Google I/O 2026、Karpathy 加入 Anthropic，以及 Cerebras 950亿美元 IPO｜第256期

Dave

If you had said 5 years ago, “Hey, Google’s going to 6X its CapEx and the stock will go up,” nobody in their right mind would have said that was even possible.

Quadrillions, billions, hundreds of billions, trillions. It gets numbing after a while.

David Blumberg

There was a lot of conversation that Google was cooked, Google was not going to make it, and that its revenue engine was being massively disrupted. And here they are, sort of disrupting the disruptors.

Andrej Karpathy joins Anthropic. He was a co-founder of OpenAI. He left in 2017 to run full self-driving for Elon. He’ll start a new initiative focused on using Claude to accelerate Claude’s own pre-training research.

Cerebras’s record IPO closes up 68%, with a market cap of $95 billion. Andrew Feldman is the CEO of Cerebras.

It’s like a lifetime achievement, kind of like a Nobel Prize or an Olympic gold medal, where you carry it for the rest of your life.

All right, I see Andrew Feldman has entered the room. Andrew, a pleasure to have you here.

Thank you for having me on your show. I appreciate it.

Now that’s a moonshot, ladies and gentlemen.

Today, we have our extraordinary group of Moonshot mates: David Blumberg, our emperor of AI investing; Salim Ismail, professor of all things exponential organizations; and our very own artificial superintelligence, our Moonshot mate, Ari Wallach. Gentlemen, a pleasure to have you here.

Dave

And Peter, you’re a birthday boy as well. Shall we sing “Happy Birthday to You”?

Everybody’s going to be signing off the pod right here, right now.

David Blumberg

That’s right. We are not singers. Stick to the verbal.

AWG

Well, one of us was. My first career was in the New York City Opera Company.

Really? You just keep impressing all of us. I would like to hear an aria sometime. All right, maybe I’ll do an outro.

Ari Wallach

Okay, for sure.

We had such a fun surprise birthday party for you.

Oh my God, we did. It was crazy. I have never been so surprised in my life. Honestly, you, Salim, are the most giving individual. Salim was visiting with Lily and his son Milan. We had a birthday dinner for him Saturday night, but his birthday was Sunday, and he walks me along the beach to a surprise birthday party where there were 50 people. I walk into this room, and I’ve never been more surprised in my life. I literally dropped to my knees in the level of surprise.

It was a perfect decoy. It was a great decoy.

It was awesome. Oh my God. Very, very special. And Dave, you sent a great message over.

David Blumberg

It was great. It was wonderful. Thank you all for that.

Our job here is to get you pumped about the future. No politics, no doomerism—just the science and technology driving us along the singularity.

Today we have a special episode: our annual recap of Google’s mega-event, Google I/O. We’re covering news about Andrej Karpathy joining Anthropic and Elon’s defeat in the trial against OpenAI. Finally, we’ll be joined by Andrew Feldman, the CEO of Cerebras, after an epic IPO.

Dave, it was a blast to be with you at Google I/O. Here’s an image of us along with Tyler Donahue. What do you think of it?

Dave

Yeah, it’s great to be where it all started. The vibe on AI is global, but within this epicenter where everything began, it’s off the charts. The first hour alone had an incredible amount of stuff to talk about compared to a year ago or 2 years ago. We’re taking it all for granted, but it’s just crazy—the number of new Google brands and new AI products is pretty baffling.

We’ll go through it all. We’ve got it all beautifully cut up today, so you can analyze every piece of it.

And Alex, you were watching online, weren’t you?

Alex

I was. I was watching in real time, dissecting it all for my newsletter, The Innermost Loop. I thought there were some high points, some low points, and some midpoints, if I may say, and I’m eager to dive in. And probably some no points.

And of course, where’s Waldo? Today, Salim, you’re a probability function on planet Earth. Where are you today, Salim?

I just landed in Brazil, and I’ve got a bunch of meetings and presentations here. I flew from Los Angeles with you, Peter, directly here.

Of course you did. Let it not be said that you’re not a traveling dude.

Ari Wallach

Peter, I’m at the Cerebras headquarters today, too. The vibe here—you know, the third-biggest tech IPO in history—is just epic. I couldn’t resist the opportunity to feel it right after an IPO at a company. It’s a once-in-a-lifetime kind of thing for most people, so the vibe is just epic.

Yeah, a record IPO until the next record IPO, until the next record IPO after that.

Ari Wallach

I know. It’s going to be 10X this year, so you’ve got to savor it while you have the record.

All right, let’s jump into all things Google I/O. I want to kick it off with the opening summary by Sundar. Quite the year. Let’s listen to Sundar, and then we’ll continue on and dive in.

Sundar Pichai

2 years ago, we were processing 9.7 trillion tokens a month across the surfaces. It’s a huge number. Last year at I/O, that grew to about 480 trillion tokens. Fast-forward to today, and that number has jumped 7 times to 3.2 quadrillion tokens per month.

Over 8.5 million of you are now building new apps and experiences with our models monthly, and our model APIs are now processing around 19 billion tokens per minute.

We are, of course, also seeing incredible demand across our products. We now have 13 products with over 1 billion users each. 5 of those have more than 3 billion users. AI Overviews now has over 2.5 billion monthly users, and AI Mode has been a revelation—our biggest upgrade to Search ever. People love it. In just a year, it’s already surpassed 1 billion monthly users.

Last year at I/O, the Gemini app had 400 million monthly active users. Today, we have surpassed 900 million, more than doubling in a year. And today, more than 50 billion images have been generated with our Nano Banana models.

In 2022, we were spending $31 billion annually in CapEx. This year, we expect that number to be about 6 times that—approximately $180 billion to $190 billion.

We can now seamlessly distribute training across multiple sites, scaling across more than 1 million TPUs globally. This gives us the ability to create the largest training cluster in the world. Both chips are more energy-efficient, delivering up to 2 times better performance per watt.

