AWS bets on NVLink Fusion and Trainium4 to reshape AI infrastructure at hyperscale

At AWS re:Invent 2025, Amazon Web Services and NVIDIA Corporation unveiled a significant expansion of their longstanding partnership, centered around deep integration between NVIDIA NVLink Fusion interconnect technology and AWS’s custom silicon platforms. The announcement marks a strategic convergence of AWS’s infrastructure vision with NVIDIA’s scale-up compute architecture, as both firms aim to meet the surging global demand for high-performance, secure, and sovereign AI compute environments.

AWS confirmed that NVLink Fusion will be embedded across its next-generation Trainium4 accelerator, Graviton CPUs, and Nitro System. This multi-layer integration is designed to create a flexible and high-throughput computing backbone capable of powering generative AI, agentic models, robotics workloads, and hyperscale inference tasks. Industry observers see this as a transformative move, positioning AWS to compete aggressively not only on GPU-as-a-service offerings but also on vertically integrated AI infrastructure that blends proprietary silicon with industry-standard interconnects.

How is NVLink Fusion changing AWS’s approach to AI compute architecture?

With NVLink Fusion, AWS gains access to NVIDIA Corporation’s high-bandwidth, chip-to-chip communication protocol that enables CPUs, GPUs, and custom accelerators to operate as a unified, rack-scale system. This scale-up architecture allows Trainium4 to perform large-model training and inference without traditional bottlenecks such as PCIe latency or limited memory sharing across devices.

AWS also revealed that the integration is not confined to Trainium4. Its Graviton CPUs and the AWS Nitro System, both of which underpin general-purpose and virtualized workloads respectively, will adopt NVLink Fusion capabilities over time. This move broadens the impact of the partnership, suggesting that even non-GPU-centric applications will benefit from NVIDIA’s hardware and system engineering approach.

The adoption of NVIDIA MGX rack architecture alongside NVLink Fusion will allow AWS to simplify deployment logistics, reduce cooling and power overheads, and unify its internal silicon roadmap across multiple generations. By leveraging NVIDIA’s full supplier ecosystem — from chassis and power systems to interconnects — AWS is effectively building its next wave of data center infrastructure on top of a pre-integrated, modular compute fabric.

What role will Trainium4 and Blackwell GPUs play in AWS’s AI Factories?

As AWS scales its hardware offerings to meet global demand, it also introduced a new AI infrastructure product line branded as “AI Factories.” These are dedicated AI compute environments deployed inside customer-controlled data centers but operated and maintained by AWS. The goal is to offer high-performance compute with data sovereignty and regulatory compliance in mind, especially for governments, public-sector institutions, and large enterprises managing proprietary datasets.

At the heart of these AI Factories will be NVIDIA Corporation’s latest Blackwell GPUs, including the HGX B300 and GB300 NVL72 configurations, which support high-throughput AI training and inference. The soon-to-launch RTX PRO 6000 Blackwell Server Edition will also be available for visually intensive applications. All of these GPUs will be deployed alongside Trainium4 and Graviton chips, creating a heterogenous stack capable of handling both agentic AI and physical AI workloads.

The unified deployment strategy also includes NVIDIA Spectrum-X Ethernet switches, ensuring low-latency networking across all layers. AWS emphasized that this architecture provides customers with full stack visibility, traceability, and performance guarantees necessary for production-grade AI deployments. Customers can scale from prototyping to enterprise rollouts without reconfiguring infrastructure or software environments.

How does this support sovereign AI requirements for regulated sectors?

AWS and NVIDIA Corporation are jointly emphasizing the sovereign AI dimension of the new platform. As regulatory scrutiny increases in sectors like healthcare, defense, finance, and public administration, AWS AI Factories are being positioned as a secure alternative to public cloud deployments. The ability to operate these systems in physically isolated locations, with full customer control over data residency, access policies, and compliance frameworks, directly addresses some of the most persistent barriers to AI adoption in sensitive sectors.

Public sector customers will be able to integrate AWS’s Nitro-enforced isolation, encryption, and hardware root-of-trust features with NVIDIA’s full-stack AI software, including NeMo for LLM development and Riva for speech recognition. This blend of hardware and software provides a compliance-ready solution while maintaining high performance. Analysts monitoring government cloud procurement believe this architecture could disrupt traditional supercomputing contracts and offer a fast track for sovereign LLM training and inference capabilities.

