TL;DR — Key Takeaways
- AWS plans to deploy 2 million additional NVIDIA GPUs in 2027 and 2028, on top of more than 1 million GPUs previously planned beginning in 2026.
- The expansion will include Blackwell Ultra, Rubin and Rubin Ultra GPUs, reflecting continued strong demand for AI infrastructure.
- AWS is balancing its investment in NVIDIA hardware with development of its own Trainium accelerators, including tighter integration through NVLink Fusion and NVHBM.
Amazon Web Services plans to deploy 2 million additional NVIDIA GPUs in 2027 and 2028, a new deployment that adds to AWS’s previously announced plan to add more than 1 million NVIDIA GPUs beginning in 2026.
That earlier commitment was announced at NVIDIA GTC 2026, but customer demand moved beyond AWS’s and NVIDIA’s expectations. The new GPUs will include NVIDIA Blackwell Ultra, Rubin and Rubin Ultra processors.
In short, the massive expansion demonstrates continued robust demand for the infrastructure required to support AI workloads. The deal also reveals the difficult balancing act facing hyperscalers: They are developing their own AI chips to control costs and reduce dependence on NVIDIA while simultaneously buying enormous numbers of NVIDIA processors.
Deeply Integrated Infrastructure
The expanded partnership includes a wide array of infrastructure. AWS and NVIDIA are integrating their technologies more closely across chips, networking, memory, AI models and robotics.
AWS has invested heavily in its Trainium line of custom AI accelerators. Rather than treating Trainium and NVIDIA GPUs strictly as competing options, the companies are working to make the two technologies operate more closely together.
AWS previously announced that next-generation Trainium chips would support NVIDIA’s NVLink Fusion, which allows third-party processors to connect to NVIDIA’s high-speed interconnect architecture. Amazon’s Annapurna Labs is now also working with NVIDIA on NVHBM, a custom high-bandwidth memory technology.
The combination will allow Trainium and NVIDIA GPUs to operate within a common rack-scale architecture. This gives AWS greater flexibility to use its own silicon while continuing to offer NVIDIA hardware for workloads that require it.
AWS and NVIDIA are also extending their partnership into data processing and enterprise AI software. NVIDIA’s Nemotron open models are available through Amazon Bedrock and SageMaker. The companies are working on GPU-accelerated vector indexing for Amazon OpenSearch Service and data processing for Amazon EMR.
Additionally, Amazon Robotics plans to use NVIDIA technologies including Jetson, Omniverse and Isaac for tasks like simulation, robot training, synthetic data generation and route optimization.
Part of the expanded partnership supports government AI. AWS and NVIDIA plan to deploy 100,000 GPUs for federal and national security workloads, including systems to handle Impact Level 6 and higher classifications.
This type of aggressive investment was reflected in NVIDIA’s recent earnings report. The chipmaker reported another sharp jump in sales, posting quarterly revenue of $96.2 billion, up 106% from a year earlier, with data center revenue reaching $89 billion.