NVIDIA and AWS Team Up to Run Enterprise AI at Production Scale
Running AI at production scale just got a clearer path on AWS. NVIDIA and Amazon Web Services have deepened their collaboration across multiple layers of the cloud infrastructure stack, targeting the real friction points enterprises face when moving AI from testing into production at NVIDIA AWS AI production scale.
The most visible piece is the new Amazon EC2 G7 instance family, now powered by NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs. Compared to the previous G6 generation, G7 delivers up to 4.6x faster AI inference performance and up to 2.1x better graphics performance, while also accelerating GPU-powered data analytics through the NVIDIA cuDF library on Amazon EMR. The instances support up to eight GPUs with 256GB of total GPU memory, 700 Gbps of EFA-enabled networking, and up to 7.6TB of local NVMe SSD storage, giving teams real flexibility to right-size compute rather than overprovision it. G7 instances are already available through AWS Deep Learning AMIs, Amazon EKS, Amazon ECS, and Amazon Deep Learning Containers, with Amazon SageMaker AI support coming soon.
On the search and retrieval side, the next generation of Amazon OpenSearch Serverless is now defaulting to GPU-accelerated vector indexing, powered by the NVIDIA cuVS library. For teams building retrieval-augmented generation systems, semantic search, and agentic AI applications, this is a meaningful shift. What was once a specialized optimization project becomes a standard capability out of the box, with vector indexing up to 10x faster at a quarter of the cost compared to CPU-only setups. That makes billion-scale vector databases something teams can realistically build in under an hour.
AWS has also earned NVIDIA Exemplar Cloud status for the NVIDIA GB300, meaning it meets the strict performance benchmarks NVIDIA uses to validate training workloads against its reference architecture. This comes out of deep co-engineering work between the two companies and gives AI teams a stronger signal that they are getting consistent, peak-optimized infrastructure when training large models on AWS. The designation is designed to help organizations evaluate cloud providers with more confidence and reduce the gap between planning and actual deployment.
Taken together, the announcements cover inference, retrieval, and training, the three layers where production AI deployments most commonly hit bottlenecks. The goal across all three is the same: infrastructure that performs at scale without piling on operational complexity for the teams managing it.
