New Delhi: At CES 2026, NVIDIA did not talk about a single chip. Instead, the company talked about an entire AI machine. The announcement focused on Rubin,
a new computing platform meant to power the next wave of large AI systems, from model training to real world inference.
The message from the stage was clear. AI workloads are getting heavier, models are getting larger, and running them is getting expensive. NVIDIA says Rubin is its answer to that pressure. It replaces the idea of buying one powerful processor with the idea of building AI as a full stack, from silicon to software to networking.
What NVIDIA Rubin actually is
Rubin is not one chip. It is a platform built using six tightly linked components. NVIDIA says these parts are designed together, not separately.
The Rubin platform includes:
- NVIDIA Vera CPU
- NVIDIA Rubin GPU
- NVLink 6 switch
- ConnectX 9 SuperNIC
- BlueField 4 DPU
- Spectrum 6 Ethernet switch
NVIDIA says this extreme codesign cuts inference token cost by up to ten times compared to its Blackwell platform. It also reduces the number of GPUs needed to train mixture of experts models by four times.
Jensen Huang, founder and CEO of NVIDIA, said, “Rubin arrives at exactly the right moment, as AI computing demand for both training and inference is going through the roof.”
Why Rubin matters for modern AI
Large AI models now do more than answer one question. They reason across long conversations, remember context, and act like agents. This type of work pushes data centers harder than before.
Rubin is built for that style of AI. NVIDIA says the platform focuses on:
- Faster training for large models
- Lower cost inference at scale
- Better efficiency per watt
- Long context and reasoning workloads
One detail NVIDIA highlighted is inference context memory. Rubin introduces a new storage platform that helps models reuse memory across sessions. This matters for agent style AI that needs to remember past steps.
Inside the Vera Rubin systems
NVIDIA will ship Rubin in two main forms. The first is the Vera Rubin NVL72 rack. It combines 72 Rubin GPUs and 36 Vera CPUs into one rack scale system. NVIDIA says the NVLink fabric inside this rack moves 260 terabytes of data per second.
The second option is the HGX Rubin NVL8. This version links eight Rubin GPUs and targets more traditional server setups.
Both systems support confidential computing. NVIDIA says this protects AI workloads across CPU, GPU, and networking layers.
Networking and power efficiency
Rubin also brings a new networking push. NVIDIA Spectrum X Ethernet Photonics is designed to improve uptime and power efficiency. The company claims five times better power efficiency and longer uptime for AI workloads.
This matters for data centers that now run nonstop. Power and cooling costs are no longer side issues.
As someone who has watched cloud bills explode during model testing, this part feels very real.
Industry backing and cloud plans
Rubin is not launching into a quiet market. NVIDIA listed a long group of partners, including cloud providers, AI labs, and hardware makers.
Sam Altman said, “The NVIDIA Rubin platform helps us keep scaling this progress so advanced intelligence benefits everyone.”
Satya Nadella said Microsoft will use Vera Rubin systems in future AI superfactories.
Sundar Pichai said Google Cloud will bring Rubin to customers needing scale and performance.
NVIDIA says Rubin based products will arrive through partners in the second half of 2026. Early deployments will include AWS, Google Cloud, Microsoft Azure, Oracle, and CoreWeave.
A shift in how AI hardware is sold
Rubin shows a shift in NVIDIA’s thinking. The company is no longer selling just GPUs. It is selling ready made AI factories.
The platform approach reduces custom work for data centers. It also locks hardware, networking, and software into one system.
That may simplify deployment for enterprises. It also deepens reliance on NVIDIA.
