Illustration of GPU infrastructure at AWS and NVIDIA scale. The image is a conceptual picture, not a photograph of a real data centre.
AWS and NVIDIA plan to add two million GPUs to AWS infrastructure during 2027 to 2028. The investment extends beyond GPUs to processors, high-speed networking and AI platforms. For enterprises planning AI workloads, the important question is what this expansion could make possible, which capabilities are already available and what should influence infrastructure decisions today.
Two million GPUs and what they mean for enterprises
In August 2026, AWS and NVIDIA announced plans to deploy two million additional NVIDIA GPUs across AWS infrastructure during 2027 to 2028. The expansion includes NVIDIA Blackwell Ultra, Rubin and Rubin Ultra GPUs. It builds on an earlier collaboration involving more than one million GPUs.
The scale of the announcement is significant, but the number alone does not explain what it means for organizations building or operating AI applications. A company training its own AI models has different infrastructure needs from one using managed AI services. A robotics company developing autonomous machines faces different demands again.
AWS and NVIDIA are investing across these areas, expanding not only accelerator capacity but also processors, networking, AI models and data-processing technologies. The companies identify growing demand for agentic AI, physical AI, scientific research and other computationally intensive workloads as reasons for the investment.
For enterprise architects, the useful question is not whether two million GPUs is an impressive number. It is what infrastructure AWS is preparing to offer, which workloads could benefit and what remains uncertain until the new capacity is actually available.
Source: AWS and NVIDIA announcement, August 26, 2026
What AWS and NVIDIA are building
The partnership covers several connected parts of AI infrastructure. The largest visible component is GPU capacity. GPUs perform the large volumes of parallel computation used in training and running many AI models. For demanding workloads, their memory capacity and communication speed can be as important as raw computing performance.
The planned expansion includes three NVIDIA GPU generations: Blackwell Ultra, Rubin and Rubin Ultra. But a large computing system cannot operate efficiently through GPUs alone.
When a model is distributed across many accelerators, those processors must exchange information quickly. Delays in moving data can reduce the benefit of adding more computing power.
AWS and NVIDIA are therefore deepening their collaboration on networking, including NVIDIA Spectrum technology and AWS Elastic Fabric Adapter. They are also planning infrastructure based on NVIDIA Vera CPUs to support processing tasks around AI workloads. Additional collaboration covers NVIDIA Nemotron models on Amazon Bedrock and SageMaker, accelerated data processing and technologies used in robotics development.
| Infrastructure | Planned development | Relevance |
|---|---|---|
| Blackwell Ultra, Rubin and Rubin Ultra | Additional GPU capacity | Large-model training and inference |
| NVIDIA Vera CPUs | CPU infrastructure on AWS | Supporting computation around AI workloads |
| Spectrum and AWS networking integration | High-bandwidth communication | Distributed computing across accelerator clusters |
| Nemotron and AI software | Model and platform integration | Additional options for building and operating AI applications |
| Accelerated data processing | GPU-assisted data and vector processing | Preparing and retrieving information used by AI systems |
| Physical AI collaboration | Robotics simulation and development | Training and evaluating AI for physical environments |
These developments are connected, but they do not serve identical workloads. A cloud architect planning a large GPU cluster may need to examine networking and accelerator communication closely. An organization using managed AI services may never directly manage those components. The partnership is designed to support both kinds of demand through AWS’s broader infrastructure and service offerings.
Source: NVIDIA technical announcement
Why agentic and physical AI are part of the investment
AI workloads are becoming more varied. Training a large language model, running an AI assistant and simulating a robotic system place different demands on computing infrastructure.
Agentic AI is one area AWS and NVIDIA expect to drive additional demand. An AI agent may use a model to interpret a request, retrieve information, call software tools and work through several steps before producing a result. Some of that activity requires model inference, which may run on GPU-backed infrastructure. Other parts use CPUs, databases, storage and external services.
This helps explain why the announced investment extends beyond GPUs. Physical AI introduces another set of requirements. Systems being developed for robots and autonomous machines may need to learn from simulated environments, process visual information and respond to changing physical conditions.
Building and testing these capabilities can involve extensive computation and large datasets. AWS and NVIDIA’s collaboration includes technologies supporting this work, alongside more conventional model training and inference.
However, the announcement does not establish how much of the new capacity will ultimately be used by agentic AI, physical AI or other workloads. Those are areas of anticipated demand rather than a published allocation of the two million GPUs.
What is available today and what is still planned
An important distinction is that the two-million-GPU announcement describes future infrastructure. It does not mean the entire capacity is already available to AWS customers. Some of the underlying technologies are already entering AWS’s infrastructure.
In August 2026, AWS announced additional regional availability for EC2 P6-B300 instances, which use NVIDIA Blackwell Ultra GPUs. A P6-B300 instance provides eight GPUs, approximately 2.1 TB of combined GPU memory and up to 6.4 Tbps of networking through Elastic Fabric Adapter. These capabilities are relevant to organizations running large models that require substantial accelerator memory or distributed computation.
Rubin and Rubin Ultra, by contrast, are included in the future expansion. Their presence in the roadmap should not be interpreted as general availability.
