Written by: Techub News Compilation

In a lengthy conversation with Stanford students, Nvidia founder and CEO Jensen Huang made a very clear core judgment: computer science is not undergoing a localized upgrade, but is experiencing the most profound restructuring since the IBM System/360. He believes that over the past sixty years, people's understanding of computers, software development, company organization, chip design, system building, and application deployment have undergone several waves, including PCs, the Internet, mobile computing, and cloud computing, but the underlying computing paradigm has largely remained unchanged; with the emergence of generative AI, reasoning models, and intelligent systems, this stable structure has been disrupted for the first time.

Jensen Huang first emphasized that today is an excellent time to learn computer science, not only because AI is hot, but because "computation" itself is being redefined. In his view, traditional computation resembles a "pre-recorded" world: software is written in advance, content is stored in advance, and computers play more of a role in retrieval, execution, and display; whereas in the age of AI, computation is increasingly shifting towards "real-time generation," where systems continuously generate answers, content, actions, and even decisions based on context, user intent, and environmental changes. This means that computers are no longer just passive tools executing commands but are beginning to become active systems capable of processing semantics, understanding context, invoking tools, and operating continuously.

This change is not a point upgrade in technology, but a full-stack reconstruction. Jensen Huang mentioned that changes have occurred at almost every layer, from software development methods, programming tools, organizational structures, to the operation of neural networks and traditional compilers, and then to networks, storage, cloud services, and final application forms. He particularly pointed out that deep learning and generative AI not only have made tasks like image generation, summarization, and translation feasible but, more importantly, have clearly shown for the first time that machines can not only generate content but also have the potential to form "thought processes" and further evolve towards intelligent systems with step-by-step reasoning and tool invocation capabilities.

From "Layered Optimization" to "Full-Stack Collaboration"

When discussing "co-design," Jensen Huang provided a very representative system perspective. The past computing industry has been accustomed to dividing problems: those who create microprocessors design processors, those who work on compilers study compilers, those involved in programming languages research languages, collaborating through abstract boundaries. This division of labor was very effective during the era of general computing, but when workloads turn into highly computationally dense problems such as deep learning, autonomous driving, quantum chemistry, and fluid mechanics, it becomes increasingly difficult to achieve orders of magnitude performance leaps by optimizing a single layer alone.

He illustrated this with Stanford's tradition in RISC: when processor architecture and compilers are considered "cooperatively," simplifying instruction sets and enhancing compilability can be more effective than achieving local optima independently. This approach has been pushed to the extreme in the AI era. Jensen Huang believes that Nvidia's true difference lies not in just making GPUs but in jointly optimizing algorithms, frameworks, compilers, chips, CPUs, GPUs, networks, switches, storage, and even data center-level systems. In his words, this is a form of "extreme co-design," meaning not improving a single component, but rewriting the entire machine around the target task.

The numbers he provided for comparison are also striking. According to the traditional imagination of Moore's Law, achieving tenfold improvement in five years and a hundredfold improvement in ten years is already quite remarkable; however, from his perspective, if relying solely on the natural iteration of semiconductors without touching the software stack and system architecture, the actual gains over the past decade could even be limited to between tenfold and a hundredfold. With the help of full-stack collaborative design, Nvidia has achieved generational improvements of tens of thousands to millions of times in certain AI computing scenarios. It is this superlinear system progress that has transformed the training method of "feeding the entire Internet to the model" from an unrealistic fantasy into an engineering reality.

In the AI Era, Why Computers Are No Longer Just Extensions of Old Machines

Jensen Huang repeatedly reminded students not to view AI as just a new layer of software added on top of old computers, as it is changing "what a computer is." In the traditional cloud computing era, computing was mostly demand-triggered: users initiate requests, systems process tasks, and resources are reclaimed after computation ends, which is a typical on-demand model. However, systems in the agent era begin to exhibit characteristics of continuous operation, continuous observation, continuous planning, and continuous tool invocation, where computing resources are no longer allocated on a per-instance basis but resemble a long-term online "digital workforce."