Wow. Quadrillions, billions, hundreds of billions, trillions. It gets numbing after a while.

David Blumberg

It does, but you have to step back. If you said 5 years ago, “Hey, Google’s going to 6X its CapEx and the stock will go up,” nobody in their right mind would have said that was even possible. And there it is.

Those TPUs are the linchpin. They’re 2X the power efficiency, but they’re in the hunt now, competing vertically from the transistor all the way through the user experience. No one else can say that. It’s just a lot.

If you look at the array of logos that have over 1 billion users, you have to rewind the video and go back and look at that again. It’s just a litany of Google things now.

Yeah, it’s relentless. I think having 900 million users is pretty incredible because that’s pretty close to ChatGPT. I found that very striking. Alex, what’s your take on these numbers?

Alex

Well, it’s inevitable in some sense. I’m reminded of a conversation I had about 20 years ago with Larry Page during his interregnum, when he wasn’t CEO. I was reminded that he was asking me for advice on how to get Google interested in spending $100 million to work on AI.

If you can believe that, it’s unconscionable by today’s standards that Google wasn’t interested 20 years ago in AI. And now here they are. It’s become the central focus of the company, full stack, from chips and data centers all the way through applications.

All of these numbers, I think, were inevitable. It’s been widely remarked that if Google hadn’t leaned in at every layer of the stack to try to own, or at least lead in, AI, they would have been toast. The original model of Google, based on search and ads, would have been cooked. So what do you do? You lean into it.

Gemini

900 million-plus users. I think that’s perhaps just slightly less remarkable than it might seem, given that Gemini is basically being swapped for Assistant. Assistant already had quite a bit of traction, but nonetheless, it’s nice to see that Gemini usage is taking off.

It’s nice, I think, on balance, to have something other than a duopoly between OpenAI and Anthropic. If Google doesn’t aggressively lean into consumer and enterprise Gemini adoption, I think that’s the default outcome.

And Alex, I do as well, because, again, I’ve known Larry since 2003 or 2004. AI was always his focus. He wanted to build an AI company from the very beginning, and he had a lot of difficulty for a while.

It wasn’t obvious to everyone else within Google for the first part of its life that AI was where this was all going. And, of course, Eric—the adult supervision in the room—came in and built the revenue engine. Google is only able to do what it can do today because of the massive revenue.

The other thing that people don’t realize is that part of Larry’s vision early on was also BCI.

Alex

He wanted to connect the brain to AI. And space stations, Peter—do you remember that whiteboard that Google used to maintain with their long-term tech tree? They were going to have a Google space station, BCIs, and all of these things. That’s all happening now. Finally, three decades later, the Google whiteboard vision is playing out.

That whiteboard actually had the tethered satellite on it. They were going to make an elevator into space.

A space elevator?

Alex Hormozi

A space elevator, yeah. They needed a carbon-nanotube wire for things to go up on. They managed to manufacture about a meter of it, for a lot of money. And that’s a 20,000-mile cable.

We’re finally catching up with the whiteboard. I was with Jack Hidary at SandboxAQ yesterday, talking about what large quantitative models—the LQMs—are going to be able to do. One of his objectives is finding new materials with the tensile strength to give you a space elevator. So, again, what do we say? What did you say, Alex?

Alex Hormozi

Over the next 10 years, every sci-fi trope, everywhere, all at once over the next 10 years. That’s the singularity.

I want to give a big shout-out and congratulations to Sundar, Sergey, Josh, and the team there. They’re hitting their numbers. Again, you have to remember that a year and a half ago, there was a lot of conversation that Google was cooked, that Google wasn’t going to make it, and that its revenue engine was being massively disrupted. Here they are, sort of disrupting the disruptors.

This is an AI-native operating system company because they’re now in a constant cycle of continuous sensing, execution, and adaptation, and they’ve kind of hit an inner loop there. This was the company that birthed the Transformer. This was the company that my friend John Smart, I think perhaps insufficiently famously, pointed out had an average Google query that was following an exponential curve in terms of the number of words, or tokens, over a period of 15 years. This was before everything hit its inflection point in 2017 or so with the Transformer.

If you look at the number of words per Google query, it was inevitably going to end up with people having full conversations with AI. Google had the exponential trajectory of user interaction, the Transformer, and the compute. It was just a matter of putting all of the institutional pieces together, and it seems like they’re finally coming together. Yeah, amazing.

Moving us along after that epic intro by Sundar, Google is launching an entirely new family of AI models called Gemini Omni. It’s capable of generating video clips from prompts that include a variety of inputs, including text, photos, videos, and audio. Google says Omni will be the “create anything from any input” product.

Let’s take a look at the video being introduced by Demis. Demis was a rock star on stage. It was so much fun to see him there.

Demis Hassabis

I’m excited to announce Gemini Omni. Models like Veo, Nano Banana, and Genie are able to create extremely realistic videos, images, and interactive simulations. It’s a step change in simulating things like kinetic energy and gravity.

Gemini’s world knowledge and reasoning really shine in Omni. It can translate complex ideas into highly accurate videos. For example, you can give it a simple prompt like, “Make a claymation explainer of protein folding,” and get this:

Proteins start as chains of amino acids. They fold into patterns like the alpha helix and flat sections called beta sheets, forming a perfect three-dimensional shape.

Omni gives you a more natural way to edit video with conversational language.

Wow. What’s really cool is that you can give it your own videos—for example, this selfie—and change reality in a really fun way.

Dave, you were going to say?

Dave Blumberg

Demis had, by far, the best visuals, and the crowd reaction was strongest. First of all, the crowd is hugely into medicine and science as the use case that everybody cares about, and Demis is the spokesperson for that. But then his visuals on the video stuff were the best, too.

The worry about AI that’s going on globally makes us lose the fantasy and the cool factor: You could look forward to this your whole life, and now you can suddenly play with it. I really encourage everybody to get in there, build some stuff, and play with it. Put yourself in a movie as a character and change the backgrounds.

