After adding $100 billion a month, Jensen Huang says NVIDIA is selling the data center

Sarah Guo and Elad Gil open their 2024 highlights reel with a number that barely fits in a sentence. Since their 2023 conversation with Jensen Huang, NVIDIA’s stock had tripled, adding almost $100 billion of value each month of 2024 and crossing into the $3 trillion club. The clip they lead with is not a victory lap about chips. It is Huang explaining why NVIDIA is no longer a chip company in the old sense, but a data-center ecosystem that builds the whole machine before it sells the parts.

The rest of the year, on their show No Priors, they sat with people who were busy making the rest of that stack. Andrej Karpathy, from OpenAI and Tesla and then from teaching the field how to train networks, talks about AI as an exocortex. Bret Taylor, OpenAI board member and founder of Sierra, talks about branded company agents. OpenAI’s Sora team talks about video as a world model. Waymo co-CEO Dmitri Dolgov talks about the nines of autonomy. Figma CEO Dylan Field talks about whether chat kills the interface. Scale AI CEO Alexandr Wang talks about why AGI will not arrive in one jump.

Cut together, the seven conversations are a map of 2024’s real fault lines, not a consensus memo. Some clips complete each other: Huang’s data-center-scale software and Taylor’s shovel-ready agents are both arguments that the bottleneck has moved off the demo. Some clips collide. The Sora team thinks predicting video will grow more general intelligence; Wang says that story is a nice narrative without strong scientific evidence yet. Guo and Gil do not try to referee. They just put the arguments in the same half hour and let the year sit there.

Guo’s question is concrete. NVIDIA’s unit of customer support has grown from a chip to a server to a rack, the NVL72. What is next — does the company do the full data center? Huang’s answer is that they already do. “We don’t build PowerPoint slides and ship the chips,” he says. “We build a whole data center.” Software has to be validated on the computer in its full manifestation. Until the building is up, you do not know whether the fabric works, whether the efficiencies are real, or whether the thing still runs at scale. That, he argues, is why someone’s actual performance is so often dramatically lower than the peak number on the slide.

He puts it as a change in what computing even is. The new unit is the data center; that is what you have to deliver, and that is what NVIDIA now builds. They do it in every combination they can think of: air-cooled and liquid-cooled, x86 and Grace, Ethernet and InfiniBand, NVLink and no NVLink. The company is running five supercomputers internally and plans to build easily five more the following year. “If you’re serious about software, you build your own computers.” Then, in the move that looks like a contradiction, they vertically integrate, optimize the full stack, disaggregate everything, and sell it in parts.

The reason for the insane complexity, Huang says, is grafting that architecture into GCP, AWS, Azure, and OCI. Each cloud’s control plane, security plane, and idea of cluster sizing is different, and all of them still have to accommodate NVIDIA so CUDA can be everywhere. What he is protecting is an old software principle he does not want given up: write once, run everywhere, modulo the 10 percent that each operator has optimized differently. Software is the expensive investment, and you maintain it for as long as you shall live. He closes the loop with a size comparison: look at the hardware industry, then look at the world’s industries sitting on top of it. One is about a trillion dollars. The other is an order of magnitude that makes the software, not the board, the thing you cannot afford to fragment.

Karpathy’s clip is about what happens when the model is no longer a tool you query but an exocortex, an extension of human cognition. The market structure he sees is an oligopoly of a few closed platforms and an open camp that is behind — Meta’s Llama and the rest. Guo asks what that means for access if this stuff is fundamental enough to democratize. He reaches for the crypto line: not your keys, not your coins. The transposition is “not your weights, not your brain.” If a company controls your exocortex, it starts to feel invasive. You realize you are renting a brain, “and it seems strange to rent your brain.”

Guo’s thought experiment is the other half of the trade. Is she willing to give up ownership and control to rent a better brain? She is. Karpathy says that is the trade-off, and the market will show how it shakes out. The pattern already forming is to use the closed versions by default because they are amazing, and to keep an open fallback you fully control. When closed APIs go down, customers implement fallbacks into the open ecosystem and feel empowered by it. That, he thinks, is probably how people will use a brain: closed most of the time, open when something happens. Which makes it “100 percent” important that open source keep progressing — not an obvious point, and not one everyone agrees on.

