Written by: Techub News Compilation

At the recent GTC conference held in Washington, D.C., NVIDIA founder and CEO Jensen Huang delivered a nearly two-hour keynote speech filled with information. This speech was not only a concentrated presentation of NVIDIA's technological blueprint but also a forward-looking declaration of America's leadership in technological innovation and the future industrial revolution. At a critical juncture where competition in artificial intelligence is heating up and the global technological landscape is being reshaped, Huang's speech revealed how computing paradigms are undergoing fundamental shifts and how NVIDIA is attempting to play the role of a core infrastructure provider in this transformation.

Accelerated Computing and AI Factory: The Rise of a New Computing Paradigm

Huang pointed out at the beginning that the computing industry is experiencing a once-in-60-years paradigm shift. With the slowdown and even end of Moore's Law, traditional general-purpose computing can no longer meet the exponentially growing computational demands, especially those driven by artificial intelligence. NVIDIA's focus on accelerated computing for the past 30 years has finally come to fruition.

However, accelerated computing is not simply about hardware replacement. Huang emphasized that its core lies in a complete software stack and ecosystem—CUDA and over 350 CUDA-X libraries built atop it. From the computing lithography library cuLitho (adopted by TSMC, Samsung, ASML) to the AI training framework Megatron Core, and the medical imaging framework Monai, these libraries have redesigned algorithms to enable developers to fully leverage accelerated hardware. He referred to this as the company's "treasure," as it is these software layers that have transformed NVIDIA's GPUs from graphics processors into general-purpose parallel computing engines, ultimately becoming central to the AI era.

One of the core concepts of the speech was the "AI Factory." Huang elaborated that past computing centers were general-purpose and could run various applications; future AI infrastructure will resemble a "factory" focused on a single product. The "product" of this factory is the "Token"—the digital unit generated by AI after processing information. Whether text, images, videos, 3D structures, or proteins and chemical molecules, everything can be "tokenized" and learned by AI in terms of its "language" and meaning.

"AI is not a tool, AI is a worker," Huang made a profound observation. Past software (like Excel and browsers) were tools used by humans. AI is the "worker" that can utilize these tools. For instance, Cursor AI writes code using VS Code, and Perplexity AI uses browsers to plan trips. This shift means that AI will participate directly in value creation, touching upon economic areas that past tool software did not cover, with scales reaching trillions of dollars.

The driving force of the AI factory comes from two "exponents": first, the exponential growth in computational demand needed for increasingly intelligent AI models (across pre-training, post-training, and inference/"thinking" stages); second, the smarter the model, the more users there are, which also results in exponential growth in computational demand. These two exponents place immense pressure on global computing resources against the backdrop of Moore's Law failing. Huang's answer is "extreme collaborative design"—full-stack redesign from chips, systems, networks, and software to model architecture and applications, rather than just designing faster chips.

Blackwell and Rubin: The Hardware Engine Driving Exponential Growth

Huang detailed the results of NVIDIA's extreme collaborative design: the Grace Blackwell platform, particularly its NVLink 72 architecture. He likened NVLink 72 to the "shield of Captain America," connecting 72 GPUs into a giant virtual GPU. This design allows each GPU to serve only a few "expert" model subsets instead of handling the entire large model, significantly improving computational efficiency and token generation speed.

According to third-party benchmarking, Grace Blackwell achieved a performance of 10 times per GPU compared to the previous generation H200. More critically, although the GB200 system itself is costly, the cost of generating each token is the lowest in the world. This ensures that the "virtuous cycle" of AI development continues: lower costs drive broader usage, generating more profits to invest in stronger computing, thus producing smarter models.

Huang revealed that NVIDIA's visibility into early demand for Blackwell and the next-generation Rubin platform has reached $500 billion (excluding the Chinese market). In comparison, the previous generation Hopper shipped 4 million GPUs over its entire lifecycle, while Blackwell shipped 6 million GPUs (each containing two GPUs) in just the first few quarters, with an expectation to deliver Blackwell products equivalent to 20 million GPUs over the subsequent five quarters, a growth rate five times that of Hopper. This intuitively reflects the market's fervent demand for AI computing power.

A video showcasing the manufacturing process of Blackwell chips in the U.S. was played during the speech, starting from silicon wafers in Arizona, to HBM stacking in Indiana, and final assembly in Texas. Huang specifically mentioned that bringing manufacturing back to the U.S. is "necessary" for national security and economic employment. He disclosed that just nine months after the related initiative was proposed, Blackwell has entered full production in Arizona.

Huang also previewed the next-generation Rubin platform, showcasing its third-generation NVLink 72 rack-scale computer, and emphasized that it will achieve "fully cable-free" and 100% liquid cooling. He announced that NVIDIA is building an AI factory digital twin blueprint called "Omniverse DSX," allowing partners to collaboratively design, simulate, and optimize the entire system from construction, power, cooling, to AI infrastructure in the digital world before physical factories are built, significantly shortening construction time.

