Written by: Techub News Compilation

At the recent GTC 2025 conference held in Washington, D.C., NVIDIA founder and CEO Jensen Huang delivered a nearly two-hour keynote speech. This speech is not only the most important annual AI and computing technology barometer but also attracted attention for the intensive release of heavy products and collaborations involving AI infrastructure, 6G communication, quantum computing, robotics, and other fields. Huang systematically elaborated on how NVIDIA addresses the computational challenges of the post-Moore's Law era through "extreme co-design" and outlined a next-generation industrial blueprint driven by AI.

The Era of Accelerated Computing: From CUDA to AI Factories

At the beginning of the speech, Huang reviewed the history of technological innovation in the United States and pointed out that we are in a new era opened by revolutionary new computational models. He emphasized that the turning point of accelerated computing has arrived. This is not merely a hardware upgrade but a paradigm shift that has lasted for thirty years. NVIDIA invented the GPU and the CUDA programming model to solve problems that general CPUs could not handle. However, accelerated computing requires entirely new algorithms, libraries, and even rewriting applications, which is a long and arduous process.

Huang views the CUDA-X ecosystem as the company's "treasure." This includes over 350 libraries—from cuLitho used for computational lithography, to the medical imaging AI framework MONAI, and to Megatron Core that drives large language models—each redesigned algorithms for specific fields, opening new markets. He presented a video entirely generated by simulation, featuring various industries from healthcare to manufacturing, robotics, and autonomous driving, showcasing the power of mathematics and computer science, as well as the fact that accelerated computing has permeated all industries.

However, the rise of AI has completely reshaped the computing stack. Huang noted that the past software industry was about creating "tools" (such as Excel, browsers), while the essence of AI is "workers", intelligent agents capable of completing tasks using these tools. For example, Cursor, widely used by NVIDIA's internal engineers, is an AI partner that can use VS Code tools to write code. This shift means that AI will directly participate in the global economic production worth trillions of dollars, rather than merely serving the trillion-dollar IT tools market.

This has created a demand for new computational systems. Huang proposed the concept of "AI factories." Unlike traditional data centers that run various applications and store files, AI factories are highly specialized facilities whose core task is to efficiently and cost-effectively produce "smart tokens". These tokens can be text, images, protein structures, or robotic action instructions. AI factories need to process context in real time, decompose problems, reason, and execute plans, generating massive tokens at every step, which places an exponentially growing demand on computational power.

"People used to say that reasoning is simple, and training is the tough part," Huang said, "but how can thinking be simple? Regurgitating memorized content is simple; thinking is hard." He elaborated on three technological stages of AI development: pre-training (learning basic abilities), post-training (learning skills), and "thinking" (continuous reasoning and interaction based on new knowledge). The latter two, especially "thinking," place unprecedented pressure on computational infrastructure.

More critically, the AI industry witnessed a turning point last year: models have become smart enough that people are willing to pay for them. From Cursor to 11 Labs to OpenAI and Claude, paid usage has become the norm. This has formed a dual exponential growth flywheel: the smarter the model, the more it is used; the more it is used, the greater the computational power required; more computational investment makes the models smarter. However, in the context of Moore's Law nearing its end, how can such exponential demand be met?

Huang's answer is: Extreme Co-Design. NVIDIA is the only company today that can start from a blank sheet of paper, synchronously think about new computer architecture, chips, systems, software, model architecture, and applications. The latest Grace Blackwell NVLink 72 system embodies this concept. It is not just an upgrade of a single chip, but interconnects 72 GPUs through NVLink to form a giant virtual GPU, optimized for future AI models (such as Mixture of Experts models). The result is that, despite the number of transistors only doubling, each GPU performance has achieved ten times that of the previous generation H200, while realizing the lowest token generation cost globally.

Huang revealed that NVIDIA has gained visibility of orders worth $500 billion for Blackwell and early Rubin (excluding China and parts of Asia), which is equivalent to five times the total revenue throughout the Hopper lifecycle. He played a video showing the domestic manufacturing process of the Blackwell super-scale system, from silicon wafers in Arizona to assembly in Texas, emphasizing "we are making in the U.S. again."

Looking ahead to the next generation, Huang showcased the third-generation NVLink 72 rack-scale computer codenamed "Vera Rubin." It features an all-liquid cooling, cable-free design, and introduces a new "Context Processor" to handle the trend of AI needing to process increasingly longer contexts (such as reading large PDFs and watching videos). The bandwidth of the NVLink switch connecting all these processors is equivalent to the peak traffic of the global internet.

