Reading the AI Era of Work Through Jensen Huang's "5-Layer Cake": How Our Jobs Will Change

Hello, this is Ryuta Hamamoto from TIMEWELL.

When people hear the word AI, the first thing most of them picture is a chat that answers cleverly on the other side of the screen, or an app that builds documents for them automatically. That is what we can touch with our own hands, so it is a natural feeling. But if we think about the AI era of work only through how it looks from there, I get the sense we drop something important.

What set me thinking was a metaphor Jensen Huang, the founder and CEO of NVIDIA, laid out in 2026: "AI is a 5-layer cake." The head of a semiconductor company painted AI not as a story about apps and services, but as essential infrastructure on a par with electricity and the internet. This way of looking at it turns out to be a far better handle than I expected for thinking about how each of our jobs will change from here.

This article is Part 1 of a series on the AI era of work. First I set out what Huang's 5-layer cake is really saying, correcting a common misreading as I go. Then, following the primary sources, I think through the change in work that this layered view reveals and the skills that will matter. If you have any interest in how to bring AI into your own company, it helps to first check where you stand with our free AI literacy self-check. The second half of the piece will feel more concrete once you have.

First, let me clear up a common misreading

This phrase, the 5-layer cake, is in fact starting to come up in Japan's business circles too. But I often see its contents passed on wrongly. The most common version explains the five layers as a software stack: "AI infrastructure, foundation models, tools, agents and apps." A foundation model here means a general-purpose AI base trained on a large volume of data, and an agent means an AI that plans and carries out several tasks on its own in place of a person. It looks like such a tidy hierarchy that you almost want to believe it.

But this is not what Huang is saying. There is no software "tool layer" or "agent layer" in his 5-layer cake. What he drew are the layers of physical and economic infrastructure that we have to build up in the real world in order to produce the intelligence we call AI. From the bottom up: energy, semiconductors, infrastructure, models, and then applications. Those five. What matters is that this is not a list of software features. The story starts from a base you can touch with your hands, such as electricity, factories and buildings.

Why does the mix-up happen? Probably because the only part of AI we ever touch is the top application layer. It is natural to imagine that invisible software tiers continue just above or below it. But I think the reason Huang went to the trouble of comparing it to a cake is that, underneath, there really lies a physical reality of electricity and semiconductors, and if you miss that, you read the whole picture of AI wrong. I want to start by setting this one point straight.

Incidentally, there is some small wobble in how this metaphor gets named. Reports of a January 2026 conversation at Davos introduced the middle layers with slightly different wording, such as "semiconductors and computing infrastructure" and "cloud data centres"¹. In this article I treat the definition in the signed essay Huang himself published on 10 March 2026, "AI Is a 5-Layer Cake"², as the official one. Going to the primary source is how you notice differences in phrasing like this.

Reading Huang's 5-layer cake from the bottom up

So let me go through the five layers one at a time. Because it is a cake, picture it building up from the base at the very bottom.

Layer	Name	The gist in one line
5 (top)	Applications	Drug discovery, robots, self-driving and other places where economic value is created
4	Models	The AI itself, understanding language, science, the physical world and more
3	Infrastructure	The "factory that manufactures intelligence," bundling huge numbers of processors
2	Semiconductors (chips)	Processors that turn energy into computation efficiently
1 (base)	Energy	The electricity that powers everything

The base at the very bottom is energy, that is, electricity. To borrow Huang's words, intelligence generated in real time needs electricity generated in real time. Every single answer AI returns is eating electricity. Each time we casually toss a question into a chat, somewhere a power plant is running. That this obvious fact is placed at the very foundation is telling.

On top of that sits the semiconductor layer, the chips. A semiconductor is a device that turns energy, in the form of electricity, into computation at large scale and with efficiency. The GPU that NVIDIA is known for is a leading example of such a processor. A GPU is a semiconductor originally born for image processing, one that is good at handling large volumes of computation all at once. How little energy you can waste in turning it into computation is where this layer is won or lost.

The third layer, infrastructure, is the one that landed most solidly for me. This is a giant facility that bundles tens of thousands of processors. Huang calls it not a place that stores information but an "AI factory that manufactures intelligence." If a conventional data centre is a warehouse that holds data, an AI factory is a plant that produces intelligence itself, is how he frames it. You secure the land, connect the power supply, run the cooling, put up the building, lay the network. That all of this is still not a software story but real-world construction is the important part.

