When we talk about artificial intelligence (AI), most people's intuition might still be limited to intelligent chatbots like ChatGPT or Copilot assistants that can help with coding. However, in the eyes of NVIDIA CEO Jensen Huang, AI has long transcended the realm of simple "software applications" and transformed into an indispensable modern infrastructure, much like electricity and the internet. In a recent paper...In-depth articleIn his book, Jensen Huang likened the operational structure of the AI industry to a "five-layer cake," providing a profound analysis of this industrial revolution that is reshaping the global economic system, from the bottom layer of energy to the top layer of applications.
According to Jensen Huang, the essence of AI has undergone a fundamental change: we are no longer just "retrieving" pre-written code, but enabling computers to learn to "generate" intelligence in real time.
From pre-recorded software to real-time intelligence: A shift in computing paradigms
At the beginning of his article, Huang Renxun pointed out the limitations of computer computing over the past half-century: software is "pre-recorded." Human engineers write algorithms, stuff structured data into databases, and then computers retrieve and execute them according to precise instructions (such as SQL).
But AI has completely disrupted this pattern.
For the first time, modern AI systems possess the ability to understand unstructured information—viewing images, reading articles, listening to sounds, and even inferring context and intent. More importantly, each of its responses is generated "instantly" based on the current context. This paradigm shift from "retrieving pre-stored instructions" to "generating intelligence on demand" forces the entire underlying computational stack to be scrapped and rebuilt.
Deconstructing the "Five-Layer Cake" Architecture of AI
To explain this massive industrial transformation, Huang Renxun proposed a clear five-layer architecture model:
1. Energy: The most fundamental physical limitation
At the very bottom of the pyramid is not any software abstraction layer, but real electricity. Every token generated in real time by the AI model is the result of electron flow and heat energy conversion. Jensen Huang stated that energy is the "first principle" of AI infrastructure, and also the ultimate ceiling that determines how much intelligence a system can generate.
2. Chips: Engines that convert electricity into computing power.
Above energy lies the processor, responsible for efficiently converting electricity into computing power. Facing the massive parallel computing and high-bandwidth memory required for AI workloads, technological breakthroughs at the chip level directly determine the speed of AI expansion and the "cost of intelligence."
3. Infrastructure: Manufacturing Smart Factories
The infrastructure includes land, power transmission, cooling systems, networks, and supercomputers that connect tens of thousands of chips. Jensen Huang emphasized that these systems, known as "AI factories," are no longer designed to store information like traditional data centers, but are specifically designed to "create intelligence."
4. Models: The cognitive layer for understanding the world
AI models operating with powerful hardware support have long since transcended the scope of simple language (LLM). The most transformative advancements to date have occurred in fields such as protein structure prediction, chemical synthesis, physical simulation, and robotics.
5. Applications: The final form of economic value creation.
The top layer contains the various applications we are familiar with. Whether it's platforms for drug development, autonomous vehicles (AI manifested in machines), or even humanoid robots (AI manifested in bodies), all of these rely on the stable support of the four layers below.
The explosion of open source models and infrastructure
Huang Renxun pointed out that in the past year, the capabilities of AI models have crossed the threshold of "large-scale applicability" for the first time (inference enhancement and illusion reduction). This technological breakthrough has directly triggered demand at the application layer in fields such as logistics, customer service, and healthcare.
In the article, Huang specifically named the Chinese open-source model DeepSeek-R1 as an example. When powerful and free inference models become readily available, it not only accelerates software innovation but also drives up the demand for training computing power, infrastructure, chips, and even energy.
To meet this exponentially growing demand, the world is undertaking the largest infrastructure construction project in human history. This construction not only requires programmers, but also urgently needs a large number of electricians, plumbers, steelworkers, and installation technicians, bringing new high-paying opportunities to the traditional labor market.
Analysis of viewpoints
Huang's article is less of a technical analysis and more of an "AI strategy white paper" for government and corporate CEOs around the world.
He cleverly defined NVIDIA's core hardware products (chips and server architecture) as the "wave makers" and "necessities" of this era. When he elevated AI to the same level of infrastructure importance as the national power grid, he was actually conveying a clear message: the future competitiveness of a nation will depend on the scale of its computing power and energy capacity.
This also explains why tech giants including Microsoft and Google have recently begun investing heavily in basic power infrastructure such as nuclear power. Because in the five-layered pie of AI, if the bottom layer, "energy," is cut off, even the smartest models will be ineffective. As Jensen Huang stated, this trillion-dollar infrastructure project has only just begun, and the speed at which we build it and the breadth of our participation will directly determine the final form of this era.
