OpenAI vs. Grok：构建“Everything App”的竞赛｜Emad Mostaque、Dave Blundin & AWG｜第199期

OpenAI Dev Day just occurred. Some of the most staggering things in human history got announced yesterday, and they still undersold it. Good morning, and welcome to Dev Day.

OpenAI is really trying to do a global land grab, right? Going into India, going into the UK, going into Greece, and then you've also got all of the open-source models coming out of China. We're just at this tipping point, and the tipping point is in the next 6 months.

What happens when suddenly we're able to 5x, 6x, 7x the amount of broadly accessible superintelligence across the world? I think this starts to become the foundation for transformative economic changes at a planetary scale.

Now that's a moonshot, ladies and gentlemen. We're spinning up this episode real quick with my extraordinary moonshot mates because OpenAI Dev Day just happened. I want to cover the subjects there, but there's a lot happening across the board in robotics. Tesla's FSD 14.1 just dropped, as well as other robotics updates and data center updates.

We've got my moonshot mate Dave Blundin. Dave, good to see you, pal.

Good morning.

And of course, we have AWG live from someplace in hyperspace. Good to see you, Alex. Welcome back. Then we have one of our other moonshot mates, Emad Mostaque, coming in from London. Emad, good morning to you.

Morning.

Or good afternoon, as the case may be. You know, in the rocket business, there's something called hypergolic fuel. Hypergolic fuel is when 2 chemicals come together, explode, and make a propulsive force. I think about AWG and Emad coming together as my hypergolic fuel this morning.

Better than coffee.

Completely agree. Absolutely. All right. As always, this is the news that's breaking that I think is impacting the global economy, impacting our mindsets, impacting how we teach our kids and run our companies. Nothing is more important to me, so let's jump in.

The reason we spun this up for everybody is that OpenAI Dev Day just occurred. I want to hit on this and really evaluate along the way how critical this is, how rapidly it's happening, and how Sam is manipulating the future of his company and AI in a positive way. We're going to discuss this.

I'm going to open up with a short video clip of Sam opening up OpenAI Dev Day yesterday. Let's take a listen.

Back in 2023, we had 2 million weekly developers and 100 million weekly ChatGPT users. We were processing about 300 million tokens per minute on our API, and that felt like a lot to us, at least at the time.

Today, 4 million developers have built with OpenAI. More than 800 million people use ChatGPT every week, and we process over 6 billion tokens per minute on the API. Thanks to all of you, AI has gone from something people build and play with to something people build with every day.

We think this is the best time in history to be a builder. It has never been faster to go from idea to product. You can really feel the acceleration at this point. To get started, let's take a look at apps inside of ChatGPT.

Obviously, Sam's not the best. He's not Steve Jobs on stage, but the numbers are just staggering. The increase from 300 million to 6 billion tokens per minute is the one that really jumps out.

It's going to explode from here forward, too, because I could easily consume 10,000-plus tokens myself just coding. With the number of developers coming on board and the number of home users coming on board, it's just astronomical. As we've been saying, there's nowhere near enough compute to keep up with it.

I think Jony Ive might be the guy driving the, “Hey, let's do this Steve Jobs-style. Have our very first big-stage developer day.” Their GPT-5 launch was really flat—really, really flat. They did a much better job yesterday. I've got to believe Jony is driving that. Put some money and some effort behind it. Let's go.

Some of the most staggering things in human history got announced yesterday, and they still undersold it relative to the implications. We'll see it in a couple of other videos here. I don't know if that's deliberate slow-playing because they don't have enough compute to keep up with the demand anyway, or if they're just learning how to do showbiz on the big stage.

In any event, we'll see some more mind-blowing capabilities that, if anything, are understated. Eight hundred million users is pretty extraordinary. They're tracking for a billion users, and I wonder: Is this a winner-take-most type of scenario, or is there anything that can overturn the final result? Let's go to you, and then we'll go to Alex to bring us home on this one.

To put it in context, it's a lot, but it's about as many weekly active users as Snapchat. I know which one is going to have a bigger impact on the world between the two. I think there's still so much upside to come from here, but now you're seeing their model with Sora 2 and others moving maybe toward an advertising model as tokens get cheaper and as they get faster.

To put the token numbers in context, 6 billion a minute is 3 quadrillion tokens a year. All of the humans in the world together speak 50 quadrillion tokens a year, and I expect that number to go up 10 times.

So next year, OpenAI is probably going to be at 30 quadrillion, and then, by themselves, the year after, they'll overtake—in terms of tokens—all the human words spoken every single year.

I think we're getting close to that because Google said they're doing a quadrillion on their billion active users right now because it's in search and things.

I think we're at that tipping point now where the number of AI tokens coming into the world is about to overtake humans. Maybe we should call it a something day, right?

Quadrillion here, quadrillion there. Yes.

Quadrillion here, quadrillion there.

Yeah. Alex, what's your take on this opening commentary?

I think we're really far from saturation. In addition to being, call it, at about 6% saturation by comparing the number of OpenAI-generated tokens with human-spoken tokens per minute, I think there's probably an even more important statistic: There are approximately 4 billion human users of smartphones who aren't yet using any sort of superintelligence, if you will.

Now ask yourself what happens when suddenly we're able to 5x, 6x, 7x the amount of broadly accessible superintelligence across the world. I think this starts to become the foundation for transformative economic changes at a planetary scale.

And you're limiting that to humans. Of course, humans might be the least significant users of superintelligence in the final result.

With full autonomy, superintelligence is arguably the ultimate user of superintelligence.

There's a limit to the number of words we can say. It's about 20,000 a day. Our thinking tokens are 200,000 a day. AI has no limit to the number of tokens, or economically valuable tokens, it can produce except for the GPUs. That's the only limit.

On our last podcast, we talked about Sam saying we're going to have to make a trade-off between tokens used for education for our children and tokens used for healthcare to save lives. We can't—we don't have an infinite amount of compute, and I don't want to make those difficult decisions.

