Nvidia has revealed its road map to maintain leadership in the business of hardware For artificial intelligence (AI). The firm announced the launch dates of its Blackwell Ultra, Vera Rubin and Rubin Ultra chips, which offer significant performance improvements.
During its annual developer conference, GTC 2025, the company confirmed that the Graphic Processing Units (GPU), Blackwell Ultra GB300, will be available in the second half of 2025. This chip represents an evolution of the original Blackwell, with 20 Petaflops (PF) performance at AI, but with a significant increase in its high performance and bandwidth memory capacity, from 192 to 288 Gigabytes (GB).
This indicates that the new Nvidia chip is designed to process and generate more Tokens For its predecessor, which allows you to offer IA support services for time sensitive applications and generate up to 50 times more income compared to the generation of Hopper GPU, which debuted in 2023.
The Ultra Blackwell will be available in a cluster called DGX GB300 ‘SUPERPOD’, which maintains the configuration of 288 Central Processing Unit (CPU) and 576 GPU, but now it offers the possibility of accessing a expanded memory of 240 to 300 terabytes (TB), with 11.5 Computer Exaffuting FP4 (four floating point precision bits per operation). This increase optimizes its ability to train and execute Ia models more effectively. The new chips will also be available in a version called B300, which only includes a GPU; in a variant with eight GPU on a single Blade server, and in a configuration for a Rack with 72 microprocessors.
The Blackwell Ultra platform, according to NVIDIA, is ideal for applications that include AI agents with sophisticated reasoning and iterative planning to solve complex problems with multiple steps, as well as for physical AI models. This will allow companies to generate synthetic and photorealist videos in real time for the training of applications in robotics and large -scale autonomous vehicles.
“The market is increasingly demanding with AI, since new technologies and tools, such as AI agents, require more and more processing capacity, and Blackwell Ultra is Nvidia’s response to this demand,” said Marcio Aguiar, director of the Enterprise division of Nvidia for Latin America. “We want AI to continue transforming the world, reaching new levels of innovation and progress.”
Nvidia wants to continue dominating the AI
Nvidia intends to go further with architecture Vera Rubin, which will be available during the second quarter of 2026 and will offer 50 Petaflops of inference FP4, more than double the Blackwell Ultra. The company ensures that a Rack Complete of this platform should offer more than three times the performance of an ultra comparable Blackwell.
Finally, the company advanced the launch of Rubin Ultra during the last six months of 2027. This architecture will consist of two interconnected Rubin GPU, which will allow it to provide a capacity of 100 Petaflops FP4 and a 1 TB memory.
Nvidia has generated income of $ 11,000 million with Blackwell architecture. Only its four main buyers have acquired 1.1 million chips so far from 2025, according to the corporation.
Jensen Huang, founder and CEO of Nvidia, emphasized that the industry will need “100 times more power than we thought last year” to address the growth of AI. He said that the next architecture of the company, after Vera Rubin, will be called Feynman, in honor of the theoretical physicist Richard Feynman, and will arrive in 2028.
Some financial analyzes indicate that NVIDIA currently obtains about $ 2,300 in profits per second thanks to the AI revolution. Although after its launch in January, the Chinese Deepseek R1 model generated concern among the firm’s investors, the company glimpses great opportunities around this technology. Nvidia plans to use the software to evaluate the performance of your new products.
Despite the initial concerns about competition, Huang believes that Depseek is a positive sign, since its focus on ‘reasoning’ requires greater processing power, which benefits the new Blackwell ultra chips
#Nvidia #presents #Blackwell #Ultra #Rubin #Ultra #maintain #leadership
