Amazon and Nvidia portray their enlarged alliance as the fastest route to put AI into production at scale, while AWS’s own chip programme underscores the strategic tension beneath the deal: Amazon wants more Nvidia capacity now without surrendering its longer-term independence.

The partnership had already been moving quickly. In March, AWS said it would deploy more than one million Nvidia GPUs beginning in 2026. Five months later, demand from startups, enterprises, AI labs and governments had surpassed that plan, prompting Amazon to order an additional two million Blackwell Ultra, Rubin and Rubin Ultra GPUs for AWS data centres in 2027 and 2028.¹

Nvidia chief executive Jensen Huang cast the escalation as evidence that the AI infrastructure race is no longer speculative. “Demand is running ahead of every forecast,” he said, arguing that the companies’ 16-year cloud relationship is now expanding across “GPUs, CPUs, networking, open models and software.”²

AWS chief executive Matt Garman framed the same move as a customer-choice proposition rather than a lock-in strategy. Customers, he said, want the best tools for their AI workloads and confidence that they work together; the expanded collaboration gives labs, businesses and governments more ways to build on AWS.²

The deal reaches well beyond chip purchases. AWS plans to bring Nvidia’s Vera CPUs into its cloud, serve Nvidia’s Nemotron models through Bedrock and SageMaker, and adopt Nvidia’s robotics stack—including Jetson, Omniverse and Isaac—for Amazon Robotics.¹

Yet Amazon is also building the counterweight. Its Trainium chips compete directly for deep-learning workloads, while Graviton targets server processors; AWS has said its custom-chip business passed a $25 billion annualised revenue run rate. The immediate verdict is clear: even a company investing heavily in alternatives sees Nvidia as indispensable while demand for AI compute keeps accelerating.¹