Quantum Computing, 6G, and National-Level Cooperation: Laying Out Next-Generation Infrastructure

In addition to AI, Huang also announced significant progress in the fields of quantum computing and communication infrastructure.

In quantum computing, NVIDIA released the "NVQ Link" interconnect architecture aimed at directly connecting quantum processors (QPU) with GPU supercomputers. Quantum bits are extremely fragile and require complex error correction mechanisms, which in turn need robust classical computing power. NVQ Link and the CUDA-Q open platform enable researchers to conduct quantum error correction, device calibration, and quantum-classical hybrid simulations. Huang announced that 17 quantum computing companies and 8 U.S. Department of Energy (DOE) national laboratories support this new architecture.

Following that, he announced another major news: the U.S. Department of Energy is collaborating with NVIDIA to build seven new AI supercomputers to advance national scientific endeavors. Huang praised Energy Secretary Chris Wright's determination to keep the U.S. at the forefront of scientific advancements.

In the field of communication, NVIDIA announced an important partnership with Nokia, the world's second-largest telecommunications equipment manufacturer, to jointly promote the 6G revolution. NVIDIA launched a new product line called "NVIDIA Arc," which is a software-defined, programmable wireless communication and AI processing integration platform based on Grace CPU, Blackwell GPU, and ConnectX networking technology. Nokia plans to integrate Arc technology into its future base stations and use it to upgrade millions of existing global base stations.

Huang pointed out that current global wireless technology relies heavily on foreign technology, and the fusion of 6G and AI (AI for RAN and AI on RAN) represents a "once-in-a-lifetime opportunity" for U.S. technology to return to the core of the communications field. AI for RAN can utilize reinforcement learning to adjust beamforming in real-time, enhancing spectral efficiency (consuming about 1.5-2% of global power); AI on RAN can build edge industrial robot clouds on radio communication networks, pushing computing power to the network edge.

Open-source Models, Enterprise Empowerment, and the Ecosystem of Physical AI

Huang emphasized the importance of open-source AI models for developers, startups, and national competitiveness. He announced that NVIDIA has become a leader in open-source contributions, with 23 models leading various rankings across fields such as language, physical AI, and biology. He reiterated that the U.S. needs to lead the development of open-source models while having powerful proprietary models.

At the enterprise application level, NVIDIA is deeply integrating its CUDA-X libraries and AI models into the ecosystems of major cloud service providers (AWS, Google Cloud, Microsoft Azure, Oracle) and enterprise software giants. He specifically announced two new enterprise collaborations:

• Collaboration with cybersecurity firm CrowdStrike to create a "light-speed" cybersecurity system deploying AI security agents in the cloud and on-premises.
• Collaboration with data analytics firm Palantir to accelerate its data processing and commercial insight platform to address massive data challenges for governments and enterprises.

The latter part of the speech focused on "physical AI"—making AI understand and act upon the physical world. Huang noted that physical AI requires three computers to work together: the Grace Blackwell (GB200) for training models, the Omniverse computer for simulating and training robots in digital twins, and the Jetson Thor robot computer deployed in robots or autonomous vehicles.

He showcased a digital twin simulation of building NVIDIA AI infrastructure system robot manufacturing factories in Houston, Texas, in collaboration with Foxconn. The factory completed design, verification, and robot AI training in Omniverse before physical construction began. He also mentioned collaborations with Figure, Agility, Johnson & Johnson, and Disney in the robotics domain, forecasting that humanoid robots may become one of "the largest new markets in consumer electronics."

In the autonomous driving sector, NVIDIA announced its DRIVE Hyperion platform, which provides automotive manufacturers with a "robotic taxi-ready" standard sensor and computing chassis. Mercedes-Benz, Lucid, and Stellantis have already designed vehicles based on this platform. The more significant announcement is that NVIDIA will collaborate with Uber to connect Hyperion-based autonomous vehicles to the global ride-hailing network.

Conclusion: Dual Platform Transformation and America's "Apollo Moment"

At the end of the speech, Huang reviewed the historical milestones of innovation in America, from transistors and ARPANET to personal computers and the internet. He referred to the current era driven by accelerated computing and artificial intelligence as America's next "Apollo moment."

He concluded that the world is simultaneously experiencing two major platform transformations: from general computing to accelerated computing, and from hand-written software to artificial intelligence. NVIDIA's CUDA and CUDA-X ecosystems enable it to reach almost every industry, and it has now reached a "turning point," beginning to grow like a virtuous cycle. Advances in fields such as quantum computing, 6G, open-source models, enterprise AI, and physical robotics collectively outline a future landscape redefined by AI across industry, science, and society.

The entire speech was not only a showcase of NVIDIA's technology products but also an ambitious roadmap on how to build national technological competitiveness, reshape manufacturing, and lead the next industrial revolution. The core message Huang conveyed is that through extreme full-stack collaborative design, building open ecosystems, and deeply integrating digital twins with the physical world, we stand at the dawn of a new era, its scale and impact potentially comparable to the moon landing program.

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。