Once you’ve experienced that, you really get a sense of the amazing things that are possible, starting now and for the next few years—just new things every week. It’s wild.

Alex, reality is cooked, isn’t it?

Alex Hormozi

I think reality is getting enhanced, for sure, but I also want to applaud Demis and Google DeepMind for being the only, arguably remaining, American frontier lab to still be chasing multimodality. OpenAI is in the UK, right? Well, they’re really American. They may have a lot of personnel in the UK, but it’s an American frontier lab. They’re the only frontier lab still chasing multimodality.

OpenAI cut Sora and arguably deemphasized video. Anthropic has never, arguably, been chasing multimodality. They’ve been squarely focused on CodeGen. That just leaves Google with the only credible frontier American video model, since video is arguably the hardest modality, combined with consumer demand.

Then you have all of the Chinese frontier labs. China is taking video as a modality far more seriously. My sense from some of Google DeepMind’s earliest announcements with multimodality is that they have a grand vision of modality scaling. Even though video presents as the most consumer-friendly and the most impressive demo, at the back end they’re probably treating biological sequences, like DNA or protein sequences, as another modality.

They probably have dozens of other modalities that they’re trying to fold into this omnimodal model, looking for modality scaling in a way that the other American frontier labs just aren’t. So it’s a bet. At this point, it’s almost an idiosyncratic bet that they’re going to get to some form of superintelligence that’s distinguishable because it handles all these different modalities—text, audio, video, maybe biological sequence data, maybe other crazier modalities—all in some meta-uniform way that the other labs aren’t achieving. But it’s a bet nonetheless.

Interesting. Salim, I can’t imagine a better kind of technology for education and teaching people. I so wish this had existed when I was doing organic chemistry and studying medicine. This should clear away so much of the cruft of trying to figure out how to present things and the different ways of showing things—biological models, et cetera. This could all become real-time and fully 3D. It’s incredible to see what’s going to come from this. Very exciting.

The real-time part of it is huge, too, because in a call like this, or a podcast like this, you can create real-time graphics and visuals to fit the dialogue purely with your voice. They can do it at Google. We can’t do it because we don’t have the token speed to keep up, so we have to wait a minute.

If you said something really cool right now, Salim—“Hey, Brazil, let me tell you about data center explosions in Brazil”—the graphic that backs that up would take a minute to come back. You can’t do it in real time, but they can. It’s purely a question of who has access to the compute. Yeah, it’s incredible.

But it’s going to come to a point where AI will just create a soundtrack and a visual track for your life, always present whenever you want it. Yeah, amazing.

Let’s dive into their new Gemini 3.5 Flash model. They just launched Gemini 3.5 Flash. It’s the new default for the Gemini app in AI search mode. As you’re about to hear, compared to Gemini 3.1 Pro, it’s better across almost all the benchmarks.

Importantly, Google says this new model is significantly faster—in a league of its own in terms of intelligence versus output speed. It’s better at handling agentic tasks, offering improved agentic coding and richer, more interactive graphics. Let’s take a look.

Josh Woodward

Today I’m excited to introduce Gemini 3.5 Flash, our first in a series of models. When compared to Gemini 3.1 Pro, Flash is better across the board on almost all benchmarks. It’s made huge progress in coding. And look at that extraordinary jump in GDPval, a benchmark that captures many real-world, economically valuable tasks.

Gemini 3.5 Flash is a very capable model at the frontier and comparable to the best models, but much, much faster. That’s why, when you look at intelligence versus output speed, it’s in a whole league of its own, in the top-right quadrant.

When looking at output tokens per second, it’s 4 times faster than other frontier models, and it’s an incredible delight to use.

Alex, impressive. What do you think?

Alex Hormozi

Remember when I opened by saying there were highlights, lowlights, and then, to use the colloquialism, “mid”? I would call Gemini 3.5 Flash solidly mid.

If you look at its capabilities—and others have pointed this out as well—just from a raw capability standpoint, not talking about throughput or cost, it doesn’t compare favorably with, say, GPT-5.5 High, X High, or Pro. On the other hand, this is a Flash-series model, so it’s not Pro yet. Sundar, sort of infamously at this point, has said Gemini 3.5 Pro is coming out in another month.

There were groans in the audience at the time. So, this isn't intended to be top of range. I think the strategy, if I were to play Kremlinologist here, is that Google is sort of solidly tier 1.5 at this point in the race to raw frontier capabilities.

Gemini 3.5 Flash represents perhaps pushing the optimal frontier in terms of throughput versus performance—that optimal frontier—but I think it's also very telling that Sundar is highlighting throughput versus performance instead of, like, number of tokens on the x-axis, input tokens versus performance, or rather output tokens on the x-axis versus performance on the y-axis, or some other metric. He picked the most flattering possible metric, and if you actually look—and everyone picks flattering metrics, but some flattering metrics are also more truthful than others in some global sense—it's very telling.

Gemini 3.5 Flash is being compared primarily with Gemini 3.1 Pro, less with other frontier models. Secondly, the areas and benchmarks where it's really excelling are benchmarks where tool use, in particular really aggressive tool use, is needed. So, if I had to squint at this, I would say the emphasis at Google wasn't necessarily beating the frontier with Gemini 3.5 Flash. It was probably some combination of throughput maxing and tool-use maxing—not pushing the boundaries of the frontier, but a solid release nonetheless. It's nice that Google's still in the game.

And Dave, I'm imagining you—we've talked about the labs pulling their punches. I imagine that, releasing, there'll be the next version of GPT, and then, of course, Pro will come out right after that.

Well, the scuttlebutt here in Silicon Valley is that it's a 2-horse race between OpenAI and Anthropic for the best AI in the world, and the talent is flooding into those 2 buildings in San Francisco. Nothing in that demo or in the vibe on campus at Google contradicts that.

