The 2025 NVIDIA GTC was a crucial event that provided a glimpse into the future of artificial intelligence (AI) technology. The next-generation AI chips, announced through CEO Jensen Huang's keynote, surpassed the limits of computing performance and foreshadowed revolutionary changes in various industries. In particular, NVIDIA's technology leadership was further consolidated in all areas where AI can be applied, from data centers to autonomous systems, robotics, and scientific computing. This GTC went beyond a simple product launch, presenting a new paradigm for the AI era and attracting the attention of the global technology community.
At GTC 2025, NVIDIA unveiled two core AI chip architectures. First, the Blackwell Ultra, scheduled for release in the second half of 2025, is designed to further enhance the performance of the existing Blackwell chip, processing more tokens per second. This will significantly improve the speed of generative AI workloads such as large language models (LLMs). Then, the Vera Rubin architecture, to be released in the second half of 2026, provides over twice the AI inference performance of the Blackwell chip, at 50 petaflops (PetaFLOPS), and boasts twice the speed of the Grace Blackwell CPU with the custom CPU Vera. The roadmap includes Rubin Ultra in 2027, accelerating the progress of AI computing even further.
The new AI chips go beyond simply increasing speed and maximize the efficiency of AI workloads. Vera Rubin provides 1.2 exaflops (ExaFLOPS) of performance in FP8 training, and memory bandwidth increases from 8TB/s to 13TB/s through HBM4 technology. In addition, NVLink throughput has doubled to 260TB/s, dramatically increasing the data transfer speed between GPUs. The BlueField-4 DPU, which accelerates data center infrastructure, offers 800Gbps throughput and 6 times better computing performance than the previous generation, efficiently handling networking, storage, and security tasks. These technological advances are further highlighted by a new AI performance metric called ‘token generation efficiency’, making companies reconsider their AI infrastructure.
NVIDIA is opening the era of generative AI based on next-generation AI chips. At GTC 2025, the concept of ‘AI Factory’ was emphasized, a blueprint for designing and operating data centers on the scale of several gigawatts as if they were industrial plants. The Omniverse DSX platform was announced to build digital twins of these AI factories, which can reduce design and deployment time by up to 500 times and increase efficiency by 30%. In addition, personal AI supercomputers such as DGX Spark and DGX Station based on Blackwell Ultra were introduced to support AI development from small researchers to large companies, lowering the barriers to AI development.
The advancement of AI chips also has a significant impact on the fields of autonomous systems and robotics. NVIDIA is leading robotics innovation by announcing Isaac GR00T N1, an open-source model for developing humanoid robots, and the Cosmos AI model for generating synthetic data for robot training. It is accelerating the development of self-driving vehicles and robotaxi systems through cooperation with General Motors (GM) and Uber, and the NVIDIA DRIVE Hyperion ecosystem is establishing itself as the core platform for these autonomous systems. The concept of ‘physical AI’, which understands the physical world, will be a foundational technology that helps robots operate more sophisticatedly in real-world environments.
NVIDIA is building the largest AI supercomputer in the United States in collaboration with the U.S. Department of Energy (DOE) and Oracle. The Solstice system, to be deployed at the Argonne National Laboratory, will be equipped with 100,000 Blackwell GPUs, and the Equinox system will provide a total of 2,200 exaflops of AI performance with 10,000 Blackwell GPUs, accelerating scientific research. Furthermore, it announced NVQLink, an open system architecture that closely connects GPU computing and quantum processors in preparation for the quantum computing era. In addition, it is contributing to the development of future communication technologies by integrating AI services into 6G communication platforms through a strategic partnership with Nokia.
NVIDIA's AI technology is creating new opportunities in various industries. In the field of cybersecurity, it is building an AI-based security system utilizing Nemotron-based models and NeMo tooling through collaboration with CrowdStrike. In the field of industrial automation and digital twins, it is revolutionizing factory design and operation through Omniverse DSX, maximizing productivity. The transition to ‘edge AI’, which runs AI directly on devices and networks as well as the cloud, will enable real-time decision-making in various fields that require low latency and high security, such as healthcare, finance, and automation.
NVIDIA GTC 2025 clearly showed that AI is not just a tool, but a driving force for a new industrial revolution that will define the future of all industries and countries. Next-generation AI chips like Blackwell Ultra and Vera Rubin will accelerate innovation in numerous fields such as generative AI, autonomous systems, scientific computing, and robotics through unprecedented performance improvements. NVIDIA is building AI infrastructure through an integrated approach to hardware and software, and is constantly expanding the application of AI technology through collaboration with partners, further strengthening its leadership in the AI era. This vision is giving a clearer picture of what an AI-based society will look like in the future.
A1: At GTC 2025, Blackwell Ultra (released in the second half of 2025), Vera Rubin (released in the second half of 2026), and Rubin Ultra, which is included in the 2027 roadmap, were announced as the main AI chips.
A2: The Vera Rubin chip provides over twice the AI inference performance of the Blackwell chip, at 50 petaflops, and the custom Vera CPU boasts twice the speed of the Grace Blackwell CPU.
A3: It is mainly expected to be applied in a wide range of fields, including training and inference of generative AI models, construction of large-scale data centers (AI factories), autonomous driving and robotics, scientific computing and quantum computing integration, 6G communication, cybersecurity, and industrial digital twins.
A4: ‘AI Factory’ is a concept that designs and operates AI data centers with a gigawatt scale as if they were industrial plants. It has the advantage of reducing design and deployment time by up to 500 times and increasing efficiency by 30% by building a digital twin through the Omniverse DSX platform.
0