But if FSD is coming online and all these cars are going to start driving themselves, the quality of the driving is directly tied to the amount of compute available. We're imminently going to make very difficult decisions around tolerating a very rare car crash versus giving somebody the ability to build something at home using AI.

It's just incredible, the difficult decisions that are coming immediately after all these functions that we're about to see get deployed. At the same time, we're also limited by energy, which we'll talk about in this conversation.

I'm going to move us to a few of the features from OpenAI Dev Day. We'll see a few things. I didn't show the video here, but one of the primary high points was that they're talking to apps within ChatGPT. They have an Apps SDK and the ability to say to Booking.com, “Book me this trip”; to Figma, “Diagram this”; or to Coursera, “Teach me this.”

You can just speak to Zillow. It's the ultimate interface with all the other apps out there. What's the significance of this for you, Emad?

I think attention is all they need, as it were. The battle here is that human attention is finite, and so OpenAI, Meta—everyone's making a play for who you're talking to that then enables these MCP-enabled agents to come and do the job.

What is the Tencent-WeChat-type super app that's coming together? Everyone's folding themselves into these nice kinds of things, and again, that's how they're going to try and monetize. You'll see this battle between Meta, via WhatsApp and Instagram; Google; and OpenAI to try and occupy that real estate. Then, of course, Elon's going to come in with X and all sorts of interesting things coming.

Everyone's trying to be the everything app.

Yeah, for sure.

Dave?

Well, in a second, we're going to see actually something built by voice.

Why don't we look at it, and then we can—it's actually really cool when you see it.

The Codex example.

Yeah, yeah. All right, we'll come to that. But before, I'm just wondering: when OpenAI drops this capability, are they picking winners in the final result? Are they going to be equally spreading their attention across everybody? Are they basically eating away all the entrepreneurial startups? There was a tweet that went out. I was trying to capture it, but it said, “Okay, OpenAI just eliminated a million different startups out there working on their approach.”

Remember, every platform ergonomically wants to have its own app store. I think that the notion of an app store being built on top of a new platform, where ChatGPT and presumably other frontier models as well want to become the new operating system—or certainly rhyme with the Facebook platform moment when Facebook launched that—is a very natural market movement.

But I would also caution that, at some point, I think it's reasonable to expect that every pixel is going to be generated. It's not just going to be vector art or HTML-type graphics. Every pixel is going to be generated. I would view this as almost a transitory moment where apps are floating on top of ChatGPT as the new operating system environment, but it's a passing phase. At some point, every single pixel probably wants to be generated.

That was my first thought. The other thought is, do you remember, Peter, back in 1987, when Apple, without Steve Jobs, launched their Knowledge Navigator concept?

Yes, I do. We're living in that now.

We're living in that, where a professor is having a conversation with a basically similar type of canvas that is able to pop open new apps and interact with them on demand. We caught up with the future approximately 40 years later. We're living the Knowledge Navigator future.

One of the examples they had on their live demo stage was an individual proposing a new startup. In this case, it was a dog-walking app. They said, “Okay, create me an image for it. Create me a name for it.” Then they said, “Okay, Canva, turn this into a deck. I want to raise money.”

At the end of the day, we're not too many steps removed from ChatGPT: “Start this business for me and start wiring the revenues to this location.”

I think that's the multitrillion-dollar endgame here, where at some point we see autonomous corporations.

Mhm.

I literally did exactly what you just said, Peter, yesterday at a red light in Cambridge. As I was sitting there, I created a business plan and tried to recruit a Princeton team into it via AI at the red light.

That's like when Elon said, when he was driving from SpaceX back to his home in Beverly Hills and there was traffic, “Damn it, I'm going to start a tunneling company. I'm going to call it The Boring Company.” There's literally a future in which we're going from mind to materialization. It's stating what you want to do and having the universe conspire to create it for you.

That's crazy. I think he has The Boring Company, but then he has his even cooler name of Macrohard, his new software company.

I love that.

Against Microsoft. Elon is a 13-year-old kid for sure, which is literally trying to do this. It's trying to turn ideas into full companies entirely digitally, right?

I think what you've seen is 3 phases. Consumption was expensive. It became cheap. Creation was expensive. It's becoming cheap. Now the valuable thing is curation and attention. Again, the battle is who can have that value for the pixels that you see and the noises that you hear. A lot of that creation element is going to be abstracted away, and I think all the big players realize this.

The question is, where does it end up? Where does it go eventually, Dave?

That quote that you had, I've heard it 100 times: “A million startups just died because of what they rolled out yesterday.” It's absolutely not true. Show me the names of those startups that died.

This came up when we were talking to Amjad Masad a couple of weeks ago on that other podcast. The founder of Replit had to build his entire foundation model from scratch to get to market because it was before OpenAI had the APIs. You ask him, “Do you regret that? You had to throw away all that code?” He's like, “No, I absolutely don't regret it. You constantly have to change. AI is going to move at this ridiculous, accelerating pace.”

You're not reinventing your business constantly, you're dead on arrival.

Yes, exactly.

Your team is intact. If you have a great team and you're in AI, you will succeed every single time. Maybe something you do gets crushed by the next iteration of OpenAI, but you pivot so quickly and easily, just like we're talking about right now. Show me the names of those companies, those million companies that died. They don't exist.

All right, let's jump into the next demo they had at OpenAI DevDay. It is Agent Builder, creating multistep workflows without coding. I'll just show the first few seconds of this.

To make this interesting, I'm going to give myself 8 minutes to build and ship an agent right here in front of you. I'm starting in the workflow builder in the OpenAI platform. Instead of starting with code, we can wire nodes up visually. Agent Builder helps you model really complex workflows in an easy and visual way, using the common patterns that we've learned from building agents.

All right. Emad, you're building agents left, right, and center right now for Intelligent Internet. What do you think of this?

