Jensen Huang Is Building Your Future: The Five-Layer AI Cake and History's Largest Buildout

When Ezra Klein turns his camera to Nvidia's campus in Santa Clara, his opening question is blunt: if some people truly believe they've lost control, why do they keep shipping? In his intro he notes that for weeks the loudest voices on AI have been the frontier labs — the teams behind Claude, ChatGPT and Gemini — yet Jensen Huang carries a different kind of weight. Nvidia is now the world's most valuable company at a $5.4 trillion market cap, and as Klein puts it, since 2023 fifteen cents of every dollar the U.S. stock market earned came from Nvidia alone.

Klein's framing matters here: AI is not famous and therefore Nvidia is famous; modern AI is possible because Nvidia's chips were famous first. The chips were originally built for graphics and video games. The parallel-compute architecture inside them and the way they were programmed turned out to be exactly what deep learning needed. That coincidence became the material foundation for today's model boom.

Huang's quiet influence inside the Trump administration forms the backdrop. Not only because he controls a central resource for training and inference, but because he can read the energy and manufacturing chain beneath it. Klein uses that leverage to contrast the safety-and-regulation calls of lab leaders with Huang's confidence in engineering fixes.

The five-layer cake: from energy to application

Huang describes AI not as a single model or app but as a five-layer cake. At the bottom is energy. Because intelligence is generated in real time, energy must be converted in real time; every token is electrons moving, heat being managed, power turned into compute. Above energy sits the layer Huang lives in: chips and computing infrastructure. Above that is the AI factory — the cloud and data-center infrastructure that processes, cools and distributes.

The fourth layer is models. Here Huang insists on breadth: not only language models but chemical, biological, physical, articulation, robotics, navigation and self-driving models. At the very top, where economic value is created, are applications. Legal copilots, health documentation, drug discovery, logistics, manufacturing, retail search — even a self-driving car or a humanoid robot is the same stack embodied in a different body.

The logic of the cake is pull-through: every successful application pulls on every layer beneath it, all the way down to the power plant. So when Huang sees an application rise, he does not see just software success but the entire chain being pulled from turbine to cable to silicon to cooling.

Why the largest buildout in history is happening now

For Huang this is not a product cycle but an industrial revolution and we are still early. A few hundred billion dollars of it has been built; trillions remain. Chip fabs, computer assembly plants and AI factories are rising simultaneously around the world. In Klein's relay, this is the largest infrastructure buildout in human history.

Numbers make it tangible. TSMC has disclosed plans for twenty new plants, and manufacturers like Foxconn and Wistron are building large-scale systems with Nvidia. Systems like Vera Rubin rank among the most complex computing architectures ever built. Each layer brings its own bottlenecks, vendors and permitting problems, and each layer writes its own bill.

What broke through last year was the model layer. Hallucinations fell, reasoning improved, grounding jumped. For the first time applications became useful at scale and started generating real economic value. Drug discovery, logistics, customer service, software development and manufacturing showed product-market fit. As applications pulled, models, infrastructure, chips and energy were pulled with them.

Open source is an accelerant here. Huang notes most of the world's models are free, and when a frontier open model is released it does not just change software — it lights demand across the whole stack. He cites DeepSeek-R1: making a strong reasoning model widely available accelerated adoption at the application layer and increased demand for training, infrastructure, chips and energy underneath.

Jobs, nurses, and tradecraft

The buildout is not only white-collar. Huang lists what AI factories need: electricians, plumbers, pipefitters, steelworkers, network technicians, installers and operators. These are skilled, well-paid tradecraft jobs and they are in short supply. You do not need a PhD to participate.

Health care is the sharpest example. With a shortage of roughly five million nurses in the U.S., nurses spend nearly half their time on charting and documentation. Huang argues that AI taking over charting and documentation lifts productivity, and paradoxically hospitals want to hire more nurses. Productivity expands capacity, so employment expands with it.

The Davos counterpart to this story played out with BlackRock CEO Larry Fink. Huang called AI infrastructure like electricity and roads. Every country, he said, should build its own AI on top of its own language and culture and keep that national intelligence inside its ecosystem. Accessibility underwrites the claim: no software in history reached nearly a billion people in two to three years because no software was this easy to use.

Safety as engineering, not panic

On safety Huang diverges in tone from the labs. He does not dismiss worry but frames the fix as engineering, not legislation. Being alarmed is not by itself a public service, he suggests; if you think you've lost control you should not ship until you regain it. The line is not stop, but fix and test.

The question lingers: how long does the engineering fix take and who audits it? Huang reads broad open-source diffusion as speed; the same diffusion enlarges misuse risk. Klein's questions expose that tension, and Huang answers that safety is not a one-time patch but continuous quality control at every layer of the stack.

Bubble or sensible build?

The market story shadows the cake. Nvidia's $5.4 trillion value and its outsized contribution since 2023 make a bubble debate unavoidable. Huang's answer again comes from the stack: because every layer reinforces the others, the build is sensible. As long as applications create value, they create demand downward to energy, which decouples the build from a single sector's enthusiasm.

Risks still stack. An energy constraint, a chip-supply shock, a cloud-capacity crunch, a model-quality stall or a slow application uptake happening at once would be painful. Huang implies the remedy in each case is more infrastructure and better models, but in Klein's show that optimism is not yet sufficient assurance for those who want regulation.

By the end the picture sharpens: Huang is not building your future as a model race but as a country-scale grid. Power plant, fab, cloud, model and application must rise together or intelligence cannot be produced. That is why his refrain — every company will use AI, every nation will build it — sounds less like a vision and more like a construction schedule.