NVIDIA founder and CEO Jensen Huang has laid out a “five-layer cake” framework for AI in a post published on the company’s official blog, arguing that the full stack runs from energy to chips, infrastructure, models, and applications. In his view, AI is not just software with better output. It is becoming a form of core infrastructure, comparable to electricity or the internet.
Huang says every company will use AI and every country will build AI. His essay is framed around a basic question that has drawn growing attention: what AI actually is, why it matters, and how to understand the system as a whole.
From pre-recorded software to real-time intelligence
He begins by separating AI from traditional computing. Conventional software follows rules written in advance by humans, with algorithms and instructions fixed during development. AI breaks from that pattern. It can process unstructured information such as images, text, and sound, then generate outputs in real time based on context.
Huang describes this as the production of intelligence on demand. Each response is newly generated rather than selected from prewritten logic. He argues that the scale of this shift is large enough to be compared with the Industrial Revolution.
Why the lower layers set the ceiling for everything above
The bottom layer in Huang’s model is energy. He calls it the most overlooked part of the stack and the most basic constraint on how much intelligence an AI system can produce. Every token requires electron flow, thermal management, and energy conversion. If energy supply is limited, the upper layers are capped as well. That point also explains why he has publicly described nuclear power as a reasonable option in the AI era.
The second layer is chips, which convert energy into computing power. Huang notes that AI workloads rely on massive parallel processing, high-bandwidth memory, and fast interconnects, making them structurally different from traditional CPU-centered computing. Progress at the chip layer shapes both the pace of AI expansion and the cost of each unit of intelligence.
The third layer is infrastructure, which he describes as an “AI factory.” This includes land, electricity delivery, cooling systems, construction, networking, and management systems capable of coordinating tens of thousands of processors at once. Huang draws a clear distinction here: an AI factory is built to manufacture intelligence, not to store information like a traditional data center.
Models and applications are now moving into wider use
The fourth layer is models. Huang presents them as the carriers of AI capability across language, biology, chemistry, physics, finance, healthcare, and the physical world. He singles out protein AI, chemistry AI, physical simulation, robotics, and autonomous systems as areas with strong transformative potential.
He also points to the role of open-source models, citing DeepSeek-R1 as an example. Wider circulation of open models, he says, speeds up adoption at the application layer while also increasing overall demand for training compute, infrastructure, chips, and energy.
The fifth layer is applications, where economic value is realized. Huang lists drug discovery platforms, industrial robots, legal assistants, and self-driving systems as examples of how AI capability becomes embedded in products, machines, and task-specific workflows.
On timing, Huang’s position is clear: the industry is still very early. He says model capability has only reached the point of large-scale usability over the past year, with better reasoning, fewer hallucinations, and stronger deployment potential. Sectors including drug discovery, logistics, customer service, software development, and manufacturing are starting to show real product-market fit, but much of the infrastructure has yet to be built and a large share of the workforce has not been trained.