Once this happens, many old assumptions need to be rewritten. What is a personal computer? What is cloud service? How do CPUs and GPUs divide labor? How do storage and networks connect? Is application software "one program" or "a system composed of multiple agents"? These questions no longer have answers based solely on the old era's experiences. Jensen Huang describes future intelligent agent systems as having long and short-term memory, requiring frequent tool use, needing extremely low latency CPU responses, and needing storage to connect more directly to processors and interconnected networks. With this thinking, the design focus of computing platforms shifts gradually from traditional servers to "machines born for inference, for agents, for continuous work."

He took Nvidia’s generational platform evolution as an example to illustrate this design logic aimed at future workloads. Hopper is more focused on pre-training, addressing the needs of building large-scale training clusters; Grace Blackwell significantly targets inference and token generation, aggregating memory bandwidth of multiple chips through larger interconnects like NVLink 72 to satisfy the bandwidth demands during the decoding phase of large models; and in Huang's narrative, Vera Rubin is set to optimize more clearly toward the agent era, addressing long memory, tool invocation, low-latency CPU collaboration, and direct storage connections. His judgment is clear: pre-training is not the endpoint, inference is not the endpoint, the real goal is to "do work," which means moving AI from simply answering questions to executing tasks over the long term.

Measuring Computing Power Cannot Just Focus on FLOPS

When discussing open ecosystems, computing power utilization, and performance metrics, Jensen Huang presented a very intriguing perspective: focusing solely on MFU or FLOPS is likely optimizing the wrong target. He explained that large-scale AI systems are not limited by a single bottleneck machine; they can be constrained at any moment by various factors including computation, video memory bandwidth, video memory capacity, network bandwidth, and storage access. If a system must over-configure in multiple dimensions to handle peak workloads, then having certain FLOPS idle during many time periods does not indicate failure in system design, but could rather be the necessary cost to avoid the constraints imposed by Amdahl's law.

More importantly, truly meaningful metrics should return to task results. For generative AI, "tokens per watt" is closer to actual value than simple FLOPS; however, even tokens should be assessed according to what kind of tokens are generated, in what tasks, and whether they ultimately convert into genuine intelligent output and business results. Jensen Huang's point is not overly complicated: the most common error engineering teams make is to latch onto a conveniently quantifiable metric and align all organizational efforts towards it, ultimately achieving impressive numbers but failing to deliver better products, stronger intelligence, or higher user value.

This still reflects a system thinking approach. Different laboratories, customers, and applications have their own assessment standards; if the underlying platform is overly customized for one type of task, it will lose scalability in other scenarios; but if absolute generality is pursued, it will devolve into a "machine that can do anything but excels at nothing." Jensen Huang refers to this balance as an almost "artistic" job: capturing a sufficiently large general market to support R&D investments while providing performance density that far exceeds general platforms in critical scenarios.

The Significance of Open Models Is More Than Just "Free"

Regarding the open-source versus closed-source debate, Jensen Huang's answer is not extreme. He explicitly acknowledges that the current most powerful frontier model products are very useful, and Nvidia internally uses models like those from Anthropic and OpenAI, with company engineers often working with the assistance of AI agents. From a pragmatic perspective, he does not romanticize "openness," instead recognizing that closed-source frontier models are often more mature in terms of productization, toolchains, and overall experience at this stage.

However, he also emphasizes that open models remain extremely important for at least three reasons. First, language models are a form of encoding human knowledge and intelligent structure and cannot be monopolized by a few companies regarding their foundational capabilities. Second, many languages, regions, and specialized fields will not automatically become priority investments for large companies, so there is a need for cutting-edge open models as a foundation, allowing different countries, communities, and disciplines to perform localized fine-tuning and secondary development on them. Third, and crucially, there is the issue of safety: if we wish AI to be safe, verifiable, and defensible, then the system must be as transparent as possible because a completely black-boxed capability system is difficult to audit and build a reliable defense system around.

