs, enterprises, startups, and governments all want Nvidia’s platform — CUDA, networking, and the full software stack — even as they also want the cost economics of alternatives like Trainium. “Customers want the freedom to choose the best tools for their AI workloads, and they want confidence that everything works seamlessly together,” said AWS CEO Matt Garman. The NVLink Fusion and NVHBM work makes that literal: within a single rack, Amazon’s chips and Nvidia’s will share an interconnect and memory architecture.

The earnings backdrop

The partnership news did not arrive in a vacuum. Nvidia’s quarter itself was a blowout: $96.2 billion in revenue (up 106% year over year), $89 billion of it from data centers, and guidance of $108 billion for the next quarter. The company has now committed $279 billion to secure supply and manufacturing capacity for current and future data-center projects — up from $119 billion just last quarter — including $92 billion in projected spending for the rest of the fiscal year and $87 billion more in fiscal 2028.

“AI is now doing productive and useful work,” Huang said on the call. “AI is generating profitable tokens… If we had more compute, we could generate more profitable tokens, which results in more profit for all of the services. This is the exact phase where we’re at, which is the reason why everybody’s leaning in.”

The AWS deal is the concrete expression of that thesis: when the largest cloud provider triples a million-GPU order after five months because “demand has exceeded those expectations,” the infrastructure buildout is no longer speculative — it is backlog.

The open question investors will keep pressing is whether compute reliably converts into profit across the industry, not just for Nvidia. For now, the two companies that have benefited most from the AI buildout just locked in another two years of it together.