NVIDIA and AWS deepen technical cooperation to standardize global AI infrastructures

NVIDIA and AWS are expanding their technical collaboration through deeper integration of hardware, interconnect technologies and software layers to provide a unified and scalable platform for future AI needs. The collaboration includes AWS cloud infrastructure as well as custom chips Trainium4, Graviton and the Nitro system. At the heart of this is the implementation of NVIDIA NVLink Fusion, a technology for coupling customized AI systems that AWS integrates into its own silicon solutions. The combination of NVLink Scale-up Interconnect and the NVIDIA MGX rack design creates a platform designed for low power consumption, high bandwidth and consistent latency. The integration is part of a multi-year roadmap aimed at accelerating the development of new cloud scaling architectures.

AWS is using Trainium4 in combination with NVLink Fusion for the first time, increasing the efficiency of inference workloads and agentic AI models while reducing management overhead. Already active MGX racks with NVIDIA GPUs are extended by NVLink Fusion with a unified system structure. Additionally, AWS leverages NVLink Fusion’s supplier ecosystem, which provides components such as chassis, power and cooling solutions to simplify rack-level deployment. Support for Elastic Fabric Adapter and Nitro enables the upcoming NVIDIA Vera Rubin architecture to be fully embedded into the existing AWS network fabric.

In the area of international AI infrastructure, AWS is planning the global expansion of sovereign AI clouds, which will take stricter regulatory frameworks into account. These structures are based on NVIDIA Blackwell GPUs such as HGX B300 and GB300 NVL72 as well as the new RTX PRO 6000 Blackwell Server Edition for visual applications. The systems form the foundation of the AI Factories operated by AWS, which provide companies with dedicated AI data centers without violating data sovereignty requirements. Public sector organizations in particular are expected to benefit from the combination of AWS cloud services and NVIDIA computing architectures, as it allows for extensive model training with complete control over the most sensitive data.

The collaboration will also be expanded at software level. The open NVIDIA Nemotron models are integrated into Amazon Bedrock and enable the development of generative applications and multimodal agents without infrastructural management overhead. Immediate deployment via Bedrock’s serverless model lowers the barriers to entry for companies. The first users from the security and finance sectors are already using this integration productively. In parallel, Amazon OpenSearch Service receives GPU-based acceleration based on NVIDIA cuVS, allowing vector indexes to be built faster. Early test results indicate a significant reduction in latency with simultaneous cost savings, which is particularly relevant for systems with retrieval augmented generation.

For the operation of agent-based systems, NVIDIA and AWS provide a coordinated software chain consisting of Strands Agents, the NVIDIA NeMo Agent Toolkit and Bedrock AgentCore. This combination covers the steps from development to performance tuning and scalable execution. The cooperation also includes the simulation of physical AI systems. NVIDIA Cosmos world models are available via NIM microservices in Amazon EKS and enable real-time robotics applications. The same models can be used for extensive data synthesis via AWS Batch. The scenarios generated in this way are suitable for validation in Isaac Sim and Isaac Lab. Companies in the robotics sector use this infrastructure to collect, process and train robot-generated data.

The deepened partnership marks a further step in the long-standing technological relationship between the two companies. As part of this, NVIDIA was recognized as AWS Global GenAI Infrastructure and Data Partner of the Year. Both partners are increasingly focusing their architecture on global, scalable and data-sovereign AI infrastructures that are designed to provide complex models securely and efficiently.

Conclusion

The expansion of the collaboration between NVIDIA and AWS leads to a unification of hardware, interconnects and software services, increasing the scalability and performance of AI infrastructures. The integration of NVLink Fusion into AWS silicon represents a structural change in the design of cloud-based AI systems. Complementary software optimizations and the integration of sovereign AI clouds expand the offering towards fully controllable high-performance environments for enterprises and public institutions.

Sources