Jensen Huang Unveils Nvidia’s Upcoming A.I. Chips at GTC: What to Know

At the recent GTC 2025 developer conference held in San Jose, California, on March 18, Nvidia CEO Jensen Huang unveiled two groundbreaking GPU architectures: Blackwell Ultra and Rubin. He referred to this annual event as the "Super Bowl of A.I.", emphasizing that unlike a traditional Super Bowl, "At this Super Bowl, everyone wins."

The Blackwell Ultra chip is scheduled to launch in the latter half of 2025, while its successor, Rubin, is set for a late 2026 release. An even more advanced version, known as Rubin Ultra, will follow in 2027.

What to Know About Blackwell Ultra and Rubin

Nvidia’s Blackwell Ultra series is an enhanced version of the Blackwell chips showcased at last year's GTC. These new chips provide an impressive 40 times more processing power compared to the previous-generation Hopper chips. The Blackwell Ultra will feature eight stacks of 12-Hi HBM4E memory, allowing for a total of 288GB of onboard memory. Additionally, it will utilize NVLink 72, a high-speed interconnect technology that facilitates communication between GPUs and CPUs, which is essential for handling large datasets needed for A.I. training and inference.

Jensen Huang explained, "NVLink connects multiple GPUs, turning them into a single GPU. It addresses the scale-up problem by enabling massive parallel computing."

Moreover, the conference featured the introduction of the Nvidia RTX PRO 6000 Blackwell Server Edition, designed for enterprise tasks including multimodal A.I. inference, advanced content creation, and scientific computing. This server edition comes with 96GB of GDDR7 memory and supports multi-instance GPU technology, positioning it as a powerful tool for A.I. development.

The upcoming Rubin chip is named after astronomer Vera Rubin, who played a significant role in discovering dark matter. The initial Rubin model aims to achieve a speed of 50 petaflops during A.I. model execution, while the Rubin Ultra is expected to reach up to 100 petaflops, a substantial advancement in A.I. capabilities.

In response to reports of previous overheating issues with early Blackwell chips, Nvidia unveiled the liquid-cooled Grace Blackwell 200 NVL72 system. This system can deliver real-time inference up to 30 times faster for large language models and allows for training that is four times quicker than Nvidia's previous H100 GPU. It can process up to 12,000 tokens per second, enhancing both training and inference speeds significantly.

Huang mentioned, "If you want your A.I. to be smarter, it must generate more tokens. That requires massive bandwidth, floating-point operations, and memory." He noted that advanced reasoning A.I. models, such as DeepSeek's R1, demand 20 times more tokens and 105 times more computing power.

Nvidia is also partnering with Taiwan’s TSMC to develop new data packaging technologies aimed at improving computational efficiency and thermal performance for future GPU generations.

Nvidia's Roadmap Beyond Rubin

Looking forward, Huang outlined that the next generation of architectures following Rubin will be called Feynman, named after the famous physicist Richard Feynman. The Feynman HBM is currently in development and is expected to be released in 2028. This architecture aims to further enhance A.I. performance to levels previously unseen.

These announcements come on the back of Nvidia’s strong financial performance, driven by increasing demand for its GPUs. Despite facing stiff competition from rivals like AMD and geopolitical challenges, Nvidia continues to dominate the GPU market, boasting an estimated 80 percent market share.

Nvidia, A.I., chips