NVIDIA has introduced its next-generation AI computing strategy at GTC 2025, showcasing the Blackwell Ultra and Vera Rubin AI chips while emphasizing a shift toward reasoning AI.
CEO Jensen Huang framed these developments within NVIDIA’s broader vision of AI Factories, large-scale compute infrastructures designed to handle real-time inference and autonomous decision-making AI.
The company also highlighted new enterprise AI partnerships with Oracle, GE Healthcare, and Yum! Brands, signaling a push to integrate AI-powered automation across industries. These advancements mark NVIDIA’s transition from AI training into large-scale, real-time AI deployment.
The Move Toward Reasoning AI
At the core of NVIDIA’s AI strategy is reasoning AI, a leap beyond traditional generative models. Unlike conventional AI systems that generate responses based on pattern recognition, reasoning AI models are designed to perform multi-step decision-making, improving contextual awareness and autonomy.
Leading this effort is the Llama Nemotron family, NVIDIA’s new open reasoning AI models. These are optimized for applications requiring deep contextual understanding, such as AI agents, enterprise automation, and industrial AI solutions.
The company also unveiled new Cosmos World Foundation Models (WFM), which enable real-world AI simulations and synthetic training data generation, particularly for robotics and autonomous systems.
Blackwell Ultra and Vera Rubin: AI Hardware for the Future
The introduction of Blackwell Ultra signals a significant upgrade in AI inference hardware. Expected to ship in the second half of 2025, this GPU offers 20 petaflops of AI performance and 288GB of HBM3e memory, enabling larger models to run without partitioning. It also supports FP4 precision, allowing for more efficient processing of AI reasoning tasks.
Following Blackwell Ultra, NVIDIA confirmed the development of the Vera Rubin AI architecture, set to launch in 2026. This architecture will introduce the Vera Rubin Ultra in 2027, delivering 50 petaflops of FP4 performance—a leap in computational capacity aimed at enterprise-scale AI inference workloads.
However, the increasing power of these chips raises concerns about energy efficiency and heat dissipation. While NVIDIA claims performance improvements, previous problems suggest that chip manufacturing constraints could impact availability and scaling.
DGX AI Supercomputers: Expanding AI Access
Alongside its high-end AI chips, NVIDIA introduced DGX Spark, a compact $3,000 AI supercomputer designed for researchers, developers, and startups. Powered by the GB10 Blackwell Superchip, DGX Spark aims to democratize AI computing.
For enterprise users, the company launched the DGX Station, featuring the GB300 Blackwell Ultra chip with 20 petaflops of performance and 784GB of unified memory. This machine is tailored for high-performance AI inference workloads in healthcare, finance, and automation.
NVIDIA’s enterprise AI expansion was reinforced by a partnership with Oracle Cloud, enabling businesses to deploy AI inference at scale. Collaborations with GE Healthcare and Yum! Brands were also announced, showcasing AI’s growing impact on medical diagnostics and automated retail operations.
AI Factory Strategy
As NVIDIA expands AI adoption across industries, its AI Factory vision aims to scale inference workloads for real-time applications. Throughout the keynote, NVIDIA emphasized the concept of AI Factories—large-scale data centers optimized for AI reasoning and inference.
These are designed to support the next wave of AI applications, from autonomous systems to real-time industrial simulations.
Despite the excitement surrounding these advancements, NVIDIA’s stock declined by 3.4% during the presentation, reflecting investor caution over AI infrastructure costs, chip availability, and competition from AMD and Google.
With Blackwell Ultra and Vera Rubin set to define the next phase of AI computing, industry analysts will be closely watching how well NVIDIA manages scalability, power efficiency, and market adoption.
