Amazon Web Services and NVIDIA announced a major expansion of their partnership, committing to deploy an additional 2 million GPUs across AWS global infrastructure throughout 2027 and 2028. This multi-year agreement builds on previous commitments and reflects surging demand from enterprises, governments, and AI research labs for accelerated computing power.

Why Are AWS and NVIDIA Doubling Down on GPU Deployment?

The accelerated rollout reflects three converging forces: enterprise demand for AI training and inference workloads, sovereign computing requirements from governments, and frontier research labs pushing the boundaries of artificial intelligence. AWS will integrate NVIDIA's next-generation Blackwell Ultra, Rubin, and Rubin Ultra architectures into its AI factory clusters, creating a unified computing ecosystem designed to handle everything from model training to real-time inference.

Beyond raw GPU counts, the partnership involves deep technical collaboration across hardware and software. AWS is bringing NVIDIA's Vera CPU-based infrastructure to its cloud platform, delivering dedicated high-performance processors specifically optimized for agentic AI orchestration. Additionally, Amazon's Annapurna Labs is extending integration to support NVIDIA's custom high-bandwidth memory subsystems, allowing AWS's Trainium chips to leverage NVIDIA's memory and interconnect technology within unified rack-scale configurations.

What Specific Infrastructure Improvements Are Coming?

The partnership includes several concrete hardware and software advances designed to improve performance and reduce costs for AWS customers:

• Blackwell Server Instances: AWS is introducing Amazon EC2 G7 instances powered by NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs, delivering 4.6 times the AI inference performance and 2.1 times the graphics performance compared to previous-generation G6 systems.
• Accelerated Analytics: AWS is implementing NVIDIA cuDF libraries on Amazon EMR to achieve up to 3.7 times faster processing speeds and a 30 percent price-performance gain over CPU setups.
• Vector Search Optimization: GPU-accelerated vector indexing on Amazon OpenSearch Service via cuVS delivers up to 9 times faster vector indexing at a quarter of previous costs, critical for retrieval-augmented generation (RAG) applications that power modern AI search.
• Government-Grade Security: AWS and NVIDIA are constructing dedicated AI factories for the United States government, deploying 100,000 NVIDIA GPUs across secure infrastructure engineered to process classified national security workloads at Impact Level 6 and above.

Amazon Robotics is also standardizing on NVIDIA's full-stack physical AI platform, utilizing NVIDIA Jetson, Omniverse libraries, and the Isaac robotics platform for warehouse automation, synthetic training data generation, and fleet simulation. This represents a significant commitment to deploying AI not just in data centers but in physical automation systems.

How Does This Partnership Address Competition from Alternative AI Chips?

The AWS-NVIDIA expansion arrives as competitors attempt to reduce dependence on NVIDIA's dominant GPU market position. OpenAI recently unveiled Jalapeno, an inference chip developed in partnership with Broadcom that the company claims outperforms NVIDIA processors in certain tests. However, Jalapeno is specifically designed for inference, the phase when AI models respond to user prompts, not for training new models where NVIDIA's technology remains unmatched.

Google developed its own Tensor Processing Units (TPUs) years ago, and Intel attempted to compete with Gaudi chips, though with limited market success. Elon Musk's AI5 chip and plans for a custom fabrication facility represent another competitive threat. Despite these efforts, the sheer scale of the AWS-NVIDIA commitment suggests that even trillion-dollar companies view NVIDIA's ecosystem as essential infrastructure, at least in the near term.

The technical depth of NVIDIA's advantage stems from decades of engineering work. David Kirk, often called the "father of CUDA," persuaded Jensen Huang to mandate that every NVIDIA GPU include dedicated general-purpose execution units, a massive financial gamble in the mid-2000s that ultimately enabled general-purpose GPU computing. Ian Buck then transformed this raw hardware capability into CUDA, an accessible software platform that allowed researchers to train AlexNet on NVIDIA GPUs in 2012, sparking the modern deep learning boom.

What Does This Mean for Enterprise AI Adoption?

The 2 million GPU commitment signals that AWS expects explosive growth in enterprise AI workloads over the next two years. Companies are moving beyond experimentation with large language models (LLMs) and deploying AI agents, retrieval systems, and physical robotics at scale. The partnership's focus on cost optimization through GPU-accelerated analytics and vector search suggests AWS is responding to enterprise concerns about AI infrastructure expenses.

NVIDIA's Nemotron open-model family will remain available as managed serverless offerings on Amazon Bedrock and for customized deployment on Amazon SageMaker, giving enterprises flexible options for deploying and fine-tuning AI models without building infrastructure from scratch.

The infrastructure race is intensifying globally. While AWS and NVIDIA focus on enterprise and government workloads, other regions are pursuing alternative strategies. The scale of this partnership underscores a fundamental reality: building competitive AI infrastructure requires not just GPUs but integrated ecosystems spanning hardware, networking, memory systems, and software platforms. For now, that ecosystem remains NVIDIA's domain.