So, you pointed out earlier—or Alex pointed out—that Google is the 1 remaining horse in the race to the consumer, and this is a very, very fast model that gives the consumer a much better experience. But the other labs have already pivoted to the enterprise and said, “Look, we're giving up on that. We're going totally after these massive enterprise budgets.” If you try to build something sophisticated with AI, you want the smartest AI that solves the problem, and you're not going to back off to a faster model that's not quite as intelligent. You just can't. So, they're going full bore after self-improvement at the other labs.

The other difference from a year ago is Google's unstoppable war chest: 180 billion in capex per year and rising. But the other guys, in the interim, raised a lot of money. OpenAI raised 120 billion in cash, and they'll burn that pretty quickly. Anthropic is on a similar trajectory now, so the war chests are actually not as different as they were a year ago.

So, yeah, I think that's the only disappointing thing in the whole show: the best of the best, Gemini, is not up there with Mythos, as far as we know. Now, I will say that Google does soft-sell. They don't announce what's coming in 4 months and promote and trumpet it because they don't need to. If something really, really big is cooking and coming soon, they didn't roll it out, but they don't need to roll it out. They'll wait until it's proven.

I love the naming nomenclature here. Of course, we've got GPT models—5.4, 5.5, 5.6—and Gemini jumps from 3.1 to 3.5. It's fascinating. Salim, what do you make of this?

I thought 1 thing that's clear is that you're seeing this kind of bifurcation now between premium cognition and ultra-cheap but very fast cognition. I think that's going to continue. I think Alex makes a great point about throughput. This will allow a lot of throughput, right? There's this continuous march for marginal intelligence; cost trends toward zero.

Yeah. Here was the next segment that I pulled out, and again, I want to give a shout-out to Gianluca, who clipped all these beautifully and provided them in record time for us for the show. A conversation about SynthID and content credentialing—really important, especially as we start to encroach on reality. How do you know if something is or is not AI-generated? It's going to become more important than ever before. Let's take a look here at Sundar talking about SynthID.

Sundar Pichai

Since launch, SynthID has now watermarked over 100 billion images and videos, along with 60,000 years of audio assets. We are now going a step further and adding content credentials verification across products. This will show you if the origin of the content was AI or a camera and if it's been edited with generative AI tools.

In this example, Gemini can tell this photo was captured with a Pixel camera and then edited with Google Photos. Of course, this only works at scale if more partners decide to watermark their own AI-generated content. NVIDIA signed on to SynthID last year, and today I'm thrilled to announce that OpenAI, Kakao, and ElevenLabs are adopting SynthID, too.

I love the fact that we're getting to standards and everybody's picking the best. Alex, how important is this?

Alex

The irony is that so many people were hand-wringing over the past few years that we won't know what's real and what isn't. My response was always that we're going to get, eventually, cryptographic chains of custody from reality capture to what is ultimately presented to the user, in the same sense that when you use a browser, you can maybe see a little lock icon to indicate end-to-end SSL encryption.

We're going to get the same thing for reality, and I view SynthID—which, by the way, was also just adopted by OpenAI. It was created by Google and now also adopted by OpenAI in the same breath—as sort of ironic. We're going to get, it seems, end-to-end authentication of realness, a proof of reality, if you will, not coming from the camera end, not coming from the reality-capture end, but coming from the synthetic end—from all vendors that want to claim credit, in some sense, that they were the ones who generated the reality.

Then the camera makers and all of the recording-device manufacturers are going to be downstream and the ones who adopt the same protocol. But either way, we're getting our end-to-end proof of reality one way or another.

I think there's a much broader story. I'd love Salim to get your thoughts on the bigger societal implications, because a big topic at Stanford last night, with Erik Brynjolfsson and his entire team there, was that the rate at which AI is innovating can't be kept up with by Congress. There's going to be no regulation of any value coming out of Washington.

So, the industry is starting to self-regulate. That's the only—you know, only AI can keep up with AI. We may look back on this moment as 1 of the first moves by the self-regulation community: “Okay, now we're going to start watermarking images. Hey, everybody in the community, please adopt our standard for watermarking images.” Then there'll be something else a week later, something else a week later, something else a week later, and that'll become the way that we govern ourselves in the future, much more so than any law coming out of Washington, purely because the pace can't keep up.

So, then, you think that's basically the first move in that direction?

I think that's exactly right. When you have intelligence becoming abundant, then scarcity goes toward trust, and scarcity creates value. We may end up at a point where authenticity is more valuable than creativity, and that line between something being created and knowing how it was created, et cetera, is blurred, is now merging because of the systems that are being created.

I think, Dave, the point you make is really, really important. Once you have that trust layer, now you can scale. I go back to Jure Miklavc, my community member, who said, “Scarcity equals abundance minus trust.” Right? If you can solve for trust, you solve for abundance.

One of the biggest challenges today—I thought this was really a big deal—is that we're moving from the information age to the verification age, and trust is becoming infrastructure. I think that's a very powerful, valuable pillar for the world going forward.

Wait, so let me get you to explain—playing arithmetic—does that mean that abundance equals scarcity plus trust?

[laughter]

It does. It does. Scarcity plus trust gets you to abundance. How about that?

All right. That's like grade 5 math. Even I can do that.

But because the cost of generating content is collapsing, right, the value shifts to signal filtering and authenticity. We saw this with photography, where the big problem in photography was, “How do I take the best photograph?” Each dollar, each click cost you a dollar, and you had a bunch of business models crop up around selling expensive cameras, offering courses on photography, and publishing books on composition.

Then we moved to digital photography. The cost of creating a photograph went to zero, and now the big problem everybody has in photography is, “I have 6 copies of my photographs on 7 different online services, and you can't find anything.” The value then comes in that filtering system, right?

Salim, we're moving into this intentional world that we design. Everyone's like, “What will happen next? What will happen next?” Well, whatever we design is what's going to happen next.

