NVIDIA now combines GPUs, CPUs, networking, memory, software and rack-scale systems into integrated AI computing platforms.
Its latest Vera Rubin architecture targets large AI factories designed for pretraining, post-training, inference and increasingly complex agentic workloads.
NVIDIA faces competing accelerators, enormous power requirements, expensive data-centre construction, supply-chain challenges and evolving demand from major AI customers.
The artificial intelligence boom is changing the economics of computing, and NVIDIA is increasingly positioning itself at the centre of that transformation. Once best known for graphics processors, the company now supplies much of the hardware, networking and software needed to build and operate large-scale AI systems.
NVIDIA’s fiscal 2026 revenue reached USD 215.9 billion, up 65% from the previous year, underlining how quickly demand for AI infrastructure has expanded. Its strategy now extends from individual accelerators to complete data-centre architectures designed for training and running increasingly demanding AI models.
The foundation remains the GPU, but NVIDIA’s proposition has broadened significantly. AI models perform enormous numbers of mathematical operations in parallel, making GPUs well suited to training and inference, the process of using a trained model to generate results.
Instead of selling accelerators in isolation, NVIDIA combines GPUs with CPUs, high-bandwidth memory, networking, storage connectivity and software. Its rack-scale systems are designed so these components operate as a coordinated computing platform.
The Blackwell generation illustrates this approach. The GB200 NVL72, for example, combines 72 Blackwell GPUs with 36 Grace CPUs and high-speed networking in a single rack-scale system.
Modern AI systems rely on clusters containing hundreds or thousands of accelerators. A single large model can be divided across these processors, allowing many GPUs to work simultaneously.
Memory is equally important. NVIDIA’s Blackwell systems use HBM3E, a type of high-bandwidth memory positioned close to the GPUs. The GB200 NVL72 provides 13.4 TB of HBM3E with aggregate memory bandwidth of 576 TB/s.
NVIDIA is now advancing this architecture through its Vera Rubin platform. The Vera Rubin NVL72 combines 72 Rubin GPUs and 36 Vera CPUs with NVLink 6, ConnectX-9 SuperNICs, and BlueField-4 DPUs. NVIDIA describes it as a rack-scale system designed for reasoning and agentic AI workloads.
Adding more GPUs does not automatically make an AI cluster faster. The processors must constantly exchange model parameters, intermediate results and other data. If the network becomes a bottleneck, expensive GPUs can sit idle.
That makes high-speed networking a critical part of NVIDIA’s infrastructure strategy. NVLink connects GPUs within a system, while NVIDIA’s InfiniBand and Spectrum-X Ethernet technologies connect systems across larger clusters.
With Vera Rubin, NVIDIA says sixth-generation NVLink provides 260 TB/s of aggregate bandwidth within an NVL72 rack. The company is also expanding its networking business, which recorded USD 14.8 billion in data-centre networking revenue in fiscal 2027’s first quarter, up 199% year over year.
Also Read: NVIDIA DLSS 5 is Coming to RTX 40 Series GPUs Soon
Hardware alone cannot run modern AI efficiently. NVIDIA’s software ecosystem has therefore become a major part of its infrastructure proposition.
CUDA provides the programming foundation that allows developers to use NVIDIA GPUs for accelerated computing. Above it sit libraries, frameworks and tools for training, inference and deployment.
NVIDIA AI Enterprise brings together components including NIM microservices, NeMo, TensorRT, Triton Inference Server, GPU drivers and Kubernetes-based management tools. NIM, for instance, packages optimised inference software so organisations can deploy AI models more easily across data centres and cloud environments.
This software layer helps NVIDIA turn hardware into a repeatable platform rather than a collection of individual components.
NVIDIA uses the term “AI factory” to describe infrastructure designed to convert data and computing capacity into AI outputs. Unlike a conventional data centre, an AI factory is optimised around the entire AI lifecycle, from data processing and model training to high-volume inference.
The concept is becoming increasingly important as companies move from experimenting with generative AI to operating AI services at scale. NVIDIA’s platforms integrate compute, networking, software and management so operators can build large systems using validated architectures.
The company is also extending the model geographically. NVIDIA and Australian data-centre partners announced plans to add up to 2 gigawatts of AI-related capacity in Australia by 2027, using NVIDIA’s DSX platform.
NVIDIA’s position is powerful, but it is not unchallenged. AMD is expanding its Instinct accelerator portfolio, while hyperscalers including Google, Amazon and Microsoft are developing or deploying their own AI chips.
Energy is another constraint. Large AI clusters require substantial electricity and increasingly sophisticated cooling systems. Building them also involves expensive data-centre construction, advanced networking equipment and access to semiconductor supply chains.
NVIDIA also depends heavily on large cloud and technology customers whose spending decisions can influence the pace of infrastructure expansion. Recent investor concerns about a possible slowdown in AI spending show that the current buildout is not immune to changing expectations.
At the same time, NVIDIA is trying to make its architecture more open to specialised processors. Chip startup d-Matrix plans to use NVIDIA’s NVLink Fusion to connect its inference processors into NVIDIA-based systems, with integrated systems expected in 2027.
NVIDIA’s transformation is significant because the AI race is increasingly becoming a systems race. The winning infrastructure must move data quickly, feed processors with enough memory, run models efficiently, and deliver results at a commercially viable cost.
NVIDIA is attempting to provide that entire stack, from accelerators and memory to networking, servers and software. Its advantage therefore extends beyond the performance of any single GPU.
If AI continues shifting towards larger models, real-time inference and autonomous systems, the infrastructure underneath those applications will become just as important as the models themselves. NVIDIA’s strategy suggests that its biggest role in the AI revolution may ultimately be not simply designing the chips that power AI, but building the computing architecture on which the next generation of AI operates.
Also Read: Nvidia’s $96 Billion Quarter Shows the AI Infrastructure Boom Is Still Going
NVIDIA is integrating computing, networking, memory and software into complete platforms designed to efficiently train and deploy increasingly sophisticated artificial intelligence models.
An AI factory is a data-centre architecture designed to transform computing resources and data into AI outputs through training, inference and agentic workloads.
Vera Rubin is NVIDIA’s next-generation AI computing platform, combining new GPUs, CPUs, networking and other components for large-scale agentic AI infrastructure.
Large GPU clusters constantly exchange data, making high-speed networking essential for preventing communication bottlenecks and ensuring expensive processors remain efficiently utilised.
NVIDIA faces competition from alternative accelerators, enormous energy requirements, costly data-centre expansion, supply constraints and potential changes in hyperscaler AI spending patterns.