AI Is a 5-Layer Cake

Jensen Huang’s thesis is deceptively simple: AI is not a product. It is infrastructure. And like all infrastructure, it has layers, each dependent on the one beneath it.

From Prerecorded to Real-Time

The article opens with a distinction that reframes how we think about computing. Traditional software is “prerecorded”: humans write algorithms, computers execute them deterministically. AI breaks this model entirely. For the first time, computers process unstructured information, images, text, sound, and generate intelligence in real time. Each response is freshly created based on context, not retrieved from a predetermined script.

This is not an incremental upgrade. It is a categorical shift in what computers do.

The Five Layers

Huang presents AI as a stack of five interdependent layers:

Energy sits at the foundation. The key insight here is the symmetry: “intelligence generated in real time requires power generated in real time.” Unlike traditional software that runs on fixed infrastructure, AI’s computational demands scale directly with usage. Every query, every inference, every training run consumes power proportional to its complexity.

Chips transform energy into computation. These are not general-purpose processors but specialized silicon designed for massive parallelism, high-bandwidth memory, and fast interconnects. The architecture of AI chips is fundamentally different from CPUs because the workload is fundamentally different.

Infrastructure is where chips become AI factories. This layer encompasses land, power systems, cooling, networking, and orchestration systems that coordinate thousands of processors working in concert. It is industrial-scale engineering, not a server closet.

Models are the intelligence layer. Huang makes a point worth noting: language models are just one category. AI systems now understand biology, chemistry, physics, and other domains. The model layer is broader than the chatbot discourse suggests.

Applications are where economic value materializes. Drug discovery, robotics, autonomous vehicles, AI-powered professional tools. This is where the infrastructure investment converts into outcomes.

The framework’s power is in its interdependence. Weakness at any layer constrains everything above it. An energy bottleneck limits chip production; a chip shortage constrains infrastructure; insufficient infrastructure limits model training; fewer models mean fewer applications.

The Open-Source Catalyst

One of the more counterintuitive points: open-source models like DeepSeek-R1 reaching frontier performance levels do not undermine the infrastructure business. They accelerate it. When capable models become freely available, they activate demand across the entire stack, more applications need more infrastructure need more chips need more energy. Democratizing the model layer expands the market for every layer beneath it.

The Scale of What’s Coming

Huang frames the current moment as the beginning, not the midpoint, of the largest infrastructure buildout in human history. Hundreds of billions have been invested so far; trillions remain. This requires skilled workers across electrical, plumbing, construction, and technical fields.

On job displacement, he offers an empirical counterpoint: radiology. AI now assists with routine analysis, yet demand for radiologists continues rising. The pattern is productivity amplification rather than replacement, at least in domains where demand is elastic.

Some Thoughts

The five-layer framework is useful less for its novelty than for its clarity. It forces a conversation about AI out of the abstract (“will AI take my job?”) and into the concrete (“where are the bottlenecks in the stack?”).

A few things worth sitting with:

• The energy-intelligence symmetry is the most underappreciated constraint in AI discourse. Real-time intelligence requires real-time power, and that is a physics problem, not a software problem.
• Framing open-source models as demand activators rather than value destroyers inverts the usual competitive logic. NVIDIA benefits when the model layer commoditizes because it drives infrastructure spending.
• The “prerecorded vs. real-time” distinction is a cleaner way to explain the AI paradigm shift than most technical explanations manage.
• The radiology example is doing a lot of heavy lifting on the job displacement question. Whether it generalizes beyond demand-elastic professional services remains an open question.