But Dario and Demis are 2 probably dominant architects in the future of how we live. It's just great to hear both guys go back and forth. But if you do a raw word count from Dario Amodei, CEO of Anthropic, and go back, the words were all about transformer architecture, speeds, intelligence, and benchmarks. Then they transitioned to UBI and ethics, and now they're talking about the way the world should be governed going forward and writing papers on it.

If you just track the word count, it's also on this exponential change rate. Same with Demis. Demis has to be a little more cautious because, technically, he's an employee of Google, even though he acts very independently. But those are the 2 guys very much determining the future of all humanity right now.

This watermarking is just move 1, act 1 of the whole future of the way we live. Look at what's happening, right? Our trust in legacy institutions is collapsing, and at the same time, AI is building up the capability, infrastructure, and foundation for delivering trust. So, hopefully, if that happens elegantly, we'll have an elegant shift from scarcity to abundance rather than a messy one.

Everybody, you may not know this, but I have an incredible research team. And every week myself, my research team study the the trends that are impacting the world. Topics like computation, sensors, networks, AI, robotics, 3D printing, synthetic biology. And these meta trend reports I put out once a week enable you to see the future 10 years ahead of anybody else. If you'd like to get access to the meta trends newsletter every week, go to dmandis.com/metatrends. That's dmandis.com/metatrends.

The next product that they dove into, and a central part of Google's plans, is Antigravity. They released Antigravity 2.0, a standalone desktop app built to orchestrate multiple agents to execute tasks in parallel. Let's take a listen, and then, Alex, I'm coming to you for your evaluation: mid-tier, high-tier, no tier?

Josh Woodward

At the core is Antigravity 2.0, a new standalone desktop application that delivers fully on that original glimpse of a truly agent-optimized experience. The new Antigravity is unabashedly agent-first, focusing on the core agent conversations, agent-produced artifacts, and multi-agent orchestration. Like I said, unabashedly agent-first.

As Sundar mentioned, this is the exact experience teams here at Google have been using to drive massive value. Let's take this live and actually show this operating system in action. Try running Doom right now. It just doesn't work. It turns out that the OS is currently missing some necessary video and keyboard drivers, so let's just try and fix it in the new Antigravity.

I have a prompt prepared. I'm going to paste it in. Antigravity ended up doing a whole host of research, writing over 100 lines of code, and then finally building the operating system. Let's take a peek and see if it works. Amazing.

Alex

So, first off, what is Antigravity 2.0, and what are you thinking about it? Let's remember where Antigravity came from. Do you remember Google's acquisition of Windsurf during that debacle? Antigravity is basically Windsurf rebranded from the Windsurf team that was hack-a-hired by Google.

As for Antigravity 1.0 versus 2.0, you were asking earlier, Peter, is this high, mid, or low? This is sort of mid, in my mind. If you look at what Cursor has been doing by contrast, Cursor was much more aggressively leaning from their old interface, which was sort of a reskinned Visual Studio Code-centric editor, to their more recent interface, which is agent-first.

I view this almost as a copycat, fast follow, slow follow, or somewhere in between from the Windsurf team within Google DeepMind. They're basically following the same metaphor of saying, "No, we're no longer about direct code-editing access. Now, the primary metaphor is orchestrating fleets of code agents that are doing all of the hard work."

I would say Google's almost hamstringing themselves a little bit by announcing this now as part of I/O versus, say, in a more timely, fast follow, or even leading when Cursor was doing this months ago. I've used Antigravity quite a bit. I was certainly using it even more when Google first announced it after the Windsurf acquisition.

I would say it was not super impressive. It was very buggy. I think 2.0—I haven't had a chance to use 2.0 yet—but hopefully it's a good deal stronger. But really, I don't know anyone who's doing their primary development work with Antigravity at this point. The development is happening either with Claude Code, Codex, or maybe Cursor. Antigravity—I don't know anyone who's using it.

You know, everybody except every engineer at Google. Leapfrog Google. So, Dave, you were trying to get on what?

Dave Shapiro

Even that's not true, though. It's been publicly reported that within Google DeepMind, they're all desperate to get Claude Code access for everything.

We talked about that a couple of pods ago. Dave, you were playing with this during Google I/O yesterday.

Dave Shapiro

Yeah, Tyler and I both installed it in real time as they were rolling it out, which maybe wasn't the smartest move in the world because there were 16,000 people behind us in the crowd, probably all trying to do the same thing. That was not a great first experience, but I don't blame Google for that.

But this morning, it worked fine. It installed great, and I completely agree with Alex's assessment. It looks almost identical to the new Cursor agent-first interface. I almost can't tell where I am. Am I in Cursor, or am I in Antigravity?

What they did do, which is a little more extreme than Cursor, is it completely replaces Antigravity 1.0 so that you can't even see the code anymore. You have to go and launch the old thing if you want to actually edit code. Cursor didn't go quite that far, but it's really obvious where the puck is going.

If you want to build things in the future, you're not even going to look at code. You're going to describe what you want, and you're going to debug at this much higher level of evals and functional comparisons: "I don't like where that button is. Move it." I think, in the future, nobody's going to want to go back to an autocomplete code-editor view.

They leapt ahead and said, "We're just going to eliminate that entirely. If you really want to hack, we'll give you a way to get back to it. But we're going where the puck is going, and not where it was." It really is exactly catch-up, like Alex said. Everybody's got the same thing—Codex, Claude Code, and Antigravity—and they're just going to be leapfrogging each other.

I'm just curious if there's going to be some new sort of breakout approach to this that's going to materialize. Alex, do you think there's anything in the future?

Alex Finn

Code is clearly going away as a human endeavor. It's all being abstracted away by code agents that handle all the code, and humans may not be trustworthy enough in the near-term future to even be allowed to write their own code.

I think recursive self-improvement is another arrow of time. Maybe older-generation models are trustworthy enough to rewrite themselves and generate newer, better models, but code's going away. I think that's the obvious trend here.

Dave Shapiro

I think also, Peter, to answer your question on the next paradigm, we've only had this paradigm for a couple of months. Let's let it settle in for a little bit.