He further extends "openness" beyond language. Jensen Huang mentioned that Nvidia is promoting not just general language models, but also foundational models in multiple fields such as biology, autonomous driving, robotics, and climate science, since the data structures, representation methods, and training methods in these areas differ completely from Internet text, making current general language models insufficient to naturally cover them. In his narrative, the true mission of open models is not to create an alternative downloadable from GitHub, but to establish the first infrastructure for scientific and industrial fields that have not yet been fully marketized but are highly valuable.

Education Is Being Restructured by AI

If the computer system layer is being reconstructed, then the educational layer is also undergoing structural changes. Jensen Huang candidly stated that AI should not only enter the curriculum as a "learning object," but should become a "learning tool" embedded throughout the entire educational process. The traditional textbook writing cycle is lengthy, updates are slow, and it's hard to keep pace with the rapid knowledge changes in the AI field, which may occur on a weekly or even daily basis; thus, future education is likely to take the form of a combination of "first-principles textbooks" and "AI real-time research assistants."

His statement is very down-to-earth: many people today still see AI as a "tool to summarize documents," but in reality, AI is increasingly becoming like an ever-available research partner in reading, synthesizing, relating to other papers, answering questions, and simulating research discussions. This means that students' learning methods will shift from "absorbing linearly along textbook lines" to "dynamically exploring around problems," and the role of teachers will also transition from merely imparting knowledge to helping students develop problem awareness, judgment frameworks, and academic taste.

However, Jensen Huang does not deny the value of classic textbooks for that reason. On the contrary, he particularly defends the first principles, believing that foundational methodologies will not become obsolete due to the AI wave. He recalled his time at Stanford, where he received systematic theoretical training while practicing chip design in the industry; it is precisely this "real-world problem + foundational theory" dual pathway that allows one to truly understand technological changes more easily. According to his view, the ideal educational role of AI is not to replace basic education, but to further connect foundational principles with contemporary real problems.

Computing Power Is Not Only Explained by "Shortage"

Regarding the "computing power shortage" commonly felt by universities and research institutions, Jensen Huang provided a controversial yet insightful judgment: the issue may not be that chips do not exist, but that institutions themselves have not established budgets and governance methods to acquire and organize large-scale computing resources. Discussing Stanford, he noted that funding within the university is typically dispersed and independent by departments, research groups, and grants; this mechanism is suitable for small-scale laboratory research but not for building a large AI supercomputing platform that requires inter-disciplinary sharing and peak demand allocation.

His point is not to deny the supply-demand tension but to emphasize that the organizational model itself has become outdated. Over the past few decades, many universities and research institutions have shifted from centralized computing to a decentralized model of "one computer per person," which has been sufficient for traditional research but fails to meet the needs of cutting-edge research relying on centralized computing pools in today's environment that requires large model training, simulation, bio-computation, and multi-agent experimentation. Therefore, he advocates that universities should rebuild campus-level shared supercomputers as they did in the past for large scientific installations, viewing computing power as a new foundational research facility.

The practical implications of this judgment are strong. While the AI competition may seem to be a battle of model capabilities, it is essentially also a competition of organizational capabilities: whoever can redesign their budget, infrastructure, software stack, and talent utilization methods is more likely to unleash innovation first. For universities, this is not just a matter of buying more GPUs, but whether they are willing to shift from a "decentralized project structure" to a "shared platform structure."

Energy Constraints Will Become the Main Battlefield in the Next Stage

Jensen Huang further expanded the discussion to a more macro level: if future computing is about continuous generation and continuous operation, then energy will become one of the hardest foundational constraints of the entire AI era. He predicts that the total energy required for future computing may exceed the current levels by several orders of magnitude, even if the exact multiples will be adjusted, this direction will not change. Under this premise, what chip and system manufacturers can first do is continuously improve energy efficiency, that is, to increase "effective intelligent output per watt" through architecture, interconnect, and co-design.

