Amazon and Nvidia announced Wednesday an expanded partnership that adds another 2 million Nvidia GPU chips to Amazon Web Services data centers, with deliveries scheduled for 2027 and 2028. The order includes Blackwell Ultra, Rubin, and Rubin Ultra GPUs.
This comes just five months after Amazon agreed to deploy more than 1 million Nvidia GPUs across AWS infrastructure starting in 2026. Nvidia said demand has exceeded expectations since the earlier agreement. Neither company disclosed financial terms, but based on GPU unit costs the deal is worth tens of billions of dollars.
What's new
- 2 million additional GPUs for AWS data centers in 2027-2028, on top of 1 million+ GPUs already committed for 2026 deployment
- GPU models: Blackwell Ultra, Rubin, and Rubin Ultra
- Vera CPUs from Nvidia will also ship to AWS — some integrated with Rubin, others standalone — with shipments already underway to lead partners including Oracle and SpaceX AI
- Full physical AI stack adoption: Amazon plans to use Nvidia Omniverse, Cosmos, Isaac, and Jetson for its warehouse robot fleet
- Enterprise integration: Nvidia Nemotron open models will be available on Amazon Bedrock and SageMaker
Context: competing while partnering
The partnership deepens even as Amazon ramps up its own AI chip efforts to reduce dependence on Nvidia. Amazon's Trainium accelerators target the same deep learning workloads as Nvidia's H100 and Blackwell chips. Its Arm-based Graviton CPUs challenge traditional server processors from Intel and AMD.
AWS AI chief Peter DeSantis has said the company is in talks to sell Trainium chips to other data center operators — a potential $50 billion opportunity, according to CEO Andy Jassy's April 2026 shareholder letter. Amazon's custom silicon division reportedly crossed a $25 billion annualized revenue run rate in its most recent quarter, driven by $225 billion in total commitments from AI labs including Anthropic and OpenAI.
Yet the company simultaneously tripled its Nvidia order, signaling that internal silicon cannot yet meet surging customer demand.
Why it matters
The deal cements Nvidia's position as the primary merchant silicon supplier for the largest cloud provider even as that cloud provider builds competing chips. Nvidia reported $96.2 billion in Q2 sales with data center revenue of $89 billion, up 117% year over year, and guided for $108 billion in Q3. The company has committed $279 billion to secure supply and manufacturing capacity for current and future data center projects.
For AWS customers, the expanded partnership means continued access to Nvidia's latest GPU generations and software stack. For the broader market, it illustrates that hyperscaler custom silicon programs and merchant GPU purchases are growing in parallel rather than substituting for each other.
Our take
Amazon's simultaneous pursuit of custom silicon and a record Nvidia order reveals the practical limits of vertical integration at current AI demand levels. Trainium3 may approach Blackwell performance at the rack level, but Nvidia's CUDA ecosystem, manufacturing priority at TSMC, and now Vera CPU expansion give it compounding advantages that a $50 billion merchant chip ambition from AWS has not yet dented.
