Amazon Web Services and NVIDIA expanded their partnership at AWS re:Invent with new hardware, cloud services and software for training and running generative AI. The announcement connected NVIDIA’s Grace Hopper systems and AI stack with AWS technologies including the Nitro virtualisation system, Elastic Fabric Adapter networking and UltraCluster scaling.
The collaboration covered several layers at once: GH200 superchips on AWS, DGX Cloud training as a service, three new EC2 instance families and a planned supercomputer called Project Ceiba. It also extended into application software for retrieval, summarisation, drug discovery and digital simulations.
Grace Hopper Systems Bring Tighter Compute Integration
AWS said it would become the first cloud provider to offer NVIDIA GH200 Grace Hopper Superchips with multi-node NVLink technology. The GH200 combines CPU and GPU capabilities for data-intensive AI workloads, while the NVL32 platform is intended to scale across thousands of superchips.
Connecting these systems through AWS networking and virtualisation aims to give customers supercomputer-class capacity without owning a facility. The value depends on more than peak hardware performance. Large training jobs require reliable communication among nodes, storage that can feed data fast enough and software that recovers when components fail. AWS and NVIDIA are aligning these parts rather than presenting the accelerator as an isolated product.
DGX Cloud and EC2 Offer Different Paths to Capacity
NVIDIA DGX Cloud was set to run on AWS using the GH200 NVL32 platform. This provides AI training as a service through an environment designed by NVIDIA while using AWS infrastructure. Customers that want more direct control could use new EC2 options, including P5e instances powered by H200 Tensor Core GPUs for generative AI and high-performance computing.
The range gives organisations a choice between a more integrated NVIDIA experience and familiar AWS primitives. A managed service can reduce setup time, while EC2 instances may fit teams that already operate custom cloud pipelines. Buyers still need to compare availability, pricing, data movement and software portability. Access to advanced hardware is useful only when the workload can be scheduled and funded predictably.
Project Ceiba Demonstrates the Partnership’s Scale
AWS and NVIDIA planned Project Ceiba with 16,384 GH200 superchips and a stated 65 exaflops of AI processing. The companies described it as a candidate for the world’s fastest GPU-powered AI supercomputer. NVIDIA would use the system for its own research and development, making AWS both a customer platform and the host of major internal NVIDIA workloads.
Projects at this scale show how frontier AI development is concentrating around a small number of infrastructure providers. Building and operating the system requires specialised hardware, networking, power and engineering expertise. The resulting research may improve tools available to customers, but concentration also creates dependencies. Organisations should understand which parts of their stack can move elsewhere and which are tied to a particular cloud and accelerator ecosystem.
Software Extends the Alliance Beyond Hardware
NVIDIA announced software on AWS including a NeMo Retriever microservice for chatbots and summarisation and BioNeMo tools intended to accelerate drug discovery. Amazon’s robotics and fulfilment teams were already using NVIDIA Omniverse to simulate and optimise warehouses before changing real facilities.
These applications demonstrate the commercial logic of the partnership. Hardware demand grows when customers can reach it through software tied to practical tasks. Retrieval systems, scientific models and digital twins turn abstract compute capacity into workflows for enterprises. They also add more layers that must be secured, monitored and updated.
The alliance can shorten deployment by providing components tested together. At the same time, a tightly integrated stack may make later substitution difficult. A team should decide whether speed to production outweighs the value of portability and document the interfaces that would be needed to change providers.
Security responsibility is distributed across the stack. AWS can protect the cloud foundation, NVIDIA can maintain drivers and AI software, and customers must still configure identities, networks and data access. Integrated products reduce some compatibility work but do not turn a multi-party environment into a single security boundary.
Energy and utilisation are also relevant even though the source focuses on technical capacity. A supercomputer or large GPU cluster creates value when it is used efficiently on worthwhile workloads. Cloud access can improve sharing compared with every organisation building private infrastructure, but demand for training and inference can still grow rapidly.
Procurement teams should therefore ask for utilisation and cost evidence rather than reserve the largest instance by default. Smaller experiments can identify bottlenecks before a project scales, and model optimisation may reduce hardware demand. Advanced chips are a scarce resource; matching them to justified workloads is both an economic and operational discipline.
The announcement captures how generative AI competition is becoming a systems contest. Accelerators, CPUs, interconnects, virtualisation and model software must work together. Neither NVIDIA nor AWS can deliver the complete environment as effectively alone: NVIDIA supplies specialised compute and AI tooling, while AWS provides global cloud operations and customer access.
For enterprises, the practical question is not whether the infrastructure is powerful. It is which service level matches the workload and what commitments come with it. A research team training a large model has different needs from a company adding retrieval to an internal assistant.
AWS and NVIDIA’s expanded relationship offers both groups more options inside one ecosystem. Its success will be measured by availability, reliability and cost as much as exaflops. The partnership makes frontier infrastructure easier to consume, while reminding customers that convenience and concentration often arrive together.