However, merely improving the efficiency of individual machines is far from sufficient. Jensen Huang believes that the demand for computing power and energy in the market has become so strong that it is pushing for a new round of grid upgrades, sustainable energy investments, and larger-scale infrastructure construction. In his view, many past sustainable energy projects highly relied on subsidies, but now, AI's real payment capability for electricity is providing new economic drivers for these projects. In other words, AI is not just an energy-consuming industry; it might also force the entire energy system to modernize at an accelerated pace.

What Is Learned from Failure Is Not Just Technology

When discussing personal and company mistakes, Jensen Huang did not shy away from early errors. He acknowledged that Nvidia's first-generation graphics architecture "almost got everything wrong" in many technical choices, but it was precisely in such near-destructive setbacks that the company learned how to make strategically correct shifts on top of technical mistakes. He emphasized that while technical choices are undoubtedly important, whether a company can truly survive often depends on how well it understands the market, allocates resources, and retains enough strategic maneuverability amid uncertainty.

He also cited the example of entering the mobile market, which he later defined as a "real strategic mistake": the company had invested a significant amount of resources into mobile devices, achieving a billion-dollar business at one point, but was locked down during the transition from 3G to 4G, ultimately leading to this line's near-zero status. This experience deepened his understanding that not every opportunity should be chased; entering a seemingly massive market, where the core value chain is not controlled, might just consume strategic resources. However, the low-power and high-efficiency design experiences accumulated during that investment were later redirected into new fields like robotics, becoming the starting point for another technical lineage.

This retrospective approach reflects Jensen Huang's management philosophy: failure does not automatically translate into value, but if it can be distilled into a deeper understanding of strategic boundaries, resource constraints, and future directions, it becomes part of the company's long-term competitiveness.

How to Judge the Future: First Observe, Then Return to First Principles

When discussing "how to make judgments when the shape of the future is still unclear," Jensen Huang's answer is consistent: first spot anomalous signals in reality, then break it down to first principles. He took AlexNet and the breakthrough in deep learning as examples, arguing that what is truly important is not "seeing a new model," but realizing that this model broke through the ceiling of the computer vision trajectory that had been established for many years in a single leap; this magnitude itself is evidence of a "paradigm shift."

What needs to be asked next is not how good it is today, but what old problems it will reopen, what old systems it will force to become ineffective, and what kind of new computing platforms it will require. By continuously deriving along this logic, companies can gradually establish a mental model of future workloads, future market scales, and future system forms, and make large directional bets even when things are not completely certain.

Jensen Huang also acknowledged that this deduction cannot always be accurate. In the real world, the correct strategy is often not about "guessing the future one hundred percent," but rather retaining enough room for adjustment while being roughly correct in direction with lower opportunity costs, consistently making corrections in action. This is why he repeatedly emphasizes the importance of "optionalities": in an uncertain era, it is more important to maintain the ability to continue moving in the right direction than to perfectly lock in on all details at once.

What This Conversation Truly Conveyed

If this lengthy conversation could be condensed into one sentence, what Jensen Huang truly wants to convey is not "why Nvidia succeeded" but rather "engineers in the AI era must relearn how to think about systems." In his view, the most valuable individuals in the future are not the ones who can only write a bit of model-calling code, nor those who focus solely on extreme optimization of specific metrics, but those who can understand the interplay between workloads, system bottlenecks, organizational structures, energy constraints, and industrial structures.

This is also why he repeatedly encourages students: entering computer science now is because you are perfectly positioned at the starting point of a foundational paradigm shift. When the entire computing stack is being rewritten, young engineers are not just those who can adapt to changes, but they have the potential to redefine machines.

For the average reader, this conversation also provides an important insight: AI competition has never been just a string of scores on a model leaderboard; it is simultaneously a comprehensive competition about system design, infrastructure, open ecosystems, educational organization, and energy supply. Those who can connect these seemingly disparate issues are the ones truly close to the entrance of the next era.