Authored by: Techub News Compilation

In this dialogue, NVIDIA founder and CEO Jensen Huang systematically outlined his judgments on the next wave of artificial intelligence (AI) revolving around generative computing, AI infrastructure, re-industrialization, energy, intelligent agents, physical AI, open source, and the technology competition between China and the United States. Rather than viewing AI merely as a "chatting robot," it's better to consider it as a new computing paradigm: past computing was mainly about "storing and retrieving," while today's computing is shifting towards "understanding context, generating results, and executing tasks." This is not conceptual rhetoric, but a redefinition of industrial structure, talent structure, infrastructure, and national competitiveness.

Jensen Huang suggested that the mainstream computing method of the traditional internet era is essentially "retrieval-based computing." People would first write, record, and store content, then use mechanisms like search, recommendation, and advertising distribution to match existing content with users. The core resources behind this model are storage, indexing, and distribution capabilities; hence, data centers have long resembled "data warehouses." However, generative AI has changed this logic: systems are no longer just retrieving existing content but are generating previously non-existent answers, summaries, images, videos, and even action plans in real-time based on user intent, context, and goals.

This is what Huang refers to as "generative computing." Under this framework, computers are no longer just passively executing fixed instructions; they can perceive multi-modal information such as text, images, and videos, understand human needs, conduct reasoning and planning, and return results best suited to the current situation. Essentially, AI allows "computing" to shift for the first time from static information processing to dynamic intelligent generation, thus bringing about entirely new demands for computing power, system software, networks, energy, and data center forms.

From "Chip Company" to AI Infrastructure Company

For many members of the public, NVIDIA's most widely recognized label remains GPU or "chip-selling company." However, Huang emphasized repeatedly in the dialogue that NVIDIA does not only produce chips; it creates "the computational infrastructure for modern AI." This infrastructure includes chips, systems, system software, algorithms, middleware, and a complete technology stack tailored to various industrial scenarios, covering diverse environments such as cloud data centers, on-premises deployments, factories, communication base stations, and automobiles.

This definition is crucial because it explains why AI competition is never a point competition but an ecological competition. When AI enters the industrialization phase, what determines victory is not just how fast a certain generation of chips is or how many leads a model has on the rankings, but who can truly connect hardware, software, toolchains, development frameworks, industry solutions, and partner networks. The NVIDIA that Huang describes is more like a platform company building a "computational fabric" for the AI era, rather than a semiconductor company in the traditional sense.

Because of this, he views AI as a complete new industry and not merely a technological invention. In this industry, chips are the foundation, but far from everything; system integration, infrastructure operation, model development, industry applications, and large-scale adoption collectively form the value chain. If one only focuses on the most visible model layer, it often underestimates the depth and breadth with which AI truly changes society.

AI "Five-Layer Cake": A Good Framework for Understanding AI Industry Structure

Huang provided a highly summarized analytical framework during the dialogue, namely the "five-layer cake" of AI. These five layers, from bottom to top, are: energy, chips, infrastructure, models, applications, and adoption. This framework is significant because it places AI back into the coordinates of industrial systems and national capabilities, rather than merely as a "software hotspot."

The first layer is energy. Generative AI requires immense computational power, which cannot do without electricity supply. As large model training and inference become a continuous industrial activity, electricity is not just a cost item but a strategic resource that determines whether the industry can expand. The second layer is chips and systems, which are the physical devices that carry out computation. The third layer is infrastructure, including land, power access, server housing, cloud service software, etc., which enables computational resources to be genuinely deployed, operated, and continuously called upon.

The fourth layer is models. This is the layer currently receiving the most attention in public discourse, but Huang particularly reminded that models should not only be understood as large language models. AI can represent not only language and numbers but also various information structures including biology, chemistry, physics, and motion control. Therefore, the truly important model layer is a system of general or specialized models capable of serving fields like science, industry, healthcare, and robotics.

The fifth layer is applications and adoption, which refers to the diffusion process of technology into real society and real industries. Huang clearly stated that this is the layer he is most concerned about. Because even if a country possesses outstanding chips and models, if enterprises, institutions, and workers are unwilling to use AI, afraid to deploy AI, or do not incorporate AI into processes, then technological advancement cannot automatically translate into productivity, competitiveness, and economic advantages. According to his judgment, the United States must not fall behind in the application and adoption layers, otherwise it risks losing its first-mover advantage in a new round of industrial revolution.

AI and Re-industrialization: Huang's Strongest Policy Proposition

If there is a central theme throughout this dialogue, it is that "AI is not only a digital technology revolution, but also an opportunity for re-industrialization." Huang stated bluntly that the United States needs to re-industrialize and bring an entire manufacturing sector and corresponding jobs back to the homeland. He believes that the social structure over the past period has overly favored high education and knowledge-based professions, leaving many without a four-year college degree, master's or doctorate background feeling marginalized, which is neither necessary nor healthy.

