Nvidia unveils updated AI solutions amid design flaws, industry shifts

Nvidia CEO Jensen Huang showcased the company's latest advancements in artificial intelligence (AI) hardware and software at its annual developer conference in San Jose, California, aiming to solidify its position in a rapidly evolving market. The announcements, spanning new chips, networking solutions and robotics frameworks, underscored Nvidia's push to address growing demands for efficient AI training and inference.

Huang introduced the Blackwell Ultra graphics processing unit (GPU), set for release in the second half of 2025, emphasizing its expanded memory capacity to support larger AI models. However, the current Blackwell products are facing manufacturing delays due to a design flaw complicating its rollout as the industry shifts focus from training AI systems to deploying them for real-world inference tasks.

Nvidia also introduced the Vera Rubin computing system, which combines a custom-designed processor with next-generation GPUs to outperform the Blackwell architecture. Slated for launch in late 2026, Vera Rubin will be followed by the Vera Rubin Ultra in 2027 and Feynman architecture in 2028.

Named after pioneering astronomer Vera Rubin, the system targets hyperscale AI workloads, with improved chip-to-chip data transfer speeds critical for complex models.

For developers, Nvidia unveiled DGX personal AI computers powered by Blackwell Ultra chips and built by partners like Dell, HP and Lenovo. The desktop systems aim to rival high-end consumer devices, enabling local inference of large models. "This is what a PC should look like," Huang remarked, holding a motherboard during the presentation.

Software updates included Dynamo, a free tool to accelerate AI reasoning, and Isaac GR00T N1, a robotics framework featuring a dual system "fast and slow thinking" model. Developed with Google DeepMind and Disney Research, GR00T N1 integrates Newton, an open-source physics engine for advanced robot simulation.

Despite Huang's confident assertions that Nvidia remains "well positioned" for AI's transition to inference-heavy workloads, investor skepticism lingered. Shares fell 3.4 percent following the presentation, reflecting concerns over competition and delays. Huang dismissed doubts, arguing that agentic AI and reasoning-driven tasks could require "100 times more computation" than previously anticipated.

(With input from agencies)