Yeah, I think you've gone from creation to composition and the multistage process for image generation and media. We built something called ComfyUI, which again is this node-based process. But where we're going, we don't need nodes and spaghetti.

Exactly.

The future of these things—you can look at our Common Ground platform, for example—is that it flips between Kanban, workflows, Gantt charts, and things like that. It will just show you what you need to show. The way that you'll interact with agents is like interacting with Jarvis in Iron Man.

I think in a year or 2, that's what Agent Builder is going to be. You'll just have a nice chat, and it will show you all these things and mock them up instantly. In fact, Claude had this with its latest release for Pro users: this instantly generated app-desktop-type thing that literally programmed things on the fly without code, because code is just a human translation layer, and that can be removed completely.

For sure, Dave.

Yeah, it's funny because the people succeeding in AI are overwhelmingly really young, really, really smart, with very limited business experience, and they keep recreating the same mistakes from 20 years ago. It's okay because AI is such a great tailwind.

But this graphical programming language—those lines—is the stupidest thing in the world in the age of AI, where you can talk to the AI. It's very similar to the Cursor interface. If you want to upgrade your account and you're talking to Cursor, like, “Hey—” or to Claude 4.5, “Hey, Claude, upgrade my account,” and it says, “Well, go to the menus, navigate to the settings...” What are you talking about? I'm talking to you right now. You have MCP. Just do it.

So that'll all get fixed very quickly, but it's crazy. The whole interface to AI is going to be voice—

Voice and images.

—and the idea that you're going to design programs by drawing boxes and connecting them with lines, which has been around since around 1980? No, no, no, no. It'll get cleaned up very, very quickly. It's just kind of funny to see this transition phase and all the same old mistakes being made.

Alex, any other points you want to make on Agent Builder?

Yeah, I'm reminded, almost by analogy, of the early days of Hollywood that were shaped around vaudeville-type design patterns. I think that's the stage that we're at. This is the vaudeville-on-Hollywood-screens stage of AI for software development.

On the one hand, it's great that it exists and provides probably a comforting safety net for enterprises that are migrating to end-to-end agentic workflows. On the other hand, it very much feels like a passing phase. I think it's an interesting modality for specifying software development, but really, why not take a full leap? Instead of specifying flowcharts for individual workflows, why not just dump in an entire org chart of a human organization and, rather than individual tasks or roles, port over an entire company work chart and ask for an entire enterprise to be replicated?

The Vogill analogy is great. If you look at an old, original black-and-white movie, the actors are on an actual stage with actual stage lights, doing a play, and they're just recording it on film. It's very similar to when the internet took off and the newspapers just copied the exact newspaper over to the internet and said, “Oh, we'll just read it over here instead.” Same font, same everything.

We humans are so stuck in the way we do things. We have such a high cost of switching. All right, continuing on this theme of voice mode over everything, here's a little video of controlling Codex with voice mode. Take a listen.

Let's see if we can actually control this camera now with this Xbox controller. There we go. It works. That's pretty awesome. What's funny is, I didn't even specify to Codex which button should do what, but it figured that the joystick was probably a good idea. So that's actually perfect. But let's do the real test now. Let's enter voice mode and see what happens. So, hello.

Hi there. I can hear you loud and clear. Let me know what you need.

Yeah. What do you see on the camera?

I see a large audience in front of you, with rows of attendees seated and ready. The room is well lit, and the crowd stretches far back, filling the space.

That's right. Could you shine the lights toward the audience?

There we go. I've just lit up the audience with a nice, bright wash. They should be perfectly highlighted now.

All right. This is an unusual demo for Codex, to say the least, just showing our ability to control physical items in our environment. But I have to say, I love this for one reason: I've always said I will invest in the first company that's able to make AV foolproof. How many of us are in conversations or giving presentations and we can't get this to connect to that? I just want to be able to say to my AI, “Show that video on that screen. Connect me over Zoom to that screen.” So there's hope still.

And that's a business plan. It's not hypothetical. That's a business plan that, if someone's listening right now and they get together a team, and then Peter's seed investment invests in it to give it credibility, will become yet another one of these massive success stories. It's really that simple.

I'm on stage and I say, “Something fails. Okay, AI is easy, AV is hard.” Well, hopefully AI can solve that. But also, controlling things: right now, you kind of wave to people backstage and they push some buttons or whatever. It's crazy because the AI can now recognize your hand gestures and respond to your voice. It's much more engaging for the audience if you're talking to the AV and it's changing the lighting, changing the slides, and pulling up things from the internet in real time.

Very doable. You could get that product out the door in 6 months or less and just crush it. And, of course, it self-demos. Then Peter will bring it into the podcast, and it's just that simple.

There you go. Let's start with Alex. Alex, what's your thinking?

I think it's sort of interesting because so many facets of Codex are open source and available for review on GitHub, so you can actually trace where it appeared.

Tell us what Codex is, first of all.

Codex is an OpenAI brand that seems to cover a number of different, independent software tools. It covers their code-generation-specific AI model backend. It's also used as a web front end for agentic software development. It's also used in connection with a command-line interface tool. So they use it as an umbrella brand, as it were.

But in this case, at least one of Codex's associated projects is up on GitHub. You can review the source history. Pulling the thread on the story, it was interesting, I think, to discover that at least some aspect of this functionality appears to have originated as a feature request from a third party, from the Carnegie Mellon-affiliated Software Engineering Institute, back in April. That was where a user was complaining—or really pointing out—that human prompt-typing speed had increasingly become a limiting factor for software development.

By the way, I want to read a quick tweet here that came over from an OpenAI employee. It says, “Agent Builder, which we released today, was built end to end in under 6 weeks, with Codex writing 80% of the PRs. This matches the AI 2027 report, in that it was forecasted in 2026: coding automation goes mainstream. Agents will work like teammates. AI R&D is 50% faster from algorithms.” I mean, we're seeing science fiction—or, I want to say, science predictions—tying much closer to reality.