In his view, AI offers an unprecedented market drive strong enough to promote re-industrialization. The reason is that AI will bring a whole host of new "factory" demands: first chip factories, then factories for computers and related systems, and further up, the demand to build and deploy "AI factories." The so-called AI factory does not produce traditional consumer goods but is a new type of production facility that converts computational power, models, and data into "token output." The economic logic of such a facility is the continuous transformation of power and capital input into intelligent output, which is then further converted into enterprise services, industrial decisions, software capabilities, and automated productivity.

Huang mentioned in the interview that NVIDIA has committed to large-scale procurement and supply chain investment to shift some capabilities from the East back to the West, supporting chips, packaging, computers, and other aspects to land in the United States. He described this round of investment as potentially bringing trillions of dollars in manufacturing activities and a large number of high-skilled, high-income jobs. Regardless of whether the outside world fully agrees with this scale judgment, this statement at least conveys one clear signal: in Huang's eyes, the ultimate goal of AI is not "smarter chat applications," but a reconstruction of the foundational industry.

Energy Issues: The First Hard Constraint of the AI Era

In the "five-layer cake," Huang places energy at the very bottom, and this is not accidental. His logic is very straightforward: whether it's training models, operating data centers, or manufacturing, fundamentally, they all rely on energy input. Manufacturing implies transforming the form of matter, which requires large amounts of energy; likewise, the operation of AI supercomputers also signifies continuous and large-scale electricity consumption.

Therefore, on the question of "whether more energy is needed," he hesitated little. The real issue is not whether to have more, but whether the nation decides to become a manufacturing powerhouse again and whether it is willing to reform its energy system around this goal. Once the answer is affirmative, upgrading the power grid, optimizing supply mechanism, introducing more flexible service level agreements, and developing nuclear energy, solar energy, and other sustainable energy would all become part of industrial policy.

Notably, Huang does not simplify the energy issue into a single-path battle. His focus is not on betting on a particular energy source, but on using the strong market demand brought by AI to push the United States to simultaneously accomplish two things: first, to rebuild manufacturing capabilities, and second, to upgrade the national energy system. From this perspective, AI is not a burden on the energy system but may become a catalyst for modernizing energy infrastructure.

Agent AI: The Real Breakthrough is Not Just in Models, But in "Orchestration Systems"

When discussing the next wave of AI, Huang believes that one of the most critical leaps in recent years is from large language models to chatbots, and then from chatbots to intelligent agent systems. In his view, the leap from models to chatbots hinges on "reinforcement learning from human feedback," making the models more usable and aligned. The leap from chatbots to agents is not solely about increasing model parameters, but about "harness"—the entire orchestration mechanism built around the model that connects real-world scenarios and tasks.

This orchestration mechanism enables models to possess several key abilities: accessing real information sources, utilizing browsers, conducting research, employing tools, retaining memory, collaborating with other systems, and continuously working through longer task chains. In other words, an agent is not a "better speaking" model, but a "more capable doing" system. This also explains why the progress of agents has been so significant in the past six months: what has truly been amplified is not the quality of single-round responses but the capability for complex task automation.

Huang particularly noted that the software development field has already been deeply affected by this change. He believes that most software tasks can now be highly automated, and programmers no longer need to personally complete substantial coding. But this does not mean software engineers will disappear; rather, it means they will spend more time understanding problems, defining objectives, coordinating resources, and driving innovation.

Will AI Eliminate Jobs? His Answer is Negative

Throughout the dialogue, Huang expressed a very clear opposition to the narrative that "AI will massively eliminate jobs." He believes such statements are not only distorted but can cause real harm to society, as they may scare young people away from professions that will still be extremely needed in the future, or even more so. He cited two examples: software engineering and radiology; AI is indeed altering specific tasks within these professions, but the purpose of a profession is not to complete a single action but to solve problems, create value, and bear responsibility.

One of his core distinctions is: tasks do not equal professions, and skills do not equal missions. For software engineers, writing code is merely one task; the true purpose of the profession lies in innovation, solving complex problems, collaborating with teams, and creating new products. For radiologists, reading images is just a means to achieve goals, and the essence of the profession is diagnosing diseases and supporting treatment decisions. When AI automates certain tasks, humans do not lose their occupational significance; rather, they have the opportunity to shift their focus to higher-level judgments and creativity.

Huang further pointed out that a common erroneous assumption is that the total demand for a task is fixed. For example, people may imagine that "only a fixed amount of code needs to be written," leading to a reduction in human programmers once AI can write code automatically. But his judgment is precisely the opposite: the real limitation on innovation is not too few problems but rather the limited manpower and time. When coding costs decrease, society is likely to tackle more previously unaddressed problems, leading to new products, companies, and employment needs.

