A few days ago I watched Jensen Huang give his keynote in Taipei, on a stream like most of us, expecting a hardware launch. For the most part that is what it was. New chips, new racks, the supply chain thanked by name, two hours of it. The coverage afterwards agreed. NVIDIA had a huge day, the chips are extraordinary, the share price can look after itself.
Two of the things he did in that keynote looked like opposites, and were in fact the same move. He gave his newest models away. The weights, the methods used to train them, even the data behind them, released for anyone to run for nothing. And then he held up a chip for the home computer, built with MediaTek, and said it would take thirty three years to build.
Thirty three years is not how long the chip takes to make. It is the age of his company. The chip is worth what it is worth, he was saying, because everything already runs on it. Three decades of software, the CUDA layer underneath most of modern computing, every library and application NVIDIA has built since it began, physics and biology and genomics and graphics and now AI, all of it sitting on that one piece of silicon. You could copy the chip and you would have almost nothing, because the value was never the chip on its own. It is the thirty three years of work poured into it and onto it that no one else has, and cannot simply build in a hurry.
Put the two moves side by side and the message is hard to miss. The intelligence, he gives away. The thing it runs on, he sells you. He wants that chip in your house and on your desk, as many of them as possible, because the chip and the thirty three years packed into it are the part he can charge for. The model, the part most people are fixated on, he is content to hand out for nothing. That is not a contradiction. It is a map. The man at the centre of this whole industry is showing you, as plainly as he can, which layer he believes the value sits in, and it is not the model.
It is worth following his reasoning, because no one is better placed to read this than he is, and he has no reason to dress it up.
Why give the models away
The picks and shovels reading is true as far as it goes. NVIDIA sells the tools every digger needs, and the digging is not slowing down. But it does not explain the giveaway. Why hand out models you could have charged for, while guarding the chips and the software beneath them so closely?
The answer is one of the oldest moves in business. You make the thing next to your thing cheap. A company that earns its money from one layer wants every layer beside it abundant, because cheap neighbours drive demand for the part it sells. Cheap models mean more people building more agents. More agents mean more of the small units of computation called tokens. More tokens mean more demand for the chips that make them.
He said the line that holds it together more than once. Compute is revenue.
There is an old pattern here. Cheaper steam engines did not cut Britain’s coal use, they raised it, because cheaper power was worth using for more things. Make anything cheaper and we use more of it. Make intelligence cheaper and the world finds far more to do with it. NVIDIA does not need models to be valuable. It needs them everywhere. Giving them away is the fastest way there.
The smiling curve
To see what this does to everyone above him, it helps to borrow a map that was drawn, as it happens, in Taiwan.
In 1992 Stan Shih, the founder of Acer, described the smiling curve. Looking at the personal computer business, he saw that value did not spread evenly along the line that turns an idea into a product in someone’s hands. It gathered at the two ends. At one end, the design. At the other, the brand and the customer. In the middle, where the assembly happened, value was thin and getting thinner, because anyone could assemble a PC. Drawn on a graph, high at both ends and low in the centre, the line looked like a smile. Shih acted on it, steering Acer out of the squeezed middle and toward the ends.
The curve keeps coming back. It shows up wherever a kind of work becomes something anyone can do, because the moment everyone can do a thing, no one can charge much for it. The value moves to whatever stays scarce.
The same curve is forming again in AI, and Jensen drew it for us without naming it. The model is the assembly line of this industry, and it is becoming standardised, abundant, and free, exactly as PC manufacturing did.
One slide made the point in a single image. It showed the personal computer he is building with Microsoft, and the row of models you could run on it was eight names long, every one open: DeepSeek, Gemma, GPT-OSS, Kimi, MiniMax, Nemotron, Qwen, and more, lined up like interchangeable parts in a catalogue. Even the biggest labs appeared only in their free form. On the machine meant for your home, the model had become a slot you fill, not a thing you choose. And the loop drawn above that row, the part doing the actual work, started with one word at the top. Context.
He sits at one end of the smile, the chips, the scarce physical layer, and he is working hard to make the middle cheap. The question is what holds the other end this time, and whether you are anywhere near it.
The middle keeps rising
That is where the keynote stopped. The part it left out is the one that has been on my mind for years. The middle of the smile is not a fixed place. It is a line that moves, and it moves upward.
It moves because the models keep getting better and cheaper faster than almost anything in the history of technology. Epoch AI, which tracks this closely, found that the price of reaching a given level of model performance has fallen between nine and nine hundred times a year depending on the task, with the pace picking up since the start of 2024. Enterprise data from Ramp put the cost of a million tokens dropping from about ten dollars to two and a half in a single year. The capability that was the frontier in 2022 now costs a tiny fraction of that, by some measures close to a thousandth.
Put that next to the curve and the conclusion is plain. Every year the models improve and the price falls, and the band of work that anyone can do for almost nothing rises. Work that sits comfortably in the skilled, well paid middle today gets pulled down into the cheap centre tomorrow. Building a website. Writing decent code. A first draft of a brand, a report, a contract, an app. A great deal of what is sold today as professional services, creative work, and consulting is being pulled toward the free tier, one task at a time. And if you are honest about your own week, some of it is probably yours. Not because the work stops mattering, but because the ability to do it stops being rare.
