AWS and Nvidia Expand Infrastructure Partnership with Commitment to Deploy 2 Million Additional GPUs by 2028
Amazon Web Services and Nvidia have announced a major expansion of their cloud computing partnership, outlining plans to deploy two million additional Nvidia graphics processing units across global data center infrastructure between 2027 and 2028. The multi-year initiative expands upon an earlier agreement to introduce more than one million chips starting in 2026, driven by enterprise and developer demand for artificial intelligence compute that has continuously outpaced industry projections.
The scale of the expansion illustrates how cloud infrastructure providers are planning capital investments years in advance to support the escalating processing requirements of frontier AI models. By establishing a multi-phase rollout spanning from 2026 through 2028, AWS and Nvidia are structuring their hardware pipeline to accommodate sustained growth across enterprise, public sector, and research workloads.
Full-Stack System Architecture: Vera CPUs and Interconnect Upgrades
The updated infrastructure plan extends beyond raw graphics processor capacity to address system-level bottlenecks that frequently arise in massive computing clusters. To optimize throughput, AWS will integrate Nvidia’s newly developed Vera-based central processing units into its cloud environment. In demanding artificial intelligence training and inference workloads, legacy server CPUs can become data transfer bottlenecks, limiting overall processor efficiency. Incorporating Vera-based CPU architecture aims to balance system architecture and maintain steady data delivery to adjacent hardware accelerators.
Alongside processor updates, the cloud provider is upgrading its cluster interconnects through extended NVLink Fusion technology combined with custom high-bandwidth memory configurations. As artificial intelligence models expand across thousands of interconnected nodes, inter-chip communication latency becomes a primary constraint on computational speed. The combination of NVLink interconnects and memory bandwidth enhancement is engineered to allow large-scale GPU deployments to function as unified, low-latency computing fabrics.
Quantifying Performance Gains in Analytics, Search, and Inference
Beyond model training infrastructure, the hardware and software optimizations target data analytics and search services that underpin enterprise applications. Updates to Amazon’s analytics environment utilizing Nvidia’s cuDF software library are projected to deliver processing speeds nearly 3.7 times faster than standard computing configurations. From an operational cost perspective, this software-hardware integration is expected to yield approximately a 30 percent improvement in price performance compared to non-accelerated systems.
Search infrastructure performance is also being upgraded. Index construction for vector search capabilities within Amazon’s search platforms can now complete roughly nine times faster when executed on GPU infrastructure compared to standard setups. Vector indexing serves as a critical operational component for modern generative systems, semantic search, and retrieval-augmented generation architectures. Furthermore, recent hardware iterations deployed across AWS infrastructure have already demonstrated measurable gains, delivering up to 4.6 times faster inference processing and 2.1 times stronger graphics output compared to previous-generation hardware instances.
Public Sector Allocations and Warehouse Robotics Simulation
A distinct portion of the planned hardware deployment is targeted at government and defense requirements. Out of the overall chip commitment, AWS has designated 100,000 GPUs specifically for sensitive government and defense computing workloads nationwide. Dedicated public-sector infrastructure allows government bodies to execute complex data analysis, intelligence processing, and secure modeling within isolated cloud environments that adhere to strict regulatory and national security requirements.
In parallel with enterprise cloud offerings and public sector allocations, Amazon’s robotics division is collaborating with Nvidia to build advanced simulation environments. These digital simulation environments utilize high-performance computing to accelerate the training and testing of next-generation warehouse robots. By training robotic navigation and operational algorithms inside physics-accurate virtual simulations prior to real-world deployment, Amazon aims to shorten development cycles for automated systems deployed across its logistics and fulfillment networks.
Executive Perspectives and Long-Term Market Trajectory
Addressing the expanded agreement, Matt Garman, Chief Executive Officer of AWS, highlighted the necessity of providing flexible, integrated infrastructure for diverse enterprise workloads. Garman noted that technical co-engineering with Nvidia aims to optimize performance across every layer of the cloud stack, from networking and physical deployment to hardware security, providing enterprise customers, frontier research laboratories, and government bodies with specialized tools to deploy advanced models.
Jensen Huang, founder and CEO of Nvidia, pointed out that compute demand across global cloud infrastructure has repeatedly exceeded early forecasts. Huang noted that expanding the partnership across hardware, CPUs, networking, open software models, and custom software stacks is designed to accelerate the deployment of both software-based agentic systems and physical robotics applications at global scale.
The multi-year hardware commitment reflects sustained optimism regarding the long-term economic trajectory of artificial intelligence. As enterprise software transitions toward autonomous agentic workflows and automated physical systems, demand for high-density compute capacity continues to reshape cloud infrastructure planning. The ultimate impact of these infrastructure investments will become clearer as the hardware enters active service and commercial utilization tests the limits of the newly deployed capacity through 2028.
