Nvidia Unveils Next-Gen AI Chips and Innovations at GTC Conference

At its annual GTC conference, Nvidia announced a new generation of chips designed to advance artificial intelligence capabilities. CEO Jensen Huang introduced Blackwell Ultra, a powerful family of chips set to launch in the second half of 2024, along with Vera Rubin, the company’s next-generation GPU architecture expected in 2026. Also, Deepseek’s AI capabilities were addressed.

Nvidia’s rapid expansion in the AI sector has been driven by the increasing demand for its high-performance GPUs, particularly since the rise of OpenAI’s ChatGPT in 2022. The company dominates the market for AI model development, a process known as training, and is closely watched by investors and software developers eager to see how its latest advancements stack up.

New Chip Innovations and Nvidia’s Growth Strategy

Huang emphasized that the global demand for AI computing is accelerating at an unprecedented rate. With its latest annual release cycle, Nvidia aims to introduce new chip architectures every year—a shift from its previous biannual launch schedule.

The GTC conference, held in San Jose, California, serves as a key platform for showcasing Nvidia’s industry influence. This year’s event is expected to attract 25,000 attendees and feature participation from major companies like Microsoft, Waymo, Ford, and General Motors.

As part of its long-term roadmap, Nvidia revealed that the chip architecture following Vera Rubin will be named after renowned physicist Richard Feynman and is expected to launch in 2028.

Vera Rubin: A Leap in AI Processing

Set for release in 2026, the Vera Rubin system consists of two main components:

• Vera (CPU) – Nvidia’s first custom CPU, built on an architecture called Olympus. Unlike previous designs that relied on Arm’s off-the-shelf processors, Vera is fully tailored for enhanced performance. Nvidia claims it will be twice as fast as the Grace Blackwell CPU from 2023.
• Rubin (GPU) – A next-generation graphics processing unit designed to deliver 50 petaflops for AI inference, more than double the performance of current Blackwell GPUs. It can also support 288GB of high-speed memory, a crucial factor for AI developers.

A major shift in Nvidia’s terminology was also introduced. While previous Nvidia GPUs were composed of multiple chips fused together, Rubin will be described as two GPUs rather than one. The company plans to release an upgraded “Rubin Next” chip in 2027, which will combine four GPU dies, further doubling performance.

Blackwell Ultra: Pushing AI Speed and Revenue

Nvidia also introduced Blackwell Ultra, an upgraded version of its Blackwell family of chips. The new chip is optimized to generate AI tokens per second at a significantly higher rate, meaning it can process and generate content much faster than its predecessors.

This enhanced speed offers a major advantage to cloud providers, enabling them to offer premium AI services for time-sensitive applications. According to Nvidia, these improvements could allow cloud companies to generate up to 50 times more revenue compared to the 2023 Hopper generation.

Blackwell Ultra will be available in multiple configurations:

• GB300 – A version with two GPUs paired with an Nvidia Arm CPU
• B300 – A GPU-only version
• Eight-GPU blade server configuration
• Rack-mounted version with 72 Blackwell chips

Nvidia noted that major cloud providers have already deployed three times the number of Blackwell GPUs compared to Hopper.

DeepSeek and AI Reasoning Capabilities

Nvidia also addressed concerns about China’s DeepSeek R1 model, which initially raised fears among investors. Released in January, DeepSeek R1 was seen as a potential threat because it reportedly required fewer chips to achieve strong AI performance.

However, Nvidia has embraced DeepSeek’s advancements, arguing that the model relies on AI reasoning, which actually increases computational demands at runtime. The Blackwell Ultra chips are optimized for reasoning-based AI, making them well-suited for handling these sophisticated models.

“In the last 2 to 3 years, a major breakthrough happened, a fundamental advance in artificial intelligence happened. We call it agentic AI,” Huang explained. “It can reason about how to answer or how to solve a problem.”