Nvidia strikes massive GPU deal with AWS, deploying over a million chips through 2027

Nvidia and Amazon Web Services have locked in one of the largest GPU supply agreements in cloud computing history, with Nvidia set to deliver over one million GPUs to AWS starting in 2026 and running through 2027. The deal spans Nvidia’s latest Blackwell and Rubin chip architectures and signals a deepening dependency between the world’s largest cloud provider and its most important silicon supplier.

AWS first revealed the arrangement during Nvidia’s GTC event on March 16, 2026. Nvidia VP Ian Buck confirmed the timeline three days later, noting that shipments would begin in 2026 and continue through the end of 2027.

More than just GPUs

What makes this deal particularly notable is its scope beyond graphics processors. The agreement includes Nvidia’s full networking stack: Spectrum networking chips, ConnectX adapters, and Spectrum X gear. That’s a meaningful shift for AWS, which has historically leaned on custom-built networking solutions for its data centers.

The deal also incorporates Groq’s inference-optimized chips, a byproduct of Nvidia’s $17B licensing agreement with Groq established in late 2025. Including Groq hardware alongside Nvidia’s own GPUs gives AWS a more diversified toolkit for handling different types of AI workloads, from training massive foundation models to running the inference queries that power consumer-facing AI products.

The inference shift is real

The timing of this partnership reflects a fundamental change in how the AI industry consumes compute power. Inference now accounts for roughly two-thirds of the sector’s total compute demand. Just three years ago, that ratio was flipped, with training consuming the lion’s share.

Nvidia CEO Jensen Huang has framed the opportunity in characteristically ambitious terms, projecting a $1 trillion revenue opportunity for the Blackwell and Rubin product families through 2027. Even accounting for Huang’s flair for big numbers, the AWS deal represents a concrete down payment on that vision.

What this means for the competitive landscape

AWS committing to over a million Nvidia GPUs is both a vote of confidence in Nvidia and a competitive signal to the rest of the cloud market. Microsoft has its own deep Nvidia relationship, and Google has increasingly pushed its custom TPU chips as an alternative. But AWS choosing to go all-in on Nvidia’s compute and networking stack, rather than doubling down on in-house alternatives like its Trainium and Inferentia chips, tells you something about where the performance benchmarks currently stand.

That said, AWS hasn’t abandoned its custom silicon efforts. The company continues to develop proprietary chips for specific workloads where cost optimization matters more than peak performance. The Nvidia deal is better understood as AWS building out its premium AI infrastructure tier, the one enterprise customers with the biggest budgets and most demanding workloads will gravitate toward.