We're close to the point of recursive self-improvement, and it can go in the other direction as well. We can get negative speed, where the software is just written preemptively.

Interesting. So we're not smart enough to realize we need the software, but the AI is, and it's prepped for us in advance.

Exactly. I love that.

Emad, your thoughts on Codex here?

Yeah, if you get enough tokens. I think this is the thing: most people are just using half a million or a million there, and to build something like that is 5 bucks with the new Grok model. It's 50 cents. You're seeing a crazy thing.

I want to come to that after we close on this: how would you be disrupting OpenAI if you were going to? Dave, do you want to comment on controlling Codex with voice mode?

Well, something Alex said really sparked a thought. When Codex launched, it was a way to run 5 or 10 different coding processes in parallel and have status checks, and it made you much more efficient. But now they use it to bundle 5 or 10 different things together.

This is going to be a real problem because we're used to products having a very specific name and brand and doing a very limited number of things. But with AI, the explosion of capabilities is exponential, and you can't even keep up with the names. Now it's going to be much more like thematic branding. Codex is a grab bag, GPT is a grab bag—but what else can you do? There's just so much going on. Keeping up with the names of things, and naming things in general, is going to be—

We had that conversation with Kevin Weil at OpenAI: naming protocols are kind of insane.

If we can add one thing: 6 months ago, Dario Amodei from Anthropic said that 90% of code would be written by AI. I think he meant it can be written by AI, and this is a really great example of that.

They decided to embrace it, so you see Codex—the CLI, the command-line interface tool—literally gets 2 updates a week, which, for a multibillion-dollar, half-trillion-dollar company, is unheard of. And so, again, as Alex said, I think you're going to get these recursive self-improvement cycles, first with humans in the loop, but then the software might just upgrade itself and respond to what people might need, continuously.

What's interesting is how that interacts with the interface. If you said, “My iPhone is going to update itself twice a week,” you'd be confused as all hell. You'd never know where anything is. But now that you have an AI interface on everything, it's okay because it's self-explaining. It's seamless.

I just want JARVIS. I just want an AI I talk to, and it does everything I need to get done. I'm going to assume that anything is doable, and my AI is going to enable it or find the capability. I don't need to know all the hard work it's doing on the back end. I don't need to know what it's calling up or getting access to. It's just making it happen.

I'm telling you, Peter, within the virtual world—not within the physical world, but within the virtual world—that's today. No one's productized it yet, but all the technology and capability exist right now. The robotic version, where it makes your Iron Man suit, might be 1 or 2 or 3 years out. In the virtual world, “Build me a video game, build me whatever”—that's right now.

Yeah.

Someone needs to go and get Paul Bettany's voice rights.

Let's take a look at one more video from OpenAI Day, which was the Sora 2 API, and a segment I call “Sketch to Video.” Again, this is going from mind to materialization. If I can imagine something, can I make it real? Take a listen.

Today, we're releasing a preview of Sora 2 in the API.

Mattel has been a great partner, working with us to test Sora 2 in the API and seeing what they can do to bring product ideas to life more quickly. One of their designers can now start with a sketch and then turn these early concepts into something that you can see, share, and react to. So let's take a look at how this works.

If you're listening to this podcast, what we're seeing here is basically a hand sketch being developed into a photorealistic video of a Mattel Hot Wheels or Matchbox toy. Super compelling.

Being able to go from that—and I've talked about how, in the future, I'm going to be able to describe verbally what I want: “I want a device that can hold a hot liquid. I want it to have a handle. I want it to be this color.” As I'm describing it, it's visually materializing on my AR glasses in front of me. I say, “No, can you make it a little bit larger? Can you stretch the dimension?” Just in plain English.

Then I ask, “How much would it cost to make?” It gives me a price. “Can you give me an alternative that's cheaper or that has better thermal insulation?” They say, “Yeah, that's it. Please print it for me, manufacture it for me, and put it up on the web so anyone can grab it.”

I mean, this going from—again, I call it mind-to-materialization—is super powerful. Emad, do you want to open up on your thoughts on this?

I mean, the holodeck is getting closer, right? Not with hard lights, but as you said, that aspect is there. These models learn physics; they learn material. So, in the video that was just shown, the car goes down these ramps and it's transformed into 3D. You can have 3D extensions from it.

You can actually do a storyboard where you show, scene by scene, how every single thing changes, and you have that as the input, and it'll generate that clip. It doesn't do that by thinking or breaking it apart. It literally interpolates the concept to the video.

We're actually only scratching the surface of how powerful these models are at the moment. As they get more and more use, you'll see that they are genuinely world models that can create anything you can imagine and then adapt on the fly as well, with audio to match—with perfect audio to match.

Yeah. I think everybody has access to this, right? They just launched the API version of it yesterday for large-scale use, but anyone can do this. And if you haven't done it, you're crazy. Do it. It's so mind-opening and compelling. Do exactly what Emad said. Do it as a series of scenes.

Right now, you've got to wait about 5 or 6 minutes to get your video back, which is really annoying. It shows you the compute bottleneck, though. I'm sure when they do it internally, it comes back in a millisecond. You know, it is physically possible to do it very, very quickly. But again, there are way too many users for the capability.

You've got to try it because, again, it's mind-opening. When I first saw this video, I didn't get it because I couldn't tell the video was actually synthetic. I thought, “This guy's sketching a toy, a Mattel toy, and here's the toy. So what?” Then I thought, “Oh, wait—that toy doesn't even exist. That's actually been synthetically created.” It's just so good. The video has perfect physics. You just would never know that it's synthetic. Alex, what does this mean in the final result? Where are we going here?

This is mechanical design getting solved. MIT's mechanical engineering department has an entire set of courses devoted to training the next generation of mechanical engineers how to do product design like this. We're seeing, right before our eyes, an entire discipline—or subdiscipline—get solved in bulk by generative AI.