The fourth layer is finally the model. This is the part many people think of as AI itself. A model tries to understand not only language but a wide range of information, from biology and chemistry to physics, finance and medicine, and even the behaviour of the physical world. And the top layer is applications. Drug discovery, robotics, self-driving and other places where economic value is actually created belong here. What we touch every day is the surface of even this top layer.

The five layers only work if they grow together

The central claim of this metaphor lies not in the number of layers but in the relationship between them. Huang says the five layers are mutually dependent and only work if they expand at the same time. A cake will topple if you thicken just one tier. From the base to the top, it has to grow together while keeping its balance, is the idea.

To borrow his phrasing, a successful app pulls all the layers beneath it. If one app becomes a hit and usage grows, the demand for model computation grows by that much, more infrastructure has to be added, semiconductors run short, and in the end the load reaches all the way down to the power plants that keep it running. What happens up top cascades into the physical layers below. Conversely, if the lower layers are thin, no matter how clever an app you conceive up top, you cannot run it. If you grasp AI as a story about a single breakthrough or a single app, this whole interlock goes invisible, and that, I think, is what troubles Huang.

That is why he described AI as "the largest infrastructure build-out in human history." On scale, too, he offers a view that what has been put in so far is on the order of hundreds of billions of dollars, and that trillions of dollars of infrastructure will be needed from here. That said, this figure is Huang's own estimate, not an objectively confirmed statistic². It is also a forecast from someone in the business of selling semiconductors. Rather than the numbers themselves, the right way to read it is to take in the sense of scale: that AI is not a story that closes inside a few giant tech companies, but one that moves by drawing in the whole physical economy of power, construction and manufacturing.

This picture also overlaps with movements in Japanese policy. On the effort to grow computing infrastructure and models domestically, I have written a piece organising Japan's home-grown AI base and NEDO's support. Of the five layers, how much of which layer Japan holds within its own borders is itself a central question of economic security going forward.

Implication for work, part 1: jobs spread into the physical economy

From here I think about what this layered view means for how we work. What makes Huang's 5-layer cake interesting as a story about work is that it does not draw AI's job creation as something confined to the world of software engineers.

He says building this vast infrastructure will require an enormous amount of labour. And he named specific trades: an AI factory needs electricians, plumbers, pipefitters, steelworkers, network technicians, installers and operators. On top of that, he says you do not need a PhD in computer science to take part in this shift. When we hear "AI boom" we tend to imagine a world where only the most advanced researchers and programmers reap the benefit, but in reality, the closer a layer is to the base, the more it needs a wide range of hands, is the view.

This observation carries a lot for Japan's regional and small-to-mid-sized manufacturing sites too. When the AI era arrives, won't the skills we hold grow old and lose their value? I hear that anxiety often. But by Huang's account, the real base that runs AI is held up by exactly those physical skills. The work of running electricity, keeping cooling turning and installing equipment is needed more, not less, the smarter AI gets. If anything, it is a story about a shortage.

Of course, applying this to Japan as-is calls for some care. What Huang has in mind is countries and regions where large-scale AI factory construction is advancing. How much demand of the same scale arises in Japan depends on future siting and investment, and it cannot be told through optimism alone. Even so, the perspective that jobs in the AI era can spread not only to top-layer software but out to the base of the physical economy is worth holding on to when you rethink the value of your own work.

Implication for work, part 2: less taken away, more shifted to judgment and care

Another point Huang makes repeatedly is that AI augments skilled professional jobs more than it takes them away. The example he reaches for here is the radiologist.

A radiologist is a specialist who reads X-ray and CT images to find signs of illness. Part of the work of reading images is also something AI is good at. So radiologists are sometimes spoken of as a representative job AI will take away. But Huang's view is the reverse. If AI takes on the routine image reading, the radiologist can concentrate on the parts only a person can do: judgment, communication with patients and other doctors, and care. As a result, the productivity of the whole hospital rises, it becomes able to see more patients, and hiring, if anything, goes up, runs the storyline.

This may sound like optimism, but as a mechanism it holds up. When a task becomes efficient and can be done cheaply and quickly, demand for that task can actually increase. If image reading becomes fast and accurate, more people can get examined, and the total volume of healthcare grows. Then the jobs around it that require human judgment and response grow too. Efficiency does not necessarily lead straight to a shrinking of employment, is the line of thinking. That said, this is only one view Huang has put forward, and it is worth holding alongside the fact that there is no guarantee every job will play out this way.

