AWS Commits to Over 3 Million Nvidia GPUs by 2028 to Meet AI Cloud Demand

Amazon Web Services is dramatically expanding its artificial intelligence compute capacity through multi-year agreements with Nvidia that will bring more than 3 million advanced processors into its fleet by mid-2028, according to a report from TechRadar Pro (https://www.techradar.com/pro/1-58-million-staff-amazon-will-soon-have-more-nvidia-gpus-than-employees-aws-to-buy-more-than-3-million-additional-chips-before-2029-in-addition-to-thousands-of-existing-h100-h200). The incoming hardware inventory means the cloud provider will operate nearly twice as many high-end Nvidia accelerators as parent company Amazon has employees across its global workforce of 1.58 million.

The latest agreement, announced on August 26 alongside Nvidia's quarterly financial results, outlines plans for AWS to deploy 2 million additional Blackwell Ultra, Rubin, and Rubin Ultra GPUs across its data centers in 2027 and 2028. That commitment expands upon an earlier deal announced during Nvidia's GTC conference in March 2026 for more than 1 million processors shipping from 2026 through 2027, as confirmed to Reuters by Nvidia executive Ian Buck.

Combined, the procurement plans represent approximately 1.9 Nvidia processors for every full- and part-time Amazon worker across logistics, fulfillment, and corporate operations. While the infrastructure footprint is expanding rapidly, Amazon's broader workforce remains tempered by recent restructuring, including the elimination of approximately 30,000 corporate positions between October 2025 and January 2026. Top human resources executive Beth Galetti previously characterized those cuts as a measure to reduce organizational layers and bureaucracy.

In parallel with third-party silicon, Amazon is leaning into its proprietary accelerators. The company reported that its custom Tranium3 chips were nearly fully booked as of mid-2026. The new Nvidia purchases will significantly expand AWS's baseline GPU inventory beyond its existing base of H100 and H200 chips, which trailed Microsoft's reported 485,000 units despite an estimated 196,000 Hopper processors purchased by Amazon in 2024, according to research firm Omdia.

The absolute volume of silicon depends partly on architectural accounting. Nvidia initially presented its Rubin-generation systems by individual compute die, unveiling the Vera Rubin rack as an NVL144 system at GTC 2025 before revising its metric ahead of CES 2026 to the Vera Rubin NVL72, counting 72 multi-die packages. The upcoming Rubin Ultra architecture due in 2027 houses four compute dies within each package. The latest supply deal also allocates roughly 100,000 GPUs specifically for U.S. government workloads meeting Impact Level 6 security standards.

The multi-billion-dollar infrastructure ramp comes as AWS struggles with supply constraints rather than demand. The cloud division reported second-quarter revenue of $42.2 billion, up 37% year over year, while contracted backlog climbed to $496 billion from $364 billion in the prior quarter. On Amazon's July earnings call, Chief Executive Andy Jassy stated that compute capacity would trail customer demand through next year, highlighting striking forward demand extending into 2028. The massive capital outlay has also pushed Amazon's free cash flow into negative territory as it finances its hyperscale expansion.