But no, clearly the next paradigm is exactly the Star Trek holodeck, which Alex has been saying for a while. Right now, you're talking to it, it's building things for you, and it's incredible, but it's not natively graphical and visual, and you're not moving things with your hands.

If you say, "I want to move that button. I want to change this. I want to connect this to my..." you're not actually seeing the button move in real time. It's regenerating, and then you see a new rendering. In the future, it'll be a real-time graphical experience that's interacting in your comfortable physical space, kind of a native human environment.

That next iteration is certainly within this calendar year.

All right. Next up is Gemini Spark. It's Google's take on OpenClaw, I think, is the most obvious thing to say. It's a new, always-on AI agent that can write emails, create study guides, and keep an eye out for financial fees that you're being charged. It's Google's "We have OpenClaw at home" moment. It's powered by Gemini 3.5 Flash and offers 24/7 operation. Let's take a listen to the conversation about Gemini Spark.

Josh Woodward

Introducing Gemini Spark, taking action on your behalf and under your direction. It runs on dedicated virtual machines on Google Cloud, and it's 24/7.

Here's a task right off the bat. This is a pretty straightforward example, but it's so useful: "Help me draft an email to the team, compile everything about our recent Gemini Live launches and wins from the last week."

What's amazing here is that Spark will go through it step by step. Look at all these steps and all the time it saves you. Again, it works across the various skills and apps that you have.

What's really amazing is that it'll break it down and also be able to generate files for you. The first one here is a live RSVP tracker right in Google Sheets. You can see that it shows who's confirmed and who hasn't.

What's amazing about this is that it'll actually update because it's connected to Gmail. So, when L. Thompson in row 8 RSVPs, it'll update, which is pretty amazing.

One of the things I find fascinating is that the integration across all of the Google products is very powerful. There’s a point at which there’s such a cost to not being inside the Google ecosystem that everybody defaults to it.

This is the most ironic thing I’ve ever heard because people who are younger don’t remember that Google only exists because the FTC stopped Microsoft from killing it. Microsoft had just killed Netscape, taken total control of the browser, integrated it with the operating system, and made it impossible to do anything on the internet unless you went through Microsoft.

That triggered the FTC, and Mike Kirschner came in. The whole lawsuit stopped Microsoft cold in its tracks, and they paid a $1 fine. I don’t know if everyone remembers that. It was a hilarious outcome, but they had to unbundle, and that opened the door for Google to come into existence.

Microsoft pseudo-competed with Bing, but they were prevented from competing aggressively and tying it back to the operating system. So, here we are all these years later, and Google is coming out with a series of kind of exact copies of Cursor. But it’s perfectly integrated with these other products. You saw on the other slide what, a dozen Google products that have over a billion users? A billion users out of the world’s population. It’s a massive installed base.

If you want OpenClaw, you can be over there. But if you want an OpenClaw equivalent that works natively with Google Docs, Gmail, Android, everything else, and your Google Pixel camera, then you have to use this. So, it’s exactly history replaying itself.

It’s so ironic because they were so anti-Microsoft back then. The whole “Don’t be evil” motto was a direct attack on Microsoft, implying they were the big guys that were evil. Here we are years later. I’m not saying Google’s evil in any way. I’m saying they’re using tying as their competitive advantage in the exact same way that Microsoft used to.

Alex Finn

I think it’s a lazy copycat product. It’s obviously Google trying to take advantage of the resources that they have. Note that it’s hosted in a GCP VM, not necessarily pushed all the way to the edge, although they have aspirations tied to Gemini and Android Halo for that.

But if you’re Google and you see OpenClaw, and you see Jensen out there saying OpenClaw is the next big ChatGPT, really, what’s the smallest, what’s the minimum viable response that you could take? It’s, “Okay, we’re going to host GCP VMs with Gemini Flash that integrate together all of our products and run headlessly.” That’s the sort of minimum viable strategic response.

What I would have liked to see from Google DeepMind here was the maximum response. Show us the art of the possible. Show us what a next-generation OpenClaw or Hermes competitor actually looks like. Create the benchmarks. Show us next-generation capabilities. They didn’t deliver that here.

Well, Alex, just to be fair, they’re delivering on a lot. It’s not just one piece, right? It’s a lot that’s being deployed on Google I/O day. But having said that, I still love Skippy, which is an OpenClaw on top of my Mac Studios. I love it because it’s got a personality versus being a generic, ever-present agent.

I don’t know if you can do that with Gemini Spark, but I think the personality side of these is critically important. Salim, any thoughts?

Two thoughts. One is, I agree with Alex. They really could have gone for a little bit more bite here. But on the other hand, when you can make agents generally available to the average Google user, there’ll be hundreds of millions more people training up agents. I think that’s generally just good.

OpenClaw still has lots of room to be experimental, power-user-oriented, and very opinionated, doing the weird things like the camera stuff. But I think this is a very solid entry into that world to give people a taste of what an agent world could look like.

But it is boring. It’s safe. It’s playing it safe. Is it a safe entry? Yes, it’s a safe entry. Will a lot of people maybe use this to clean up their Gmail inbox? Yeah, probably. But it’s not pushing the frontier, which is really what I would have loved to have seen here.

Yeah, I think that’s your recurring theme on a lot of these, Alex. Is that true?

Alex

I think, look, as an accelerationist, yes, I’d love to see frontier labs pushing the frontier. To the extent that this is an avatar of Google DeepMind and not just Google corporate, I would love to see more frontier coming out of the frontier.

Dave, you want to close this out?

Dave

Two quick thoughts. There’s something very powerful happening here because this is giving everybody an operating system for their lives because of the deep integration with all the other Google stuff. I think the next productivity jump is going to come from persistence, and this will create a massive enabler across the board.

I want to go back to the earlier comment and double down on that, which is that this will train a lot of people on how to build agents and run agents. I think that’s going to enable another class of things to come forward from that.

Yeah, so there’s no doubt that this is all fast-follower, exactly the way you’re characterizing it.

