NVIDIA
Yet ANOTHER Huang chip
At CES, Nvidia CEO Jensen Huang introduced the company’s new Rubin computing architecture, describing it as the current standard in AI hardware.
The system is already in production, and Nvidia expects output to grow further in the second half of the year.
Huang said the architecture is meant to address a basic challenge: modern AI requires a rapidly increasing level of computation.
Rubin was first announced in 2024 and will replace the Blackwell platform, which itself followed earlier Nvidia designs such as Hopper and Lovelace.
The new chips are scheduled for use across leading cloud providers.
Nvidia has confirmed partnerships with firms including Anthropic, OpenAI and Amazon Web Services.
Rubin systems will also be deployed in HPE’s Blue Lion supercomputer and the upcoming Doudna supercomputer at Lawrence Berkeley National Lab.
Rubin is made up of six specialised chips that work together.
The Rubin GPU sits at the centre, supported by updates to Bluefield for storage and NVLink for interconnection. A new Vera CPU is included as part of the design, aimed at running agentic reasoning tasks.
Nvidia noted that newer AI workflows place heavy demands on memory tools such as the KV cache, especially when models handle long inputs or ongoing tasks.
To ease this bottleneck, Nvidia has created a new storage tier that connects to the compute device and allows organisations to expand their storage pools in a more efficient way.
Key points in brief
Nvidia launched the Rubin architecture at CES with chips already in production.
The platform will be used by major cloud providers and national lab supercomputers.
Nvidia reports strong speed and efficiency improvements over the previous generation.
Bottlenecks meet their match
According to Nvidia’s internal tests, Rubin offers clear gains in performance.
The company claims the architecture runs 3.5 times faster than Blackwell on model-training tasks and five times faster on inference tasks.
It can reach up to 50 petaflops of peak compute and deliver far more compute per watt.
Demand for AI infrastructure continues to rise worldwide.
On a 2025 earnings call, Huang estimated that between $3 trillion and $4 trillion will be spent on AI infrastructure over the next five years.
Nvidia naming chips after scientists keeps the nerd aesthetic strong… or maybe it’s after Rick Rubin?- MG