Maybe even more interesting than this particular instance is the API pricing. If you go to the API pricing page now for Sora 2, it's 10 cents per second for the base model. You do the arithmetic: that's $360 per hour. Assume 10× year-over-year hyperdeflation in model costs within the next year. Suddenly, it's far cheaper to outsource mechanical product design to an API call to Sora 2, or whatever it evolves into, than to a human. That's an entire hyperdeflationary field getting solved overnight.

Okay, keep those numbers top of mind because when we start talking about compute in a minute, and the cost of compute, you'll immediately recognize what Alex just said: that 10× deflation in price. We need that desperately because the demand for what we're seeing here is going to be orders of magnitude bigger than the amount of compute currently available.

What an amazing time to be a kid, right? Imagine you're sitting down with your mom and your dad, and you're just describing what you want as a toy or what you'd like your toy to do. All of a sudden, it's materialized into a video for you, and then some other enterprising company in the 3D-printing world says, “I can just manufacture that for you as an N of 1.” I mean, just amazing.

Star Trek replicators aren't 24th century. They're now—just 2025.

We are really bringing Star Trek to today. That makes me so happy. I'm so happy about that. Why wait a few centuries?

Yeah, for sure. All right, so we're going to wrap on OpenAI Day there, but I'd like you to jump in here for a second. We're seeing a half-trillion-dollar valuation for OpenAI. We're seeing OpenAI really working hard to create multiple revenue flows from advertising, from selling products, and in a multitude of other areas. What are your thoughts on OpenAI?

I think that the core business of OpenAI, in terms of the monthly ChatGPT subscription, is going to come under challenge because we've seen this breakthrough with DeepSeek, Grok 4, and others, where the cost per million tokens has literally dropped 20 or 30 times. The basic chat experience is not good enough anymore.

The chat experience is basically a couple of bucks a year when you calculate the cost reduction, down from $200 just over a year ago. So now they have to think about agentic workflows. They have to think about economically valuable workflows, and then even beyond, because the number of tokens goes from 2,000 to 20,000 to 200,000 to 2 million.

This is why, when we see Sora, they're doing likenesses, and they'll be doing advertising and more, because how do you have the cash flows to justify that? Google and Meta both have the advertising cash flows.

Mhm.

How do you monetize those 800 million users? Either by delivering excess value through your $20-a-month subscriptions or by having these new verticals, because your competitors are going to release what was your key product at the start of this year—ChatGPT—for $20 a month, for free. That's how far and how quickly token prices have dropped.

Yeah. We've talked about the notion that OpenAI is really trying to do a global land grab, going into India, going into the UK, where you are, Emad, going into Greece, and other locations. It's an interesting battle between this land grab and all of the open-source models coming out of China, which are going after a land grab as well.

Well, Paul Graham said, “Sam Altman, if you dropped him on an island full of cannibals and came back a year later, he would be running the island.” He's got to be one of the greatest business strategists of all time.

He's going after India. He's going after a massive installed base. He's got an 800 million-user installed base. He's going after the rest of the world, and he's also going after the data centers. We'll see that later in this podcast.

I think he's narrowed in on the 2 foundational points of control in this great battle: an installed base of users and massive amounts of compute. If you control the endpoints, everything in the middle will fill in. That's the way I think he sees it.

1. Meanwhile in the continuing AI wars

Let's hit on a few others. Anthropic nears superhuman computer use. Here we're seeing a graphic of performance as a percentage, hitting human performance very close to it. We're seeing, basically, over the last year. Alex, do you want to kick us off on this one?

Yeah. So maybe a comment first on what the benchmark is. In the past on this podcast, I've beaten the drum for the importance of benchmarks more broadly—not just for measuring progress, but also for accelerating progress.

In this case, the benchmark, OSWorld—Operating System World—is a really lovely benchmark that was initially developed by Salesforce and colleagues. It's a benchmark that measures the ability of a computer-use agent—an AI that has access to a keyboard, mouse, and screenshots—to conduct regular, everyday, economically important tasks on Ubuntu Linux, Windows, and macOS. Hundreds of different types of tasks.

What Anthropic is demonstrating with this chart is probably, again, by the law of straight lines, perhaps by the end of this year, in the next few months, we're going to see, at least from Anthropic—putting aside other frontier labs—superhuman performance in the ability to control computers for normal, everyday tasks.

So, Alex and Emad, I asked Perplexity Comet what this benchmark even means. It's really vague, and it came back with some complete garbage answer. Hopefully, you can fill me in. What are we measuring here?

We're measuring the ability for an AI to literally control a Windows-type interface with mouse and keyboard control and perform a variety of tasks, including web-browser navigation.

It's the ultimate verbal interface: “Do this for me,” without the verb.

I don't need to know. I mean, I just set up a new MacBook Pro, and getting all of the settings back to where I wanted them ate up half a day of wasted time.

Yeah. Yeah.

Yeah.

Peter, you want Jarvis. This is Jarvis, albeit not in the physical world—for controlling your computer across applications. There are companies that are also setting up giant science factories, controlling hundreds or thousands of experimental devices, where it's basically putting an AI layer on top of all of them and running 24/7 dark experiments to ferret out the breakthroughs of science.

Anyway, Emad, do you want to add to this?

You know, it's 360-odd tasks that take over your computer. Outside of the labs and things, it's a different kind of reinforcement learning environment. I think what this is showing is that these generalist models are getting good enough to do most human-standard tasks. Again, these models have economies of scope.

So now we're seeing Thinking Machines and others building RL environments so they can plug into the real world even more seamlessly. I don't think anyone believes that line isn't going to break through the human level. Again, this is the takeoff point. When they can control anything we can control digitally and then physically, then it's only a question of the number and quality of tokens behind that. Again, this is the takeoff point. This is why we're about to see—

What could possibly go wrong?

I would say, what could possibly go right? Quite a bit can go right.

Yeah. Okay. Thank you for bringing me back to the world of abundance, Alex. I appreciate you.

Anytime.