I have been making this argument for a while, and the keynote is the clearest proof of it I have seen, made by the one person who would profit from keeping models scarce and is instead flooding the world with them. The middle is not so much being squeezed as raised, like water, and most people are standing in it without noticing the level climb.
Where the money settles
So where does the value go. To the two ends of the smile, for different reasons, and neither is a comfortable place to hide.
The chip end holds because of physics and money. Chips, fabrication, and power are scarce, slow to build, and hugely expensive, so they earn rent. That is the end NVIDIA owns, and the keynote was a long argument for why it means to keep owning it. But even this end is stranger than it looks, because the companies buying all that compute are not obviously getting rich. Gartner put enterprise spending on generative AI above thirty seven billion dollars in 2025, more than three times the year before, while the price of the tokens kept falling. The cost of intelligence is dropping and the total bill is climbing at the same time. And the model companies, for all their revenue, mostly do not make a profit. The company behind ChatGPT is reported to have lost more than it earned in 2025, on billions in revenue. The money settles in the chips, not in the layer that rents them.
What the model was never shown
The other end is the one to care about, and it holds for a different reason than scarce materials. It holds because of what better models do to the things they cannot make.
A model only knows what it was trained on, and it was never trained on you. The way your business actually works, the data sitting in your systems, the customers you have known for years, the hard facts of your particular corner of the world, were never in the data it learned from and never will be. A stronger model does not lower the value of any of that. It raises it, because the more capable the intelligence you can point at something, the more that something is worth, so long as it is yours and not everyone’s.
Picture a firm that has spent thirty years recording why its customers stayed and why they left. No free model knows that, because it was never in the training data and never will be. Point a free model at it and the thing becomes worth more than it ever was, not less. The intelligence got cheaper. The knowledge only that firm holds did not.
There is an honest difficulty here, and it should stop anyone getting comfortable. This only helps if what you hold is genuinely scarce. A strong model that everyone can run levels the field on capability and, at the same time, hands the advantage to whoever already holds the most and best private knowledge. That is usually the big incumbent, not the small challenger. And the rising middle does not stop politely at the edge of what feels like yours. Some of what companies call proprietary is just work that has not been commoditised yet, and the level is still rising. The defensible version is narrower than it sounds. It is the data, the relationships, and the right to act that are yours and no one else’s.
Renting, or owning
There is one more shift in the keynote, and it changes where the thing only you hold can live.
Almost everything companies do with AI today, they rent. The model sits in someone else’s data centre, your words travel to it down a wire, and a meter runs. Every question has a price, and the price governs how you use it. You ration it. You do not leave an agent thinking all night, because the meter does not sleep. The cost of a single thought is always above zero, and that one fact limits what people even think to use it for.
The machines Jensen showed point the other way. That chip he held up, the one that took thirty three years to build, is called RTX Spark, and he built it with Microsoft as a Windows machine for the age of agents. The specifications are not those of a toy. One petaflop of AI performance, which is a thousand trillion operations a second, on a chip small enough for a thin laptop. A Blackwell graphics processor with 6,144 cores beside a 20 core processor, the two sharing up to 128 gigabytes of unified memory at around 600 gigabytes a second. In plain terms, that is enough to run a model with 120 billion parameters, reading a million words of context at a time, entirely on the machine in front of you, with nothing leaving the building. A few years ago that was a rack in a data centre. NVIDIA has now put a version of it on a desk, and a larger one, the DGX Station, beside it for anyone who wants to run far bigger models at home.
The point underneath the numbers is the one that matters. The capability is no longer the scarce thing. The scarce thing is what you point it at.
He paired the hardware with an agent that, in his words, runs around the clock with no meter anxiety. He showed agents built on the same open tools that hobbyists already use, working without pause on a local machine, against local files, with nothing leaving the device.
That is not a discount. It is a different game. When the model runs on hardware you have already paid for, the cost of a thought falls to almost nothing, and almost nothing is not just a smaller number. It lets you do things a meter would never let you do. An agent that watches your systems day and night. A model that thinks for an hour because no clock is running. Intelligence that is simply there, in the background, rather than something you summon and pay for and switch off. It is the coal pattern again, taken to its end. Make a thing not just cheap but free to use, and people find uses for it they never bothered to count, because they never had to.
This runs straight back to the thing only you hold. If your real advantage is knowledge no one else has, then pushing it through someone else’s metered cloud is a strange way to guard it. It travels down a wire, under another company’s terms and prices, and in some deals it trains the very models your rivals will rent next. Running the model locally closes that gap. You can point a capable model at what is yours without any of it ever leaving the building. Keeping it under your own roof is not a small technical preference. It is what makes the whole strategy safe to follow.
Notice who sells you the machine, though. The home supercomputer is the chip end of the smile in miniature, the same scarce, high margin, physical layer shrunk to fit on your desk. Whether the intelligence runs in a distant factory or your spare room, the company that makes the silicon owns that end of the curve. He is not choosing between the data centre and the device. He is selling the high ground in two sizes.