Reshaping Communication, Quantum, and Industrial Ecosystems

In addition to core AI computing, Huang announced several heavyweight plans to venture into new fields.

6G and Communication: NVIDIA launched a brand new product line—NVIDIA ARC (Aerial Radio network Computer). ARC is built on Grace CPU, Blackwell GPU, and ConnectX networking technology designed for applications, running a CUDA-X library named Aerial. Its core is to create a software-defined, programmable computer capable of wireless communication and AI processing simultaneously. NVIDIA has partnered with Nokia, the world's second-largest telecommunications equipment provider, which will use ARC as the core of its future base stations, compatible with existing AirScale base stations, meaning millions of base stations globally can upgrade to obtain 6G and AI capabilities through software. Huang pointed out that AI will be used in RAN (Radio Access Network) to improve spectrum efficiency while simultaneously building an edge industrial robotics cloud on top of the RAN, similar to how AWS builds cloud computing on the internet.

Quantum Computing: NVIDIA released NVQ Link, a new interconnect architecture that directly connects quantum processors (QPU) with NVIDIA GPUs. Quantum error correction requires extremely high data throughput and extremely low latency, and NVQ Link was designed for this purpose. At the same time, its open platform CUDA-Q has been expanded to support QPU, allowing QPUs and GPUs to work together with microsecond-level latency. Huang announced that 17 quantum computing companies and 8 U.S. Department of Energy (DOE)-affiliated national laboratories support NVQ Link. Furthermore, the DOE is collaborating with NVIDIA to build 7 new AI supercomputers to advance the frontiers of American science.

Enterprise Applications and Cybersecurity: NVIDIA is deeply integrating its libraries and models into the enterprise software ecosystem. Huang announced the collaboration with CrowdStrike, aiming to use AI to create "light-speed" cybersecurity defense systems. At the same time, NVIDIA will also collaborate with Palantir to accelerate the data processing capacity of its Ontology platform, providing real-time insights for massive structured and unstructured data for government and enterprises.

Open Source Models: Huang emphasized that open source models are essential for researchers, startups, and various industries, and the U.S. must maintain its leadership in the open-source field. NVIDIA is currently a leader in open-source contributions, with 23 leading models across multiple rankings covering languages, physics AI, robotics, biology, and more.

Physical AI and Robotics: Digital Twins Drive Future Manufacturing

Huang attributed the realization of physical AI (AI that understands the laws and causal relationships of the physical world) to the need for three computers: a Grace Blackwell supercomputer for training models; an Omniverse computer for simulating and training robots or factories in a digital twin (which requires robust generative AI and computer graphics capabilities); and a Jetson Thor robotics computer deployed inside robots or autonomous vehicles.

He showcased a collaboration with Foxconn to design, simulate, and optimize its AI infrastructure manufacturing plant located in Texas within the Omniverse digital twin. Robots in the factory (such as FANUC robotic arms) are trained in Isaac Sim, while AI visual agents monitor the entire process. This reflects the idea that "the factory itself is a robot coordinating other robots to manufacture robotic products."

Huang believes that humanoid robots are likely to become one of the largest consumer electronics and industrial equipment markets of the future. He mentioned collaborations with companies like Figure, Tesla (Elon Musk), and Agility Robotics. Notably, he highlighted the "Blue" robot project in partnership with Disney Research, which learns in a revolutionary physics simulation platform called "Newton." Huang emphasized that all movements of Blue seen by the audience are real-time simulations, not animations.

In terms of wheeled robots that are closer to commercialization, NVIDIA introduced the Nvidia Drive Hyperion architecture. This is a "robotaxi-ready" standard platform, containing a full set of sensors (360-degree cameras, radar, lidar), aimed at providing the foundation for automakers like Lucid, Mercedes-Benz, and Stellantis as well as autonomous system developers like Aurora, Momenta, and Nuro. Additionally, NVIDIA announced a partnership with Uber to connect Hyperion vehicles to a global network in the future.

At the end of the speech, Huang summarized that we are currently undergoing two significant platform transformations: from general computing to accelerated computing, and from traditional manual coding software to artificial intelligence. NVIDIA, through the CUDA and CUDA-X ecosystem, touches nearly all industries and has reached a growth inflection point. Quantum computing, open-source models, enterprise applications, robotics, 6G communication, and domestic manufacturing together constitute NVIDIA's blueprint for leading the next wave of industrial revolution.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。