AWS, NVIDIA to deploy 2 million GPUs as AI demand accelerates

The 2 million GPU commitment more than doubles AWS's previously announced Nvidia deployment plans.

AWS and NVIDIA plan to deploy 2 million additional GPUs across Amazon's cloud in 2027-2028, cementing AWS's position as the largest buyer of Nvidia accelerators while intensifying the cloud AI arms race.

"Demand is running ahead of every forecast," Jensen Huang, founder and CEO of NVIDIA, said. "For 16 years, we have scaled NVIDIA computing in the cloud together. Now, we are expanding our partnership across the full stack — GPUs, CPUs, networking, open models and software."

The expansion covers Blackwell Ultra, Rubin and Rubin Ultra GPUs, plus NVIDIA Vera CPUs, NVLink Fusion interconnect technology, and RTX PRO 4500 Blackwell Server Edition GPUs for EC2 G7 instances that deliver 4.6x AI inference performance over prior-generation G6. AWS also plans to build AI factories for the U.S. government with 100,000 GPUs on secure infrastructure rated for Impact Level 6 workloads.

The commitment comes as Nvidia faces intensifying competition from AMD's Helios platform, Google's TPU 8i, and Amazon's own Trainium3 silicon. Bloomberg Intelligence forecasts Nvidia will retain at least 70 percent of AI training workloads, but the AWS deal locks in demand at a scale that rivals' custom chips have yet to match.

Two Million GPUs and Counting

The new commitment builds on the more than 1 million NVIDIA GPUs AWS announced at GTC 2026, scheduled to begin arriving this year. The additional 2 million Blackwell Ultra, Rubin and Rubin Ultra accelerators will be deployed across AWS's global infrastructure, including AI factories, through 2028. AWS CEO Matt Garman said the expansion gives "frontier labs, enterprises and governments even more ways to build and deploy AI on AWS."

The partnership extends beyond raw GPU capacity. AWS will bring NVIDIA Vera CPUs to its infrastructure, providing high-performance CPU compute for agentic AI workloads that require more than just accelerator processing. The companies are also expanding support for NVLink Fusion high-speed chip interconnect technology to work with NVIDIA's new custom high-bandwidth memory (NVHBM), which would give Amazon's Trainium chips access to faster, more power-efficient memory.

On the federal side, AWS and NVIDIA plan to deliver 100,000 GPUs on AWS's secure infrastructure for U.S. government and national-security workloads classified at Impact Level 6 and above. The federal AI factories represent a significant expansion of both companies' government business.

AMD, Google and Etched Pressure Nvidia's Pricing Power

The scale of the AWS-NVIDIA commitment arrives as Nvidia's competitive moat faces its most serious test since the AI boom began. AMD's Helios rack-scale platform, which layers 72 MI455X accelerators with EPYC CPUs, claims up to 30 percent more inference tokens per dollar compared to Nvidia's Vera Rubin NVL72 on selected workloads. OpenAI expects to bring Helios online in late 2026.

Google's TPU 8i reportedly offers 80 percent better performance per dollar than its predecessor, while Amazon's Trainium3 — which the AWS-NVIDIA collaboration on NVLink Fusion and NVHBM will now support — delivers up to 4.4x Trainium2's performance. Broadcom and Marvell are enabling hyperscalers to build purpose-designed accelerators, with OpenAI recently showing off a Broadcom-developed inference chip.

Startups are adding pressure from another angle. Etched, an AI-chip startup valued at $21 billion after a $700 million funding round involving Jane Street, claims its technology showed nearly 10 times the performance at lower cost per die compared to Nvidia. Michael Burry, the "Big Short" investor, called Etched "serious competition for NVDA" in a recent Substack post.

Nvidia's CUDA software stack, networking, and years of developer adoption remain formidable barriers. Bloomberg Intelligence forecasts Nvidia will retain at least 70 percent of AI training workloads despite the competitive pressure. But the AWS deal — which locks in demand for Nvidia's next three GPU generations — provides a critical revenue anchor as rivals chip away at the edges.

Nvidia shares closed at $217.56 on Aug. 19, up roughly 17 percent year to date but down from a May record of $236.54. The stock trades at about 24 times non-GAAP earnings. Nvidia reports fiscal Q2 results on Aug. 26, with Wall Street expecting $92 billion in sales and $2.08 in earnings per share, roughly double year-ago levels.

This article is for informational purposes only and does not constitute investment advice.