This difference matters when planning infrastructure. An architect may be able to evaluate an existing instance type using actual specifications and pricing, while a future accelerator is still a planning consideration.
Regional availability creates another constraint. AWS may announce global infrastructure growth, but that does not establish which regions will receive particular instance families, how much capacity will be available or when customers will be able to provision it. Organizations with specific data-residency or latency requirements must therefore evaluate availability in their intended deployment regions rather than relying on the global GPU figure.
Source: AWS EC2 P6-B300 regional expansion, August 2026
AWS is also building an alternative to NVIDIA
While AWS is expanding its NVIDIA infrastructure, it is continuing to develop its own AI accelerators. In December 2025, AWS announced general availability of Trainium3 UltraServers, offering another platform for large-scale AI training and inference.
This makes the infrastructure expansion more interesting than a straightforward increase in NVIDIA GPU capacity. For organizations operating their own models, the choice between NVIDIA and Trainium involves both hardware and software. NVIDIA-based workloads commonly depend on CUDA and its supporting libraries. Moving those workloads to Trainium may require compatibility work, model validation and performance tuning using AWS Neuron.
AWS’s continued investment in Trainium means that customers may have more accelerator options, but each platform introduces its own technical and operational considerations. The practical question is whether a particular model performs better, costs less or is easier to operate on one platform than another. That cannot be determined from GPU specifications alone.
Source: AWS Trainium3 announcement
The architecture decisions that matter
The most immediate implications of the AWS and NVIDIA expansion are for organizations that operate AI infrastructure directly. For large-model training, accelerator memory and interconnect performance can determine how efficiently work is distributed across GPUs.
For inference, the considerations are different. Architects must balance throughput, response latency, model memory requirements and hardware utilization. A deployment optimized for high request volumes may differ substantially from one designed to provide low latency for a small number of complex requests.
These differences affect instance selection, cluster design and operating costs. They also influence whether an organization should manage dedicated inference infrastructure or consume a managed model service.
For example, a company operating a continuously used, specialized model may have reasons to evaluate self-hosted inference. Another organization with irregular demand and limited infrastructure expertise may find managed services more suitable. Neither approach follows automatically from the GPU announcement.
The value of the expansion is that additional hardware generations and supporting technologies could create more deployment choices as they become available.
The practical constraint: getting the capacity you need
AWS’s global GPU expansion does not remove the challenges involved in obtaining specialized compute capacity. An organization may need a specific accelerator generation, sufficient GPU memory, compatible software and multiple instances available together in one region.
For distributed workloads, allocating machines across locations that cannot provide the necessary communication performance may undermine the intended design.
AWS offers capacity reservation mechanisms, including EC2 Capacity Blocks for ML, for certain accelerator workloads. However, supported instance types, reservation periods and regional availability depend on the service offering.
For organizations planning large AI projects, capacity procurement must therefore be considered alongside the technical architecture. The 2027 to 2028 expansion could improve future options, but the announced global quantity does not indicate how much capacity a particular customer can obtain in a particular region.
Source: Amazon EC2 Capacity Blocks for ML
What remains uncertain
The scale of AWS and NVIDIA’s investment is substantial, but delivery depends on more than manufacturing GPUs. Data-centre space, electricity, cooling, networking and installation schedules all affect when infrastructure can become operational.
Future pricing is another unknown. The announcement does not establish the cost of Rubin-based AWS instances, the relative economics of different future platforms or how capacity will be distributed across regions. Those details will matter when enterprises evaluate specific deployment options.
For now, the announcement provides a technology roadmap and a signal of intended infrastructure growth, rather than a complete basis for procurement decisions.
What the expansion tells us
AWS and NVIDIA are preparing for AI workloads that demand more computing capacity and more capable supporting infrastructure. The planned expansion includes new GPU generations, CPUs, networking and software designed to support model training, inference and physical AI.
At the same time, AWS is developing its own accelerator ecosystem, creating another direction for customers evaluating AI hardware. The significance for enterprises will depend on which technologies become available and how well they serve particular workloads.
The two-million-GPU announcement is important not because every enterprise will need more GPUs, but because it represents an expansion of the computing platforms available for increasingly demanding AI applications.
Sources and further reading
- AWS and NVIDIA to Deliver 2 Million Additional GPUs and Next-Generation Infrastructure for Agentic and Physical AI, AWS, August 26, 2026.
- NVIDIA Announcement: AWS and NVIDIA Infrastructure Expansion, NVIDIA, August 26, 2026.
- Amazon EC2 P6-B300 Instances Available in Additional Regions, AWS, August 2026.
- Amazon EC2 Trainium3 UltraServers, AWS, December 2025.
- Amazon EC2 Capacity Blocks for ML, AWS documentation.
- Energy and AI, International Energy Agency, 2025.
Related reading on TechiesJournal
- Cloud Capacity Is Not the Same as Cloud Availability for why a region’s capacity is not the same as the capacity you can get.
- AI Has Money and Chips. Why Electricity Is Becoming the Real Constraint for the power side of the delivery risk.
- CPU vs GPU vs TPU vs NPU vs LPU vs DPU Explained for how the chip types in this announcement differ.