All right. Our next news item here is a major update to Grok Imagine, going from version 0.1 to version 0.9. I love the numbering protocols, everybody. Also, Elon is entering the gaming world—or at least announcing it. Elon, if anything else, is a gamer, and the video-game industry is massive, outweighing Hollywood entertainment by a long shot.

Grok Imagine can generate 15-second clips, and their comment is, “We're focusing on speed and fun.” All right, let's take a look at some speed and fun.

I love that it says, “Grok launched as a truth-seeking AI.” There you see Elon as this medieval emperor in battle. It's like, okay, this is the truth. We're seeing this. But I think even just taking that line—truth-seeking—and combining it with these models, I completely buy the notion that video, as a first-class modality when incorporated into chains of thought, is going to help us discover the truth. I think it's one of the key modalities for understanding our universe.

What does that mean, Alex? Dive in a little bit deeper, please.

When you ask a question of ChatGPT or some other frontier model, now post-ChatGPT-5, there's thinking that goes on, usually under the hood. It thinks internally in a sequence of tokens before it produces a final answer. Right now, almost all of that thought takes the form of text tokens.

But imagine a near future where the agent is able to think not just in terms of text, but in terms of video. It's able to hallucinate a short video clip—basically, visual imagination. You can introspect as well. You can pop open a little dropdown and see the little videos that it's generating as part of its chain of thought before it answers your question.

Video reasoning, I think, is going to end up being a killer app for how these video models—which right now are obviously largely aimed toward entertainment—end up delivering transformative economic value.

Amazing. In our occipital cortex, our neocortex for visual-image understanding processes much more data than our ability to bring it in through language.

Yeah. OpenAI did $4.3 billion in revenue in the first half of the year. The video-game market did $200 billion in revenue last year, so you can see that when we think about gaming and when we think about media, this is a massive market to go after. xAI and Elon are going to go after it from a first-principles basis. Whether or not the games will be any good, that's a question. I think they'll probably be quite addictive.

Again, the scarce things in the world are Bitcoin, my financial coin, and human attention. The battle for human attention is the next battle for revenue.

And everyone is basically drawing their lines and getting their GPUs ready for it.

I think we'll see this type of thing from everyone, and it's good for consumers in many ways because the quality bar will lift and access will expand.

All right, let's jump into our next segment: chips and data centers. A lot is going on there, but probably the single most important news is that AMD and OpenAI announced a strategic partnership to deploy 6 gigawatts of AMD GPUs. Dave, let me go to you, buddy.

Yeah, the stock moved 30%, which shows you Sam's ability to morph the world—or warp the world—to his perspective, or whatever he says. It had a massive impact on a huge public company. It's interesting: OpenAI is going to get 10% ownership if they hit milestones, basically for no price. How often do you get to negotiate a deal like that? Unless you're the president of the United States, in which case you can negotiate all the time.

Yeah, I guess that's true. The reason this is a serious win-win, though, is that AMD has capacity to manufacture with TSMC. Anyone can design inference-time chips and sell them, but you have to have the manufacturing capacity. So Sam's going to grab that capacity via AMD.

I'm really curious, on November 14, to look at Leopold's 13F filing from the Situational Awareness hedge fund and see if he also bought AMD and got that 30% or 40%.

He probably did.

Probably did. Yeah. I mean, Dave, couldn't we have predicted this as well? At the end of the day, talking about Intel and the criticality of that capability, you could have said the exact same thing about AMD. Who else? There's Broadcom, there's Micron. Which of these other chip manufacturers are going to be pulled into this U.S.-centric, chips-first strategy?

If you drill a layer deeper underneath the chips, there's a whole bunch of other material that will get dragged into the vortex that no one has quite realized yet. If you really want to see these 30%, 40%, or 50% pops, look a layer deeper than just the chip companies into the underlying infrastructure. You've got silicon boules, you've got glass, you've got all this underlying manufacturing infrastructure that's all just going to get sucked into this same exact vortex.

A lot of those are public companies, and some of them are smaller, too, so the movement is much bigger.

Yeah. Emad, thoughts on this one? He's probably just calling everyone now and saying, “Hey, you want to give me warrants? Your stock price will go up,” to all the companies.

Can you imagine if he ironically did this exact deal 50 times back-to-back? The amount of value that would create—oh, my God.

Just all the SaaS companies. Come on, partner up with me.

If you want to do a deal with EverQuote, I'm the chairman of that one. Just give me a call. We'll do this deal tomorrow. You're right, I do wonder if he's using GPT-6 Pro to come up with these deals, but this is massive.

If you look at the 10 gigawatts that they're doing with NVIDIA and the 6 gigawatts here, it's about $50 billion of buildout per gigawatt.

So it's about $800 billion of buildout—like a trillion that they've already got, I think. Probably more.

Amazing. And completely sold out.

Sold out years in advance. Again, the only market that can sustain this and increase the revenue is if they're going after the entire—basically, all software jobs, effectively. So I think in the next few years you're going to see OpenAI and others replicate the whole Macrohard strategy of fully autonomous workers.

That is the product that they will bring to the market, and they will cost $10,000, $20,000, $30,000, or $100,000. That’s the only thing I can see that will fill this particular massive amount.

Alex, are you going to stick with your efficient-market hypothesis from 2 podcasts ago, or are you going to start tracking the tail number of Sam’s jet and seeing who he’s meeting with next?

Well, one might imagine losing sleep as a public-market investor that maybe the singularity, as it were, happens in some private company where there’s indirect, at best, exposure via public markets. What happens if OpenAI and, call it, the 10 other largest privately traded companies suddenly have an intelligence explosion and are worth tens of trillions of dollars overnight? As a public retail investor, that’s perhaps a suboptimal outcome.

So I would actually view this through a very positive lens: through indexing, through exposure to AMD, Intel, and so forth, this is now an enormous jump in exposure to OpenAI, to the extent that an intelligence explosion happens there.