What does the software stack integration between Amazon and NVIDIA enable for developers?

Beyond the hardware announcements, the expanded partnership includes several software integrations designed to improve developer experience and reduce time-to-production for AI services. One of the major updates is the integration of NVIDIA Nemotron models into Amazon Bedrock, a fully managed serverless platform for building generative AI applications. These include Nemotron Nano 2 and Nemotron Nano 2 VL, which support text, code, vision, and video-based workflows with minimal latency and strong inference accuracy.

These models are accessible on a pay-as-you-go basis via Bedrock, removing infrastructure management burdens and enabling instant scaling. CrowdStrike and BridgeWise are among the first customers using the service to deploy custom agentic AI models at scale.

Additionally, Amazon OpenSearch Service now includes serverless GPU acceleration powered by NVIDIA cuVS, an open-source library for vector-based search and clustering. This allows up to 10 times faster indexing of vector embeddings at a fraction of the cost. The result is dramatically reduced latency for retrieval-augmented generation workflows and improved performance for real-time AI search experiences. AWS is currently the only hyperscaler to offer such serverless GPU-accelerated vector indexing with NVIDIA GPUs.

How are robotics and physical AI platforms benefiting from the AWS–NVIDIA collaboration?

AWS is also integrating NVIDIA Cosmos world foundation models (WFMs) to power physical AI development, including robotics simulation, control, and training. These models are now available as NVIDIA NIM microservices on Amazon EKS, enabling real-time inference for robotic workloads in simulation and production.

Batch-based workloads such as large-scale synthetic data generation for physical environments can be executed on AWS Batch, which supports Cosmos WFMs in containerized form. The open-source simulation and reinforcement learning frameworks, Isaac Sim and Isaac Lab, are also supported in this stack. This creates a full cycle for robotic AI development, from simulation to deployment.

Numerous robotics firms including Agility Robotics, ANYbotics, Diligent Robotics, Field AI, and Skild AI are already leveraging this integrated platform to collect, process, and simulate robot data. The flexibility to operate in both centralized and edge environments makes this solution attractive to autonomous systems developers targeting logistics, healthcare, manufacturing, and defense sectors.

What are the implications for NVIDIA’s platform strategy and AWS’s silicon roadmap?

The expanded partnership signals a shift in positioning for both companies. NVIDIA Corporation is no longer just a GPU supplier but a full-stack platform enabler whose interconnect, software frameworks, and developer tools now power custom silicon from AWS. This allows NVIDIA to maintain influence across hyperscale deployments even when customers use internally developed accelerators like Trainium4.

Conversely, AWS’s silicon roadmap becomes more extensible and competitive by aligning with a vendor-neutral, high-performance interconnect. Analysts expect that future iterations of Trainium and Graviton could benefit from closer co-design with NVIDIA’s networking and system software, reducing development friction and time-to-market for new AI infrastructure products.

AWS received the Global Generative AI Infrastructure and Data Partner of the Year award from NVIDIA, recognizing its use of NVIDIA technologies to support data storage, synthetic generation, and vector embedding workloads. This recognition highlights the maturity of the AWS–NVIDIA relationship and reinforces investor sentiment that both companies will remain central players in the AI infrastructure race.

What are the key takeaways from the AWS–NVIDIA AI infrastructure expansion?

• AWS Trainium4 will integrate NVIDIA NVLink Fusion to boost AI training and inference performance.
• NVLink Fusion will also extend to AWS Graviton CPUs and the Nitro System, aligning AWS’s infrastructure stack with NVIDIA’s MGX rack designs.
• NVIDIA Blackwell GPUs and Spectrum-X switches will power sovereign AWS AI Factories with on-premise deployment options.
• AWS AI Factories target regulated sectors with compliance-driven infrastructure that supports full-stack AI capabilities.
• Amazon Bedrock gains access to NVIDIA Nemotron models, while OpenSearch adds cuVS GPU acceleration for faster vector search.
• NVIDIA Cosmos WFMs and Isaac frameworks are now available on AWS for robotics simulation and physical AI training.
• The expanded collaboration aligns AWS’s custom silicon roadmap with NVIDIA’s full-stack platform vision and interconnect architecture.