NVIDIA CEO Jensen Huang: The largest infrastructure project in human history is coming soon

On March 10, Nvidia CEO Jensen Huang, extremely rarely, published a long-form blog post titled 《》，the seventh public long article he has released since 2016.

In the piece, Huang systematically laid out the underlying logic of the AI industry. He noted that AI has a “five-layer architecture,” and that the entire system is still in a very early stage of development. In the future, it will still require investment of tens of trillions of dollars to improve infrastructure—something that will also become “the largest infrastructure buildout in human history.” He said that over the past year, applications built on AI first began creating real economic value, and that energy will become a core element in the process of AI transformation, fundamentally determining the upper limit on the scale of intelligent production.

Below is an edited精编 summary of the content of this long-form article.

In the article, Huang got straight to the point: in the era of traditional computing, software was pre-written, and computers were merely tools for executing instructions. With the advent of AI, this model has been completely broken. For the first time, we have a computer that can understand unstructured information—it can see images, read text, understand sound, and even reason based on context. “Every response is newly created, and every answer depends on the context you provide.”

This capability to generate intelligence in real time requires that the entire computing architecture be redesigned. That also leads to his core definition of the structure of the AI industry— a complete, inseparable “five-layer cake.”

The “five-layer architecture” proposed by Huang consists, from bottom to top, of: energy, chips, infrastructure, models, and applications. This is not just a layered technical stack, but a map of how future value in the tens of trillions of dollars will flow.

Schematic diagram of the five-layer architecture, chart/ NVIDA

First is the energy layer. “Real-time generated intelligence needs real-time electricity.” Huang defines energy as the starting point of AI infrastructure, as well as the physical bottleneck that limits how much intelligence the system can generate—fundamentally determining the upper limit on the scale of intelligent production. Every generated Token is, behind the scenes, a matter of managing the flow of electrons and heat.

This view aligns closely with current industry trends. MIT Technology Review points out that the International Energy Agency predicts that global data center power consumption will double within five years. In the most data-center-dense region in the U.S., Virginia, 26% of electricity already goes to data centers. As quoted in a piece by Titanium Media, Huang’s trillion-dollar AI blueprint is facing a real-world “megawatt” test.

Above energy is the chip layer, designed to convert energy efficiently at massive scale into computing. AI workloads require extreme parallel processing capability, high-bandwidth memory, and fast interconnects. Huang emphasizes that “progress at the chip layer determines the speed at which AI can scale.” This is precisely Nvidia’s core territory, and it is also the fundamental reason behind its ongoing investment in R&D and its product iteration cadence of one year per generation.

Above the chips is the infrastructure layer, including land, power supply, cooling systems, network communications, and the systems engineering that organizes tens of thousands of processors into a single machine. This layer is the reconfiguration of the physical world. Huang calls it an “AI factory.” It is no longer a warehouse for storing information, but a production line for manufacturing intelligence.

Above infrastructure are AI models that can understand language, biology, chemistry, physics, and even finance and medicine—this is also where the public’s intuitive understanding of AI typically lies. Huang specifically noted that language models are only one category; even more transformative work is happening in areas such as protein AI, physics simulation, and robotics.

He especially emphasized the strategic value of open-source models: “When open models reach the frontier level, they change not just software, but also activate demand across the entire technology stack.” Taking DeepSeek-R1 as an example, open, powerful reasoning models will accelerate the adoption of the application layer, which in turn drives demand for chips and energy.

At the top is the application layer, where economic value is created, including drug discovery, industrial robots, legal assistants, and autonomous vehicles. Huang predicts that traditional software and app forms may disappear, replaced by ubiquitous AI agents. He believes that “every successful application will pull every layer beneath it, all the way to the power plant that keeps it running.”

Global developers’ adoption trends for open-source models, chart/Nvidia

The construction progress of the above architecture has already started. Looking across the globe, chip factories, computer assembly plants, and AI factories are rising from the ground at an unprecedented scale—this is becoming the largest infrastructure buildout in human history. In the article, Huang points out that humans have already invested hundreds of billions of dollars, but in the future, we will still need to build infrastructure worth tens of trillions of dollars.

Worth noting is that this transformation doesn’t only require top scientists; it also needs a large number of technical trades. Huang wrote: “AI factories need electricians, pipefitters, steel workers, network technicians… These are highly technical jobs with good pay, and they are currently in short supply.” This means the barriers to participating in the AI revolution are rapidly falling.

“We’re just getting started,” Huang wrote.

Most infrastructure has not yet been built, most labor has not yet been trained, and most opportunities have not yet been discovered. But the direction is already clear: AI is becoming the infrastructure of the modern world.

From the contest over energy to competition for chips—from factory construction to open-source models, and then to application breakthroughs—this nested “five-layer cake” is reshaping the path of global economic growth. As he said at the end of the piece, at this moment, our choices, our construction speed, and our deployment approach will determine where this era goes.