Let me hit on a couple of these related stories. BlackRock is buying up to 78 data centers totaling 5 GW in a $40 billion deal. We’re also seeing that Corning is poised to dominate AI data centers with optics—obviously, fiber optics for connecting everything. On these 2 topics, Corning and BlackRock, let’s get some commentary there.

Yeah, I mean, that’s the war for the downstream kind of elements here, right? BlackRock is coming in with that $40 billion. They’re coming at about 3 times what the normal multiples are. You need to deploy capital, and this feels like currently the best capital to deploy. Downstream, Corning is kind of optimal here, but half—I think something like half—of all GDP growth in the U.S. this year is AI.

Which is insane. I mean, comparing to where we were even just 1 year ago or 2 years ago—

It’s an economic transformation story for the U.S., at a minimum. Many of the campuses of the company that BlackRock is purportedly considering buying, which they’re converting to data centers, are brownfields, including, according to public reporting, a former coal plant in Ohio.

This is what economic transformation—industrial economic transformation at scale—looks like. Again, we’re on a war footing. We have to realize that we’re just pre–World War II. We’re converting automotive plants into aircraft plants. We’re tooling—I mean, you know, Santa Monica Airport, where I fly out of, was basically built out as a secret manufacturing and airport hub. It’s happening, and this is programming of the entire industrial base.

And also, it’s another investment theme, just for our investment-oriented listeners. One of our best and most prolific partners, Kush Bavaria, is starting a new company with Alex’s help that funnels money into data centers.

But it’s part of a broader theme: if this is half the GDP growth of the country and accelerating, there’s all this pent-up capital all over the world that’s not investing in things like Corning. If you can create new conduits for money into all the implications—you know, Alex has been talking about photonics for months now, and the leap from there to saying Corning is going to benefit is not a huge leap—then the capital just needs to get into these avenues to keep this engine humming.

New entities, new funds—BlackRock is obviously very, very smart money pouring into this area—but then all the other implications: data centers in new geographies, pumped hydro, and what about the equipment for pumped hydro and solar installation costs? All those things are investment opportunities.

I mean, is this an infinite sink? In other words, is it going to attract as much money, capabilities, and resources? Is there any moment where the music stops and there aren’t enough chairs for everybody who’s invested?

It’s easy—super easy—to calculate. Now, it’s an infinite demand, no doubt, but it’s limited by chip fabs. If it gets overbuilt or overinvested, it’s purely because there’s too much of X for the number of chips. But the upper bound is based on the chip fabs, and you can see those coming 4 years in advance.

It’s all bottlenecked at Intel, TSMC, and Samsung. From there, you can do all the math in both directions in terms of data centers and users and everything.

Go ahead, Alex.

My mental model continues to be that the music can continue as long as the transformative applications continue. As long as we’re driving the cost of the service economy to 0, and as long as transformative discoveries and scientific inventions pour out of these superintelligent boxes, then the music can continue.

The data center buildout can continue to the point of trillions of dollars of capex. We just need the transformation to continue and the revenue generation that results from that. The transformations are optical: nothing was optical, and now it’s all going to be optical. Corning is a huge beneficiary.

Nothing was liquid-cooled, and now it’s all going to be liquid-cooled. Jeff Markley told me he bought 1 million valves. Why did you buy 1 million valves? He said, “Well, because if water starts leaking out of a pipe, you need to isolate it quickly.” These are like $60,000 in a single 1U. You can’t have water dripping on them. But there aren’t enough valves in the world, so I bought them all.

And then there’s the underlying problem here of energy production, right? We’re about to see energy begin to spike. We’re seeing certain communities vote against opening up data centers because they don’t want them to soak up all the energy.

Are we going to get differential pricing, where data centers are paying this much per kilowatt-hour versus homeowners paying a different rate? Otherwise, we’re going to have communities basically blaming the AI tech bros for taking their jobs and hiking up the cost of electricity, and that does not bode well.

I think that the scenario where new data center deployments continue to be connected to the utility-scale grid is probably implausible at this point. There’s simply too much demand for co-located new energy output that is completely off-grid.

As long as the regulatory environment continues to be favorable—and it does continue to be—I think it’s more likely we end up in a future that looks like Colossus, where there are co-located natural-gas and, soon, SMR and fusion plants in a few years, and all of that is, by default, disconnected from the broader utility-scale grid.

Yeah, I think you had Jeff Bezos, who’s a pretty smart guy, saying we will have gigawatt data centers in space. When you actually do the math, it makes sense in a few years when you look at payload costs, chip costs, and, again, power with solar. That’s just something that’s crazy, but again, it shows the demand for these things.

Eric Schmidt has a great deal of interest in that vision as well.

All right, I’m going to move us on to our last conversation topic for today, keeping this WTF episode sharp and fun, and that’s robotics and the release of FSD 14.1. Elon has released, as promised, something that’s got 10 times more AI parameters.

I love this: navigation and routing are now handled by Tesla’s neural net, which can help you find detours and handle unexpected obstacles. I’ve had my Tesla drive me into situations that I shouldn’t have gone into, robotaxi-style, upon arrival. You can now select precisely your arrival option—where you want to park: in the street, in a garage, or curbside. In Elon’s words, V14 feels alive.

Of course, this is just the prelude to the entire automation of driving across every sector. Who wants to jump in here?

I’ll tell you one thing that’s new: with the big screen and FSD, you can watch the podcast on-screen safely while driving, so you don’t have to have Peter describe every video to you.

Yeah, I think this promises to be a big jump over 13.2.9. I think, aspirationally, it also represents the beginning of several different forms of convergence: the convergence between, obviously, robotaxi tech stacks and human-driven or supervised-driven autonomy tech stacks.

Less obviously, I think we’re going to see, over the next few years—maybe 2 to 3 years—a sequence of subsequent convergences. I would expect to see, for example, the Optimus tech stack converge with FSD, maybe in some future version.