The second half of the clip is about size. Guo wants the smallest performant model, in parameters or any other measure. Karpathy thinks it can be surprisingly small. Current models waste a ton of capacity remembering things that do not matter — SHA hashes, ancient trivia — because the data set is not curated. What you actually need is a cognitive core: a thing that thinks, and that knows how to use tools when it has to look something up. Guo guesses three billion parameters, or twenty. He goes further. Even a billion will probably surprise people. Distillation, a big model or a huge amount of compute supervising a very small one, “works like surprisingly well,” and that is why the core can get very, very small.

Bret Taylor draws a year on the wall. In 1995, existing digitally meant a website and a listing in the Yahoo directory. In 2025, it will probably mean a branded AI agent that customers can talk to in order to do everything they can do on the website: ask about products, buy, get service. The category his company Sierra works in is what he calls company agents. It is less about autonomy for its own sake than about how a firm shows up in a conversational world. He calls the work shovel-ready with current technology. Customer-experience processes are already defined. Systems of record already exist. The job is to decide what the conversational experience around the brand should be — not boiling the ocean, and not the same “agent” the AGI labs mean, which he is happy to leave as a different problem.

Two deployments make the pitch concrete. Buy a Sonos speaker, or hit the dreaded flashing orange light, and you now chat with the Sonos AI, powered by Sierra, to onboard or to debug hardware versus Wi-Fi. SiriusXM’s agent is named Harmony, which Taylor calls a delightful name, and it handles upgrades, downgrades, and the trial that comes with a new car. Branded is not a marketing adjective. The agent is part of the brand experience, the way a URL on a business card once was. Look at early websites on the Wayback Machine, he says, and you will see either a phone number or a DVD intro screen of graphics. People will look back at today’s agents with the same sense that they were quaint.

Most customers start in customer service, which he thinks is a great use case and not the destination. Fast-forward three or four years and the agent will encompass all that the company does. His running example is insurance: file a claim, compare plans, add a child to the premium when they are old enough for a driver’s license. All of the above, done by the agent. That is what Sierra is helping companies build. The 1995 website was not optional for long. Taylor is betting the branded agent will not be either.

OpenAI’s Sora team is building a video generation model, and the clip Guo and Gil keep is about what the model learned without being told. It understands 3D even though nobody baked 3D in; 3D was in the videos. It learned that a bite out of a hamburger leaves a bite mark. So much of human life is visual, they argue, that this grounding — this is the world we live in, this is how past events affect future ones — should lead to more intelligent models more broadly than video generation. Guo’s comparison is that they have almost invented a future visual cortex plus some of the reasoning brain at the same time. A team member takes the compliment: a lot of human intelligence is world modeling, playing out scenarios before you act, including in dreams. Building Sora as a world model is close to that.

Gil asks whether the human world model is only approximate, more like dropping an apple than solving a physics engine, and whether large models should aim at the same slop. The answer they give is that the slop is a human deficiency. Long-term, narrow physics is something these systems can improve on; they are optimistic Sora will supersede it and, one day, world-model better than people. Humans remain an existence proof that perfect physics is not required for other kinds of intelligence. Bill adds the scaling moral, which is the Bitter Lesson in all but the citation: methods should get better as you add compute, and what works is the simple, hard task of predicting data. Fancy objectives that live in some approximate simulated space do not pick up the scaling laws. Text was next-token prediction. Video, for Sora, is the same bet.

Waymo is the other face of the same year. Co-CEO Dmitri Dolgov’s company is completing more than 100,000 fully autonomous rides a week in cities like San Francisco and Phoenix, and Guo calls it her favorite way to travel. The question is why advancing driver assist, which works in more and more scenarios, is still not full autonomy. “It’s the number of nines,” Dolgov says. In 2009 a dozen people spent about 18 months driving ten 100-mile Bay Area routes — freeways, downtown San Francisco, around Lake Tahoe — with no intervention. No ImageNet, no ConvNets, no Transformers, tiny computers. Every technical wave has made the start of the curve steeper. Today a good computer-science graduate can take an off-the-shelf VLM, fine-tune it a little so camera frames become trajectories instead of words, and get something that drives tens or hundreds of miles and blows your mind.

