Where AI Value Accrues: Benedict Evans and Jensen Huang's Two Maps

Two talks, side by side

This piece has three parts. First it summarizes what Benedict Evans said on Lenny’s Podcast. Then it summarizes Jensen Huang’s GTC Taipei 2026 keynote. Finally it overlays the two to see where they agree and where they split.

Here is the one-line preview. Both of them treat the model itself as something that becomes a commodity. The difference comes after that. Evans thinks the value moves up, into apps and distribution. Jensen thinks it moves down, into infrastructure and CUDA. Which one is right is genuinely hard to call today.

Evans’s map: the model is a commodity, value moves up

Evans publishes an “AI is eating the world” presentation every six months. The question he holds onto in this session is one. Can foundation model companies have pricing power?

No network effects

His answer is skeptical. The starting point is an observation: the models do not seem to have network effects.

“the models don’t seem to have network effects. So there doesn’t seem to be a winner takes all effect… then why would you have pricing power?” - Benedict Evans

With no network effects, no single company runs away with a winner-take-all lead. Competition keeps going, and since the products are not fundamentally differentiated either, there is no basis to charge for the model itself.

So he treats the model as the base that other products sit on top of. What you build on top decides the product, while the model is the common layer underneath. The same Gemini powers Gemini intelligence on Android and Apple intelligence on iOS, yet the products come out completely different. The model is the same; the decisions and the distribution above it are not.

Value moves up the stack

So where does the value go? The analogy he reaches for is Windows versus the cloud. Does the model dominate the stack from above the way Windows did, or is it a swappable infrastructure like AWS, where you do not care which one you are running?

His conclusion is the cloud side. The chatbot is not the final surface, and thousands of different apps are needed that the model labs cannot all build, just as Microsoft did not build every Windows app.

Why up? His reason is that the hard part of a job is not producing the deliverable. He frames it as the difference between a task and a job. A single job is a bundle of tasks, and what AI automates is one of those tasks. The hard part, deciding what to make, stays above.

Sometimes the task is the job. An old manual elevator attendant had one task as the entire job, working the lever to line up the floor, and when the button automated it the job disappeared with it. But that is the exception.

Most jobs look more like Amazon. What Amazon does is get you the SKU. If you know which microphone you want and know the part number, you can buy it right away. If you do not know which microphone to buy, you cannot start on Amazon. Fetching the SKU got automated, but knowing what you want stays a separate job. Claude can write you the code. But which code, which feature, do you want? Who is the customer, and what is the right product for them? That is the work on the upper layer.

Consulting is the same kind of example. People mention pulling a 75-slide McKinsey deck out of Claude, but the reason you pay McKinsey in the first place is not the deck. It is walking through your company to work out why you had not done it, how the politics run, and what your real customers actually think. The deck is the task, and what the model takes is that task. Knowing what to make stays with people and apps. That is why the value stays on the upper layer.

Distribution is the moat

Once the product is a commodity, what is left is distribution, he argues. The distribution he means here is consumer distribution: brand and default.

He uses the browser wars as the example. Technically a browser is a thin wrapper around a rendering engine. An input box and an output box are most of it, and the last real innovation was tab browsing, more than twenty years ago. The engine can be better or worse, but the product the user sees is largely the same.

In that kind of commodity market, Microsoft won not on technology but on distribution. It bundled Internet Explorer into Windows as the default. When you turned on a new PC it was already there, so most people simply used it. There was no reason to go looking for a better browser.

And what happened? Microsoft held browser share for five or six years and got almost nothing out of it. The value was not in the browser but above it, in the web and the services. Even when you dominate a commodity through distribution, if the value sits higher up, the domination itself does not turn into money.

The same thing is happening with models now, he thinks. Google pushes Gemini into its own surfaces like search and Android, and Meta puts an adequate model into every service by default. For the average user the difference between Gemini and ChatGPT is small, so which one is already installed largely decides the race.

Infrastructure is a low-margin utility

How does he see the infrastructure side? The textbook example of infrastructure is a utility like electricity, and he reaches for the one he has watched for a long time: telecom. Telecom is a utility anyone uses. Lay the network, and everyone makes calls and uses data on top of it.

Over the past fifteen years or so, the volume running on top of it exploded. Mobile data consumption grew roughly 1,500 to 2,000 times since 2010. It is an almost perfect exponential curve. Whole industries run more and more on top of it. The global mobile industry has revenue around one trillion dollars a year, and to hold the network up it pours 15 to 20 percent of revenue into capex every year.