Alex

On the other hand, Peter, you love your Skippy. I love my agents that I set up, too. But when you talk to somebody on the street and you say, “Hey, have you set up an OpenClaw or a Hermes?” overwhelmingly, across the world, people say, “No, I haven’t done that.”

That install and onboarding experience is just too much friction. So, I wouldn’t underestimate the power of default behavior. Over half the world uses Google, and if Google says, “Okay, Gemini Spark is going to be one click away from a Google search,” it’s completely integrated. A massive fraction of the world is just going to click the button, and then their first experience with a personalized agent will be via that click.

I don’t think anything can slow down Google because of its massive distribution advantage. They don’t have to push the outer boundary. They can afford to be a fast follower. They just need to be as good one day later and integrated with Chrome, integrated with Google Search, integrated with Android, and they will win.

I’m equally disappointed, Alex. I’m not saying it isn’t disappointing. But from a strategy point of view, they don’t need those risks. They just need to be as good one day later, integrated with Chrome, integrated with Google Search, integrated with Android, and they will win.

I think Google’s magic potion is making it user-friendly, making it easy, and making it intuitive. I think they’re going to deliver with Gemini Spark on that particular promise.

All right, here’s another part of Google’s resurrection and dominance: the agentic powering of AI Mode. AI everywhere. Remember the conversation we had—search is dead? Search is not dead; it’s just been reinvented. Let’s take a listen.

Guest

I’m excited to announce we’re launching a brand-new intelligent search box. Before, the search box was a contained space, but now it’s totally reimagined with AI. It expands with your curiosity, and as you ask, search helps you formulate your questions with AI-powered suggestions.

This goes beyond autocomplete. It offers nuances that you might not have even thought to add. Now we’re taking an exciting step toward this vision. We’ll be able to create and manage multiple AI agents for your many tasks right in Search.

Now, let’s say you’re apartment hunting. You can do a total brain dump of what you’re looking for with all your criteria, like location, natural light, and availability. Your agent will continuously scan the entire web across sites, social media, and forums.

Persistent search here, right? This is your agent. Whenever you’ve asked a question, it is going to persistently be looking for the latest and greatest. Yes, this new apartment just became available. This product just got cheaper. Your wife loves this topic, and here is a new product delivered to her.

The other side, though, is that autocomplete function. I wonder where it’s going to take us, right? You’re going in asking, or thinking about asking, one question, and of course Google can sort of drift you into asking a different question you didn’t intend to ask. A lot of interesting perturbations here.

Oh my God, yeah. We’ll think about a vacation plan. We’re like, “I really think I should go to Barbados,” and it autocompletes to Bermuda.

Dave

And you’re rerouted to a different hotel. The revenue power of that is astounding. AI rolled out their first ads, and a lot of the companies I know have adopted it, but it’s very ham-fisted. It’s like, “Here are some ads on the side.” They’re obviously ads.

The Google version of it has to preserve $200 billion of existing, 90%-margin revenue. They haven’t quite figured out how they’re going to serve that, but their power over user decision-making is like nothing we’ve ever seen before. I’m sure they’ll find a way.

Google Ads is now going to gently drift you toward a different question that you weren’t there asking.

Dave

Absolutely. That would be amazing.

I mean, remember Google Instant as well, which also offered relatively fast suggestions. I don’t think Google ended up directly monetizing that, but Google does, to Dave’s point, have a long history of steering users toward more profitable queries. I think that’s probably quite likely.

Alex

What I probably underline here is that the shape of the rectangle changed after decades. How big a deal is that? After so many armchair commentators were saying that Google was about to be disrupted by ChatGPT with web search, it turns out Google is able to self-disrupt and change the shape of that multidecade-old rectangle. It changed the shape in the direction of building AI modalities, AI search, natively into its search experience, which I think many people were scared wasn't going to happen. They did it in the end.

You guys, my very first venture investment ever was TripAdvisor, back when it was first starting. The big quandary at TripAdvisor was, how are we going to have completely unbiased, accurate reviews and still get paid by the hotels? We need to make money somehow. How's this going to work? It turned out that just by sorting the list, the human default behavior is so dominant that they'll go over one—

Guest

You mean laziness is so dominant.

Yeah, but we're buried in decisions now. We have so many things coming at us from so many different directions that we have to be lazy. Only Alex could actually study every single pathway and make an optimal choice. Everyone else, you just have to fall into the default buckets once in a while.

So 80% of people will click on one of the first 2 or 3 hotels. You can have perfectly accurate reviews and just re-sort the list so the ones that are paying you are at the top. Then you have your cake and eat it, too. I think that default behavior will hugely benefit Google because they will steer the users, but they don't have to be super overt. They're not going to misguide you into some fraudulent products. They're allergic to that like crazy. But people will still follow the default suggestions from Gemini, and then Google will collect the revenue from whoever is willing to pay.

Salim?

Guest 2

I just love the fact that they have the courage to risk disrupting their own business. I think that's such a hard thing organizationally to do, and I have to give them full props for going after it.

Yeah, and for everybody listening to this, our goal here is to give you an overview of what Google has just done. It's so dominant on the planet. It does steer a lot of humanity's abilities. I hope folks are enjoying the summary. Please dive into these. Your mindset of curiosity is your single greatest tool, so go and play with these things. When you finish listening to this podcast, go and jump onto Google and play with the new AI search or its capabilities.

Guest

One more note, Peter, if I may, on Google's self-disruption via search. I think there's a misconception out there that the main obstacle to Google self-disrupting its search with so-called modern AI was somehow on the business side—the business risk or the advertising side. I think, actually, the main obstacle was more technical: Google engineers, for a couple of years there, were concerned that there wasn't a cost-effective or time-effective way to squeeze generative models into the very narrow, latency-sensitive, and cost-sensitive parameters of powering a search. The models were too expensive and too slow to yield search results that would be competitive.