At the core, I think what we’re seeing is the emergence of a vision-language-action model, a VLA model, from Tesla that’s just end-to-end embodiment. It works in cars; hopefully, it works in Optimus robots as well. I would expect to see, from all of the other major frontier labs, singular, consolidated VLA models that work across a variety of different embodiments.

Amazing. Speaking about VLA models, out of Google, we’re seeing the next generation of physical agents: Gemini Robotics-ER 1.5. Let’s play a little video. If you’re watching this on YouTube, you can see the model identifying everything on your desk. It helps robots think through complex real-world tasks. It reasons like a human and outperforms GPT-5 on embodied reasoning and pointing accuracy. Emad, you want to comment on this one?

Yeah, I mean, what are the brains of robots? These joint vision-language models, which can basically think and reason.

The crazy thing about this is that to do this a couple of years ago, you really needed very high-performance chips that used 1,000 watts of electricity. If you look at how efficient models like this are, you can extrapolate out. You can see they're actually going to be possible on edge compute, which just opens up the opportunity so much.

Again, I think, as Alex said, this is why you're standardizing around specific stacks, just like Dojo was stopped in favor of edge compute at X, for example. So, I think we'll see these very specialist chips and these very specialist models for them, with tremendous capabilities that can then act as a basis to learn any given task effectively.

So, I mean, as everything becomes smart, everything understands its context—where it is—and you can speak to anything and have it understand what you mean. Alex, where does this go?

It gets even better than that. Just in line with what we were discussing a few minutes ago about living in the sci-fi future, you can't make this stuff up. The safety benchmark for DeepMind's Gemini Robotics VLA model is named ASIMOV.

It's a benchmark that's semi-synthetic, but it's based on lots of different vision-language-action scenarios and the relative safety thereof. The beauty is, if you actually go and read the ASIMOV paper, the Gemini team in DeepMind are benchmarking the safety of Isaac Asimov's Three Laws of Robotics against better constitutions for the safety of these embodied robotic models.

It turns out that there are, in fact, better constitutions for constitutional AI that one can come up with beyond those three laws. But the very fact that we're now at a point in our future history where we're benchmarking the Three Laws of Robotics against other, better models—it's amazing.

I love the group at DeepMind and Google. Thank you for what you're doing.

Well, Alex, this week we'll close our investment in—I'm not sure I can say it—Andy Systems. Who cares if that leaks out? It's an incredible company that picks through all the recycling using this exact technology you just saw in the video, pulls out the precious, rare, and valuable metals—the rare-earth metals—and then gets them back into recycling for the next generation of chips and computers, right out of WALL-E.

It shows you how this human paradise is possible, where everything can be cleaned, sorted, fixed, and repaired using this exact vision capability you saw in that video.

Love it. Love it.

Alex, your investment picks so far are still 100%. So, add this to the—

No investment advice from me.

These are private. That's okay.

I'm going to show this video of Tesla's Optimus learning kung fu just because it's so cool. Let's take a quick look. Now, if you're listening rather than watching on YouTube, we just saw Optimus with a kung fu sparring partner making some impressive moves.

I've got to imagine that, for it to actually be impressive, that was not a human controlling Optimus; that was its AI model operating in the world. Does anybody have any countervailing evidence of that? Elon has actually said in connection with this that it was autonomous. It was not teleoperated.

Fantastic. We've seen our friends from Unitree doing impressive work, but Optimus towers over the G1 from Unitree. We're not too far from mech bots fighting in the ring, trained by imitation learning.

We're painfully close, I think, at this point, to unlocking physical labor and solving physical labor. Remember, in the services economy, approximately 2/3 of all service labor ultimately is connected to some sort of physical task. Think of how, in the future, so many tasks that no human would ever want to perform, for which there aren't even any jobs, can just be automated.

Yeah, I think the fact that it's all neural-network-based and imitation learning is a really important point, because people who've been working in robotics—you know, I had dinner with the founder of iRobot, and he's all cynical about robots: “Robots are slow.” It's just not true, because it's all neural-network-driven now.

The pace of development, the smoothness of the movement, and the dexterity are going to skyrocket because it's all neural-net-based. I need to jump shortly, but some closing thoughts here, pal. It's, again, the most exciting sci-fi times. Learning kung fu is going to be like a couple of megabytes, and to do any task it's probably not going to be more than another couple.

But, Emad, once again, you're going to have to wait for the Neuralink to get better for that. I think, again, we're just at this tipping point, and the tipping point is in the next 6 months across just about all of these.

We're going to need these capabilities for data-center construction. If we're going to achieve 250 gigawatts by the early 2030s, we may not have the human labor to accomplish that.

As you know, Peter, I'm always looking for what the innermost loop of the tech tree is. In this case, mixing metaphors, it increasingly, to me at least, looks like the innermost loop is going to look something like recursive self-improvement: robots building robots first, which are then building data centers, which are then putting out digital superintelligence to increase the efficiency of the materials that the robots are built out of and the efficiency of the energy used to put into data centers.

It's hyper-exponential. I can feel the singularity coming.

You're feeling the AGI.

I'm feeling the ASI. Oh my God.

Right. Yeah, Dave, thoughts to close us out.

Well, my final thought: tomorrow's my 25th wedding anniversary. When Mora sees this podcast, she'll see I bought 2 tickets to Bermuda for the weekend, so we're going to spend a ton of money, and she should see that. It'll be concurrent with the podcast. So, go ahead and open it.

Yeah. And my question ultimately is, as we hit longevity escape velocity, does “till death do us part” hold out for hundreds of years? We're going to find out.

That'd be awesome. That would be awesome.

So, I went to Tiffany's and bought something. I swear it's made of vibranium and set with Infinity Stones, given the pricing, but dealing with people in a physical store is the worst form of torture for me that I can possibly endure. So, that's the real gift.

My God, amazing.

Emad and Alex, grateful for your brilliance as always, and see you guys next time. I'm super pumped by the speed of these breakthroughs. I mean, I don't know how you asymptotically approach infinity, but we're going to watch it happen. All right, take care, guys.