Be straight about the limits. Meter free is not free. You bought the machine and you pay for the power, so this is rent swapped for ownership, not the end of cost. The real frontier still lives in the big factories, and the heaviest work stays there for now. What ends is the cloud’s monopoly on capable intelligence. The likely shape is not local instead of cloud but both, the private, always on work tied to your own data running on machines you own, the rare heavy lift reaching up to the data centre when it has to.
And there is a bigger turn the keynote leaned on harder than any other, big enough to deserve its own piece rather than a paragraph here. If an agent on your own machine, running a free model against your own data, can build the thing you need on the spot, then the fixed application starts to disappear. The software you used to open, operate, and pay a subscription for becomes something your agent simply makes when you ask. The rising middle does not stop at the model. It climbs into the software built on top of it, and computing itself begins to reorganise around saying what you want rather than operating tools. That is not a side note. It is the next chapter.
The one question to ask
All of which turns the curve into a question about you, not about the industry.
And there is a simple way to ask it. Run this test on anything you do. Picture the model as free, and everyone holding the same one. Then ask what you are left with. If the value of the thing came from the effort or the skill of making it, you are left with very little, because the machine now makes it too, for nothing, for everyone. That is the rising middle, and it is coming for that part of your work whatever you do. But if what you are left with is something only you hold, the data, the relationship, the hard won read on a particular world, then the free model is no threat. It makes that part worth more, because now there is a powerful tool to act on it and still nothing else like it to act on.
This is not a reason to abandon your craft. It is a reason to know which part of what you do is the craft the machine is learning to copy, and which part is the thing only you have, which it cannot. Most of us have spent our careers selling the first and treating the second as the backdrop. The work now is to turn that around. Take the inventory honestly. Name the parts of your value that survive a free model and the parts that do not. Then shift your weight, on purpose, onto the parts that survive, before the choice is made for you.
That changes the question everyone is asking. It is no longer which model you pick. The models are converging, improving, and heading toward costing almost nothing, and choosing between them will soon matter about as much as choosing between two good calculators. The question that lasts is what you own that a better model makes more valuable, not less.
What stays yours
Jensen gave the models away because he reads the curve better than almost anyone, and the curve says the models are not where the value sits. He was not warning anyone. He was being clear. What he could not say from that stage, because it is not his to say, is where the value went instead.
It went to the things that are only yours. The data no one else has. The trust that took years to earn. The read on one particular world that lives in your records and your head and nowhere on the open internet. The models are about to be everywhere and identical for everyone. That is the whole point of giving them away. The only thing that stays yours is the part of the world you alone have seen.
Find it. Name it. Build on it. Now, while it is still yours to build on.
Where this is going
Step back from your own corner of it, and the shape of the whole thing comes into view. Within a few years, capable agents will run on almost everything that carries a chip. The laptop and the phone, but also the car, the camera on the wall, the machine on the factory floor. Most of that intelligence will be open weight models doing the everyday work where it sits, on hardware someone already owns, because the cost of a thought there falls to almost nothing. The giant data centres do not disappear. They keep the frontier, the hardest reasoning and the next models in training, and the heavy lifts reach up to them when they have to. But the ordinary, constant, ever present work settles down onto the devices in front of us.
This is not a preference. It is arithmetic. There is no version of the future where billions of people each run an always on agent through a metered cloud and the bill stays payable. The only way to put capable intelligence in front of everyone, at a price an ordinary person or an ordinary business can afford, is to push most of it onto hardware people already own and let the meter fall away. Open weight models are what let you run it on your own machine in the first place. Cheap, powerful chips are what make it fast enough to be worth doing. The two arriving together turn this from a service a few companies sell into something closer to a utility, owned in millions of small pieces rather than rented from a few large ones.
And none of this is far off. The first machines ship this year.
Which leaves the question that sits under all of it. If intelligence becomes ordinary, as common and as cheap as power from the wall, and the one thing that separates any of us from anyone else is the private reality we each hold, then what are we really competing on? What happens to the distance between those who already own the most data and everyone who does not? What happens to the firms, the careers, and the institutions built on being the one who knew how? No one has the answer yet. But that, not the chips, is what the keynote was really about. And the machines that force the question are already on their way.
If this gave you a sharper way to see the shift, subscribe. The next piece takes up one corner of that question: what happens to computing itself when the application disappears, when an agent on a machine you own runs your work against your own data, and the software you used to open and operate becomes something it just makes on request.
Craig Hepburn is an AI strategist and Perplexity Fellow. Twenty plus years building at the frontier of digital, from Microsoft and Nokia to Art Basel and UEFA. Now building at the frontier of agentic intelligence.
If this leather-jacketed salesman's products have no meaning to a Pompeii worm living beside a hydrothermal vent I don't see why they should be taken seriously by any other organism. I honestly feel this particular evolutionary route will lead to a dead-end, albeit one festooned with the remanence of a journey of undeniable ingenuity, grandiosity and absurdity, along with all the horror and the beauty an expedition fueled by a faith in the rational individual could ever dream to conceive.
Hmmmmm