NVIDIA Rubin: Everything You Need to Know About NVIDIA’s Next-Generation AI Platform

The development of artificial intelligence requires ever-increasing computing power. While just a few years ago the primary use case for GPUs was training neural networks, today the industry faces a new challenge — scaling generative AI, agentic systems, and reasoning models.

It is precisely to address these challenges that NVIDIA introduced the Rubin platform — a new generation of AI accelerators that will succeed the Blackwell architecture.
In this article, we will explore what NVIDIA Rubin is, the technologies behind the new platform, and why it could transform the approach to building AI infrastructure.

What Is NVIDIA Rubin?

Although Rubin is often referred to as a new GPU, this is not entirely accurate.
NVIDIA Rubin is a comprehensive AI computing platform that integrates GPUs, CPUs, networking adapters, DPUs, and high-speed data transfer infrastructure.
The platform is named after American astronomer Vera Rubin, whose research played a crucial role in the study of dark matter.
Rubin includes:

• NVIDIA Rubin GPUs;
• NVIDIA Vera processors;
• sixth-generation NVLink switches;
• ConnectX-9 SuperNIC network adapters;
• BlueField-4 DPUs;
• Spectrum-6 Ethernet networking solutions.

Instead of individual servers, NVIDIA proposes using data center-scale racks, where all components are optimized to work together from the ground up.

Why NVIDIA Had to Move Beyond Traditional GPUs

Modern large language models contain hundreds of billions of parameters and process massive amounts of data.
The primary bottleneck is no longer the performance of individual accelerators, but the speed of data exchange between them.
Training and running next-generation models requires:

• enabling high-speed communication between GPUs;
• reducing inference latency;
• efficiently scaling workloads across thousands of accelerators;
• lowering the cost of generating a single token;
• optimizing data center energy consumption.

That is why Rubin is not a standalone GPU but a comprehensive platform designed for building so-called AI factories.

Key Innovations in NVIDIA Rubin

Next-Generation HBM4 Memory

Rubin will be one of NVIDIA's first platforms to support HBM4 memory.
The new memory delivers significantly higher bandwidth compared to the HBM3E used in Blackwell.
For large language models, this means:

• faster inference;
• support for longer context windows;
• fewer accesses to external storage;
• improved efficiency for agentic systems.

Third-Generation Transformer Engine

The new Transformer Engine introduces support for the NVFP4 compute format and hardware-based data compression. This increases compute density while reducing inference costs.
According to NVIDIA, Rubin delivers up to 10× lower token generation costs than Blackwell in certain use cases.

Sixth-Generation NVLink

One of the key upgrades is NVLink 6.
Each Rubin GPU provides up to 3.6 TB/s of interconnect bandwidth between accelerators. In the NVL72 configuration, total bandwidth reaches 260 TB/s.
This approach enables dozens of GPUs to function as a single computing system.

The New NVIDIA Vera Processor

Alongside Rubin, the company introduced the Vera processor, based on NVIDIA's proprietary Arm architecture.
It features 88 compute cores and uses the NVLink-C2C interface to exchange data with GPUs.
Vera is responsible for task orchestration, data preparation, and auxiliary computations.

Security and Reliability

Rubin introduces third-generation Confidential Computing and an updated RAS framework.
The new architecture provides:

• hardware-level data protection;
• isolated computing environments;
• automated hardware diagnostics;
• predictive failure detection.

This is especially important for enterprise AI systems that process sensitive information.

Rubin vs. Blackwell: What's Changed?

When Will Rubin-Based Servers Become Available?

The first commercial systems powered by NVIDIA Rubin are expected in the second half of 2026.
Among the companies planning to adopt the new platform are leading cloud providers and AI companies:

• AWS;
• Microsoft Azure;
• Google Cloud;
• Oracle Cloud;
• OpenAI;
• xAI;
• Anthropic;
• Mistral AI.

Rubin is expected to become the foundation for the next generation of AI data centers.

Where to Rent a GPU Server for AI Workloads Today

Although Rubin-based servers will not arrive for several years, many AI workloads can already be efficiently deployed in the cloud.
If you need to train neural networks, run large language models, generate images, or analyze large datasets, there is no need to invest in expensive hardware.
The Serverspace cloud platform allows you to quickly deploy a GPU server and access computing resources on a pay-as-you-go basis.
Benefits of renting GPUs in the cloud include:

• server deployment in minutes;
• hourly billing with no capital expenditure;
• flexible resource scaling;
• support for Docker, Kubernetes, and JupyterLab;
• convenient management through APIs and the control panel;
• rapid AI and machine learning project deployment.

Using cloud infrastructure allows teams to focus on developing models and applications instead of maintaining hardware.

Conclusion

NVIDIA Rubin is not just another generation of GPUs.
The company is changing the very approach to building AI infrastructure, moving from standalone accelerators to fully integrated computing platforms.
New GPUs, HBM4 memory, sixth-generation NVLink, and Vera processors are designed to solve the biggest challenge in modern AI — efficiently scaling agentic systems and reasoning models.
Over the coming years, platforms like Rubin will become the foundation of next-generation AI factories, while competition among infrastructure providers will be determined not only by GPU performance but by the efficiency of the entire ecosystem.

FAQ