And yet what happened to telecom stocks? For 25 years they went essentially nowhere. While everything running on top changed and exploded, the share prices of the companies that laid the network did not move. Because it is a low-growth, low-margin commodity utility.

You can lay down extraordinary infrastructure and the money still gets made above it. That is the seat Evans sees infrastructure sitting in. And he thinks foundation models, and further down the compute infrastructure that holds them up, may sit in the same seat.

Evans’s chain	Content
Start	The model has no network effects
Therefore	No winner-take-all, ongoing competition, weak differentiation
Conclusion	No basis for pricing power
Where value goes	Up the stack (apps)
Remaining moat	Consumer distribution (brand, default)

Jensen’s map: the agent is the new pattern, value moves down

Jensen opens his keynote with “useful AI has arrived.” Looking at the same model layer, he sees a different place.

The unit of computing becomes the agent

The change he describes is that the unit of computing has shifted. If the old unit was an app running on an operating system, the new one is an agent.

An agent is made of a model, a harness, tools and skills, and a runtime. When input comes in, it observes, reasons, plans, uses tools, and the harness orchestrates the whole flow. Managing short-term and long-term memory is part of it. The model does the thinking, and the harness ties the rest together.

He gives the model away

How does he handle the model? He releases his own model Nemotron 3 Ultra and his physical-AI model Cosmos 3 fully open source, weights, training data, and training scripts. And he highlights that NVL72 produces the cheapest token in the world.

For tools the agents will use, he points to NVIDIA’s roughly 1,000 CUDA-X libraries, now offered as skills, basically manuals, that an agent can read and use. The cheaper and more common the model, the more agents and tools run on top of it.

Compute is revenue

He nails the economics in one line. Compute is revenue, throughput per watt is revenue, and the more you buy the more you make.

“compute is revenue now. Compute is profit… Performance per watt is your revenue. The more you buy, the more you make.” - Jensen Huang

He treats even the trend of a 1GW factory cost rising from 20 to 30 billion dollars to 80 to 100 billion dollars not as cost but as a revenue engine. To run this pattern he puts the next-generation Vera Rubin into full production and introduces Vera CPU, designed for agents rather than people.

CUDA is the moat

The moat he names is the switching cost of the ecosystem. Every data center, cloud, and enterprise has already qualified Grace, and every software stack is already optimized for CUDA. With a thick ecosystem where everyone starts from CUDA, the useful life of the asset is long, which leads to a low total cost of ownership (TCO).

He also flips the assumption that agents kill software companies. With far more agents than people, more tools get used, so it is actually a good time to be a software company. In Jensen’s map, the place that makes money is not the model but the infrastructure below it.

Jensen’s chain	Content
Start	The unit of computing moves from model to agent
Model handling	Nemotron and Cosmos fully open source
Therefore	Token demand explodes, “compute is revenue”
Where value goes	Down the stack (infrastructure, CUDA)
Remaining moat	Developer ecosystem lock-in (CUDA qualification)

Overlay them: where they agree, where they split

Where they agree: the model is a commodity

There is one place the two agree. The model itself becomes a commodity. Evans says it in words, Jensen says it in action. Releasing the whole model open source is the proof.

Giving the model away is not charity. The cheaper the complement, the higher the demand for the core. For Jensen the model is a complement that lifts demand for his hardware, the AI factory that prints tokens. The cheaper and more common the model, the more tokens get printed, and the more tokens get printed, the more GPUs sell. So Evans’s “the model is a commodity” is not something Jensen argues against; it is his sales logic.

But the agreement ends there. Evans’s added claim that there are no network effects is his own further reasoning, not a point Jensen signs onto. Instead of arguing about the model’s network effects, Jensen moves the topic to token demand and infrastructure.

Split 1: up or down

Where the agreement ends, the two pictures diverge. The question is which direction value flows, up or down from the same model.

---
config:
  look: handDrawn
  theme: neutral
---
flowchart TB
    U["Upper layer: apps and distribution (Evans)"]
    M["Middle: the model is a commodity"]
    D["Lower layer: infrastructure (Jensen)"]
    U --- M
    M --- D

What is interesting is that the two reach different conclusions using the same historical analogies.