That prototype, he says, is not the problem. Is it enough to remove the driver, run millions of miles, and post a safety record demonstrably better than humans? No. The complexity is in the long tail of the many, many nines. You do not see that tail if you are building a prototype or a driver-assist system. That is where Waymo has been spending, and in his telling it is the only hard part of the problem. The Sora clip and the Waymo clip are one lesson with two temperatures. Prediction on abundant data can surprise you with 3D and bite marks. Deployment against a human safety record still lives past the demo, in the nines nobody puts on a slide.

Dylan Field, CEO of Figma, is asked where UI goes in an AI year when products are collapsing into chat and some people forecast an agentic world that does away with most interface altogether. His answer is calm. There is a lot of agent innovation, and the ways we use UI to interact with agents are still at the beginning; those interfaces will get more sophisticated. Even if they do not, new media types do not make old ones go away. TikTok did not mean you stopped watching YouTube. Even if you take as a prior — on this very podcast — that a new form of interaction is chat, you still have UI, and he suspects you have more UI and more software than before.

On multimodality, some people argue that realtime semantic understanding will push design toward verbal conversational UI and away from text. Field does not mock the experiments. There are contexts where voice matters, and voice UIs may even map onto more traditional ones. Personally, he does not want to navigate the information spaces he lives in all day via voice, and he does not want to do it Minority Report-style on a Vision Pro either. A keyboard and mouse with an amazing Vision Pro or Oculus as a monitor could be cool. The pattern he distrusts is the leap from a cool new glimpse to “useful for everything.” New interaction modes have a role. They are not going to be ubiquitous across every interaction. The mania is still healthy, because without it you never find the boundary, and that is how HCI actually moves.

The input he is genuinely bullish on is almost mundane: intelligent cameras, even in the most basic sense. It is really hard for people to describe things visually. Taking pictures and capturing video is something every user already knows how to do. If the camera becomes a first-class peripheral, that is a space he thinks will be useful, and one worth exploring while the rest of the industry tries to make chat eat the screen. The clip is a designer’s veto on two fashionable deaths — the death of UI, the death of the keyboard — and a bet on an input that does not need a prompt.

Alexandr Wang, asked for something he believes about AI that other people do not, does not hedge. The path to AGI looks a lot more like curing cancer than developing a vaccine. A vaccine is one breakthrough, then everyone gets the shot. Cancer is a long list of distinct diseases you have to zoom into and solve independently, over a multi-decade frame, until you look back and notice you have done it. AGI, in his version, is a plodding road of individual capabilities, each with its own data flywheel, and solving one does not give you much leverage on the next. A lot of the industry paints the opposite picture: eventually we will just boop, and we will solve it in one fell swoop.

He thinks that is also the more bullish case for society. Slow, consistent progress for quite some time is a world that has time to acclimate. When Guo tries to cash the analogy out — is the hard part multi-step reasoning, is Monte Carlo research a false idol, are we going to hit scaling walls — his fundamental claim is limited generality. Even across modalities, his understanding is that there is no positive transfer: training on a bunch of video does not really help your text problems, and vice versa. Each niche of capability will need a separate data flywheel. On the specific narrative that video is a basis for a world model that then helps you reason, he is polite and firm. “It’s a great narrative. I don’t think there’s strong scientific evidence of that yet.” Maybe there will be. The base case, for him, is not much generalization coming out of the models, and therefore a long list of little problems between here and AGI.

Placed against the Sora clip from the same episode, that is a direct contradiction, and Guo and Gil leave it on the table. 2024’s actual state, in this reel, is that even central practitioners do not agree whether scale-plus-video buys generality. The implied planning error is symmetric. Betting the business on superintelligence arriving next year, and betting that generality will never come, are two ways to misread a cancer-shaped road. The year-end message from No Priors is not a forecast. It is a set of tests: Huang’s gap between the slide and the cluster, Dolgov’s nines, Karpathy’s fallback weights, Taylor’s shovel, and Wang’s demand for a flywheel in every niche.