This logic does not mean that all positions will remain unchanged. But it reminds people that when observing the relationship between AI and employment, they should not just focus on a particular action being replaced but on whether the overall objective function of the entire industry has been expanded. If AI increases business growth rate, expands the number of feasible projects, and lowers the barrier to innovation, then it is likely to lead to reorganized job contents and shifting job demands rather than job shrinkage.

Physical AI: From Autonomous Driving to Humanoid Robots

In addition to intelligent agents, Huang also talked about "physical AI," referring to the capability of AI to further enter the real world from the digital world. In his view, autonomous driving has already become one of the most mature breakthroughs in physical AI, with robotic taxis arriving and related issues shifting more towards engineering and large-scale deployment. He also mentioned cars with reasoning capabilities that can break down situations, understand combinations when encountering previously unseen road conditions, and make decisions, indicating that autonomous driving is moving from "pattern matching" to "context understanding."

As for humanoid robots, Huang's assessment is that "they are not far off." He explained that if video generation models can accurately generate actions like "picking up a coffee cup and taking a sip," then from a cognitive and representational perspective, the robot's ability to learn these actions is also not far away. Of course, the real difficulty lies not only in the AI model but also in the electromechanical systems: motors, dexterous hands, materials, structures, batteries, and sensors will all affect the timeline for the practical application of humanoid robots.

This assessment reveals an important piece of information: future competition in the AI industry will not stay at the model leaderboard but will extend into the "soft-hard integration" of complex engineering. Whoever can integrate perception, reasoning, control with electromechanical systems will have the opportunity to define the next generation of robotic platforms. This poses higher demands on manufacturing, supply chains, and industrial design capabilities than in the internet era.

U.S.-China Competition, Open Source, and AI Adoption: Huang's Triple Anxiety

In the latter half of the dialogue, Huang focused on three interconnected topics: U.S.-China competition, open source, security, and American society's attitude toward AI. On U.S.-China competition, he still used the "five-layer cake" framework, emphasizing that the United States must maintain a lead on energy, chips, infrastructure, models, and adoption in all five areas. He is particularly worried that if American society overly "sci-fi" and "doomsday" AI, resulting in fear among businesses and the public towards AI, then the U.S. may lose to regions that more positively embrace AI at the adoption level.

This is also why he repeatedly emphasizes the importance of the "adoption" layer. In his narrative, parts of Asia are embracing AI with stronger enthusiasm, while there is much portrayal in American public discussion of extreme consequences like unemployment, democratic collapse, and civilizational destruction. He believes these statements not only do not help establish a reasonable governance framework but also weaken the speed of technology diffusion, impact talent choices, and ultimately harm national competitiveness.

On the issue of open source, Huang's stance is also very clear: open source is not just a democratization tool, it can also enhance security. His reasoning is that facing powerful AI attacks in the future, defenders cannot rely solely on a single closed system but need many trainable, auditable, and rapidly deployable defense agents to work in coordination. Open source enables such "defense collectives" to become possible and allows more companies to see underlying mechanisms, implement sandboxes, permissions, privacy, and policy controls.

He cited examples to mention that concerns about enterprise security arising from open source models can be addressed by establishing outer "shells" and sandbox environments to limit the information that models can access, the data they can send, and the resources they can invoke. This approach is very similar to the governance methods most valued in today's enterprise AI deployment: not outright banning models but enabling models to operate safely within institutional and technical boundaries. From this, it can be seen that Huang supports "governable open ecosystems" rather than "unconstrained open source."

Implications for Readers Outside the U.S. and China

Although the entire conversation was framed around U.S. policy and industrial competition, Huang's views are equally meaningful for readers from other countries and regions. They at least highlight three points: first, AI strategies cannot solely focus on model launches but must return to energy, infrastructure, and industrial adoption; second, the core of AI competition is not who shouts the largest parameters first but who can form a complete technology stack and manufacturing capability; third, how society discusses AI will directly affect talent flow, enterprise investment, and the speed of technology diffusion.

From this perspective, "generative computing" is not merely a software upgrade, but a redesign of the way industries are organized. It calls for governments to re-examine the relationships between electricity, manufacturing, education, research, and regulation, and it requires enterprises to rethink organizational processes, job distinctions, and technology deployment methods. For ordinary people, the truly worthy question to consider may not be "Will AI replace me?" but "What can I participate in, create, and organize after AI reshapes task division?"

Huang's continued emphasis on "not scaring people away" is not to downplay risks, but because he believes the most realistic value of AI lies not in fear narratives but in productivity enhancement, industrial upgrading, and the expansion of new job opportunities. Whether this judgment is fully realized requires time to validate. But at least from this dialogue, his bet on the future of AI is very clear: AI is not an abstract futuristic myth but a new industrial system that is on the way.