Analogy	Evans’s reading	Jensen’s reading
Windows vs cloud	The model is an AWS-style commodity, value goes up	CUDA positioned as the Windows of the AI era
Electricity / utility	Infrastructure is low-margin; the money is made by those who make something with the electricity	Compute is revenue, throughput per watt is revenue
Distribution	Consumer brand and default	Developer ecosystem lock-in
Agent / harness	The place where apps differentiate	The place that needs his runtime and Vera CPU

One thing to watch is that the two are talking about different layers. The layer Evans calls a commodity is the model layer, and the layer where Jensen wants the Windows seat is the CUDA and silicon layer below the model. So the model can be an AWS-style commodity while CUDA is a Windows-style lock-in at the same time. They are not colliding over the same object.

Split 2: same word, different moat

The most interesting overlap is distribution. Both say that once the product is a commodity, distribution is the moat. But the distribution they point to is different. Evans’s distribution is consumer brand and default; Jensen’s is the switching cost of the developer ecosystem.

	Evans’s distribution	Jensen’s distribution
What the moat is	Consumer brand / default	Developer ecosystem lock-in
How it works	Sprayed onto every surface, the inertia of the default	The switching cost of qualification and CUDA optimization
Examples	Browser wars, Gemini / Meta	Grace qualification, long asset life

Here is the trap Evans himself admits. If distribution is the moat, that moat favors not the startup building a new app but the big tech players who already hold the distribution. Of Google, Apple, Meta, and Amazon he says it is hard to see a real problem for them in all of this. So the claim that value moves up the stack does not mean value goes to small builders. The value that moved up can get absorbed into another oligopoly.

Split 3: one testable point, infrastructure margins

The analogies above can coexist because they sit on different layers. But there is one point where the two predict opposite things about the same object: infrastructure margins.

The starting point is a line from Sam Altman. He said he would sell intelligence on a meter like electricity or water. Evans does not accept the analogy.

“my dear sweet child you need me to explain the margin structure of the utility industry to you.” - Benedict Evans

When you watch TV, the broadcaster does not pay the power company a cut of your bill, and when you run a washing machine, Bosch does not pay the power company a cut of the machine’s price. Electricity became part of everything, but the money was made by whoever made something with it.

Jensen stands the same electricity analogy on its head. Throughput per watt is revenue, and the more you buy the more you make. There is one small tell here. Jensen put a competitor’s latency-tuned chip, the Groq LPU, inside his own Vera Rubin system. NVL72 prints tokens at the highest throughput, and the Groq LPX rack prints them at the lowest latency. With his own rack, he showed that a single architecture does not get to keep every workload forever.

	Evans’s prediction	Jensen’s prediction
Infrastructure margin	Converges to a low-margin utility	A high-margin revenue engine
Basis	Telco: data up 2,000x, stock flat	Co-design, efficiency, ecosystem
Weak signal	Groq LPU inside his own rack	Current gross margins still around 75 percent
Instrument	NVIDIA margin curve, token price curve	Token demand curve

This point is one that time answers. Where NVIDIA’s gross margin and token prices go over the next two to three years is the single instrument that shows whose map is closer.

Closing: both can be right, and the real question is in the middle

The upper layer and the lower layer can both win. A picture where the model lab gives value up (apps and distribution) and down (infrastructure) while losing pricing power in the middle fits both talks. So these two are less a contradiction than each person looking at the top and bottom halves of the same donut from his own seat.

The real unresolved question is in the middle. What if one of the model labs escapes the middle through memory, personalization, and distribution like ChatGPT’s 900 million users, and becomes a platform on its own? At that moment Evans’s no-network-effects premise breaks, and value gathers back into the model layer. The short-term and long-term memory management Jensen emphasized is exactly that spot. The more the model knows you, the harder it is to switch, and a switching cost appears.

---
config:
  look: handDrawn
  theme: neutral
---
flowchart LR
    Q["The model, now a commodity"] --> U["Value flows up (Evans)"]
    Q --> D["Value flows down (Jensen)"]
    Q --> E["A model lab escapes the middle"]

Evans puts a caveat on his own thesis. This is all deterministic reasoning, and if you had argued about the internet in 1997 or mobile in 2000 this way, you would have gotten most of it wrong.

“we should presume we don’t know.” - Benedict Evans

What you get from putting the two talks side by side is not an answer but a coordinate. Starting from the model as a commodity, Evans points up and Jensen points down. Which arrow runs longer is a question that NVIDIA’s margins, token prices, and the model labs’ distribution fight will answer over the next few years.

Sources

Benedict Evans, “The most rational take on AI you’ll hear this year”, Lenny’s Podcast, 2026.
Jensen Huang, “NVIDIA GTC Taipei 2026 Keynote (Full Replay)”, NVIDIA, 2026.