What I think is important is what Huang says beyond this. From here on, he says, what a job is for will be asked more than the fine detail of its tasks. The purpose of a radiologist is not the reading of images itself, but to diagnose patients correctly and care for them. Return to the purpose, and the line between the tasks you can hand to AI and the work you should keep doing yourself starts to become visible. This posture of "thinking from purpose" applies not just to radiologists but to every kind of job, I believe.

Implication for work, part 3: become someone who creates value in the top layer

Let me bring all of this back to your own work. Most of us work in the very top of the 5-layer cake, the application layer. We are not on the side that makes energy or semiconductors, but on the side that uses finished AI to create real value. What is asked here comes down to one thing: how well you can put AI to use.

So what skills pay off in this top layer? I think there are four, broadly. The first is deep expertise in your own industry and shop floor. AI is good at general-purpose answers, but it does not know the particulars of this company, this shop floor. Only the person who knows the site can fill that gap. The second is the judgment to see what AI should be made to do. It is not that you should hand everything to AI; the ability to tell apart where it works and where it does not greatly shapes the result.

The third is the ability to define a problem correctly in words. AI will not return an answer beyond the quality of the question. A vague instruction gets a vague answer back. The more clearly a person can articulate in their own words what they want to solve, the better the result they can draw out of AI. The fourth is the interpersonal and care skills of facing other people. As the radiologist example shows, the part where you finally touch a person's heart will remain human work from here too. None of these four can be replaced by AI wholesale; they are the kind of skills that show their power only when paired with AI.

And underneath all of these, I think, is the mindset of designing what you entrust to AI and what you keep for yourself. The reason we keep running an AI consulting service called WARP is precisely that we feel we are getting somewhere on this design part. Where, and how, do you put AI so that this company's this task actually gets easier? It is work you draw out together with the management. In particular, whether the top of the organisation can use AI as their own tool decides the speed of change. If the person giving the orders is not touching AI, the shop floor does not move. The top becoming AI-native: that, I feel through walking alongside companies day by day, is where everything starts.

One more thing: what AI tries to understand is not limited to words and images. The model layer of Huang's 5-layer cake also includes "world models" that understand the behaviour of the physical world. A world model is a technology in which AI holds within itself the real-world laws of how things collide and fall, so that robots and self-driving cars can move correctly in the real world. On this relationship between the physical world and AI, I would also point you to a piece on physical AI and Japan's strategy.

To sum up

It ran long, so let me organise the key points.

• Jensen Huang's "5-layer cake" is a metaphor that treats AI as essential infrastructure on a par with electricity and the internet, with five layers from the bottom up: energy, semiconductors, infrastructure, models and applications
• The common account that includes a "tool layer" and an "agent layer" is a misreading; what appears in Huang's five layers is a stack of physical and economic infrastructure
• The five layers are mutually dependent and do not work unless they expand together; if the top app grows, demand cascades all the way down to the power plants
• AI-era employment can spread not only to top-layer software but out to the base of the physical economy: electrical work, construction, manufacturing and operations
• Skilled jobs are less taken away than reassigned toward judgment, interpersonal work and care, and the rise in productivity can lead to more hiring, runs Huang's argument
• What pays off for people who create value in the top layer is on-the-ground expertise, the judgment to see where to use AI, the ability to put a problem into words, and the skill of facing other people

If you grasp AI only as a clever app, you cannot see the large ground that spreads beneath it. But the change in how we work is happening on exactly that ground. Start from working out where in your own company AI actually helps. If you are unsure how to think it through for your own case, please talk to our WARP team. Specialists who led DX and data strategy at major companies walk alongside you month by month, helping you bring AI down into your management. Through WARP ENTRE, a project run under an agreement with the Tokyo Metropolitan Government, and our experience training more than 500 people to date, we work with you to translate the large flow into the one step in front of you.

References and primary sources

Footnotes

1. "Larry Fink and Jensen Huang on AI at Davos," official NVIDIA blog (21 January 2026, a conversation at the World Economic Forum Annual Meeting 2026). https://blogs.nvidia.com/blog/davos-wef-blackrock-ceo-larry-fink-jensen-huang/ ↩

2. Jensen Huang, "AI Is a 5-Layer Cake," official NVIDIA blog (10 March 2026). https://blogs.nvidia.com/blog/ai-5-layer-cake ↩ ↩²