- Nvidia (NVDA) unveiled Blackwell Ultra chips shipping in 2024 and the Vera Rubin architecture – named after astronomer Vera Rubin who discovered dark matter – coming in 2026, with CEO Jensen Huang projecting continued advancement to Feynman chips by 2028.
- The Vera Rubin system pairs Nvidia’s first custom CPU design with a new GPU architecture capable of 50 petaflops for inference workloads—more than double the current Blackwell chips’ 20 petaflops.
- Cloud providers could generate up to 50 times more revenue using Blackwell Ultra compared to 2023’s Hopper chips, with the top four cloud companies already deploying three times more Blackwell chips than Hopper.
In a significant demonstration of its continued dominance in the artificial intelligence hardware market, Nvidia (NVDA) revealed its ambitious product roadmap at its annual GTC conference. CEO Jensen Huang announced the Blackwell Ultra family shipping in the second half of this year and unveiled plans for the Vera Rubin architecture expected in 2026, with a clear path extending to the Feynman architecture – named after physicist Richard Feynman – slated for 2028.
The announcements come at a critical juncture for Nvidia, whose sales have increased more than sixfold since OpenAI’s ChatGPT catalyzed the generative AI boom in late 2022. With 25,000 attendees expected at the San Jose conference, the company is showcasing not just its technical innovations but the ecosystem of partnerships it has cultivated with industry giants including Microsoft (MSFT), Google (GOOG), Amazon (AMZN), Waymo, and automotive companies like Ford (F) and General Motors (GM).
“This last year is where almost the entire world got involved. The computational requirement, the scaling law of AI, is more resilient, and in fact, is hyper-accelerated,” Huang noted during his presentation, highlighting the explosive growth in computational demands for AI applications.
The Blackwell Ultra chips represent Nvidia’s immediate response to these escalating requirements. The new chips are designed to produce more tokens per second than their predecessors, enabling faster content generation for time-sensitive applications. This performance enhancement could allow cloud providers to generate up to 50 times more revenue compared to the Hopper generation chips that shipped in 2023. The top four cloud companies have already deployed three times more Blackwell chips than Hopper chips, according to Nvidia.
Looking further ahead, the Vera Rubin architecture marks a significant evolution in Nvidia’s approach. For the first time, Nvidia will produce its own custom CPU design called Vera, based on a core design named Olympus. This represents a departure from its previous strategy of using off-the-shelf Arm (ARM) designs. The custom Vera CPU is projected to be twice as fast as the CPU used in last year’s Grace Blackwell chips.
When paired with the new Rubin GPU, the system can manage an impressive 50 petaflops for inference workloads—more than double the 20 petaflops capability of the current Blackwell chips. Rubin will also support up to 288 gigabytes of fast memory, a critical specification for AI model development.
Perhaps most interestingly, Nvidia is redefining what constitutes a GPU in its nomenclature. Starting with Rubin, when Nvidia combines multiple dies to create a single chip, it will refer to them as separate GPUs. The planned “Rubin Next” chip, scheduled for the second half of 2027, will combine four dies and be referred to as four GPUs, doubling the speed of the initial Rubin design. This system will be delivered in a rack called Vera Rubin NVL144, an upgrade from previous NVL72 racks.
Nvidia is also embracing new AI methodologies, particularly “reasoning” models like China’s DeepSeek R1, which initially concerned some investors when released in January. Huang has positioned these developments as advantageous for Nvidia, noting that reasoning models require more computing power—exactly what Blackwell Ultra chips are optimized to provide.
“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.”
Beyond its high-end data center offerings, Nvidia continues to diversify its product line with AI-focused PCs like the DGX Spark and DGX Station, capable of running large AI models such as Llama or DeepSeek locally. The company also announced updates to its networking components for connecting thousands of GPUs to function as unified systems, plus a new software package called Dynamo designed to optimize chip performance.
As Nvidia accelerates its product release cadence from biennial to annual, the company is betting that the exponential growth in AI computational demands will sustain its business momentum. The reception of these announcements by cloud providers and AI developers will be crucial in determining whether Nvidia can maintain its dominant position in the increasingly competitive AI chip market.
WallStreetPit does not provide investment advice. All rights reserved.
Leave a Reply