This is, I suspect, one of the reasons why you see, going back to Gemini 3.5 Flash, so much emphasis on throughput. It's reflecting Google's own internal dogfooding needs of having ultra-high-throughput models that they could use to power search and some of their own internal applications—needs that maybe OpenAI and Anthropic aren't feeling as much.

Makes sense. Yeah, makes sense. All right, next subject is Google launching a universal cart that users can add products to from YouTube, from search, from Gemini, and from Gmail. Google says this intelligent shopping cart works across a multitude of different merchants, from Nike and Target to Walmart and Shopify.

You could literally add a product when you're searching on a Nike site or a Target site and then have it monitored and bought at the same time. Let's take a look. Again, this is part of Google's incredible revenue engine. Okay, journey.

Josh Woodward

I am excited to announce the Universal Cart, a truly intelligent shopping cart. It works across merchants and across services. You'll be able to add things to your cart when you're browsing search, chatting with Gemini, watching YouTube, or even reading your Gmail.

The moment you add a product, your cart goes to work for you in the background. It finds deals, looks at price drops, gives you insights on the price history, and alerts you when something comes back in stock. So, reinventing the shopping experience.

I've got some comments, but I'd hear from you guys first. Elephant in the room. Yes, Alex.

Alex

The elephant is Amazon. I look at every announcement relating to shopping from Google through the lens of, how are they going to compete with Amazon for retail e-shopping? Whether it's trying to commodify, create virtual storefronts for individual retail vendors, crawling third-party e-commerce websites and assembling virtual pages, or now universal carts, this is all through the lens of how they're going to compete with Amazon.

I think the elephant in the room here is, is Amazon even going to contemplate going anywhere near complying with or adopting Google standards? I guess it's not.

Well, the follow-on here is that we're going to be seeing, in a few moments, the reinvention of Google Glass, where you've got imaging capability. We're probably going to see the next invention of shopping, where shopping is always on.

Wherever you're looking and you see something, your AI agent realizes, “Oh, I'm focusing on Alex's beautiful orchid in the background.” I mean, that is—that orchid real, Alex? I just need to ask.

Alex

I thought, Peter, you said reality was cooked, so you tell me.

Okay. But literally, when I look at something, my AI agent will say, “Oh, you're focusing on that. Do you want to purchase it?” Or as you're walking through the day, instead of shopping becoming something you do for an instant of time, it's a continuous function.

The Universal Cart is aggregating all the things and then probably, at the end of the day, saying, “Hey, do you want to purchase that? Just say yes.” We're going to be seeing this as an early step, but not the full instantiation of reinventing shopping. Salim, you're going to say?

Yeah, so today we go from human to website to shopping cart to checkout, right? Tomorrow we're going to go from intent to agent to transaction.

Yes. I think every CMO in the world, going forward, is going to be asking, “How do I convince 100 million agents to choose my product?” They're going to have to market to the agents, right? I think this is powerful.

Look, Google helped create a trillion-dollar company by helping people use search. Somebody's going to create a trillion-dollar company by helping agents buy. But are you ever going to market to the agent? My agent knows what I want, knows my genetics, and knows my taste from everything else. It may just be buying stuff for me all the time that could be returnable—sort of surprise and delight.

Something shows up on the front doorstep: “Oh, I thought you'd like this. Here it is. If you don't want it, I'll have it picked up and returned.”

Alex Hormozi

We'll definitely go to that. The disruption to e-commerce is not better shopping; it's getting rid of shopping. Yes, right? There's this ambient experience where, hey, your shoes may look a bit dirty from the camera I looked at—from your doorway camera. I'm shipping you new shoes. It'll be that kind of thing. Dave?

Dave

I think it's amazingly shocking, the degree to which the big guys don't care anymore about consumer shopping. Amazon was nothing but consumer shopping originally and built their entire empire on consumer shopping. Then they added AWS.

AWS is so much bigger at Amazon now than the entire Amazon we know—all of the shopping. Google was already competing with that side of Amazon with GCP versus AWS. They're already in this battle royale over compute and data centers, enterprise use, and everything.

Google had already tried to compete in retail with Froogle. Remember Froogle?

Yes, of course. Yeah, Froogle, right before Shopping. Yeah.

David Friedberg

Yet another rebrand. So I think this is another attempt to take Amazon head-on on the shopping side. But I think AI is a big game changer. It doesn't matter too much whether Amazon defends its turf or whether Google encroaches and wins.

At the end of the day, the battle on the cloud and the back end is so much bigger, and that's already raging. So this is kind of cool. I think it's just the next stage of this trillion-dollar retail battle that will rage on for a while.

All right, let's jump into a conversation about the Gemini app and NotebookLM. Here's Josh Woodward, who heads Gemini. I think we're going to get Josh on the pod here. We'll talk to him about what he's up to. Let's take a listen.

Josh Woodward

More than 900 million users are coming to the Gemini app every month. On its own, NotebookLM has now been used to create more than 1.5 billion notebooks, podcasts, slide decks, and more. It's now available in more than 230 countries and over 70 languages. It now opens up immediately and inline. Soon you'll be able to pick a regional dialect that resonates with you.

You've got a right good mix of different accents knocking about, like this one from Liverpool. Madi Bharatwaz Gemini Omni is coming right into the Gemini app.

Let's look at how this plays out in the real world. I want you to meet Sashu. She's working on a new song, and she wants to create a quick video teaser. So she shares the raw video and adds some reference visuals to it. Let's take a look at what it looks like.

The third update today is about how agents are coming to Gemini. One of our newest out-of-the-box agents is called the Daily Brief. It's a personalized digest that's designed to be your first stop every morning. Here's how it works.

You can see here that it's synthesizing information from across my inbox. And with this travel info, I can just take the next step right in line.

So this is an integration story. This is Google integrating across all of its capabilities and making it so magical that you can't afford not to be in the Google ecosystem. Dave, what do you think about this?

David Friedberg

It's amazing when you've got one guy on stage demonstrating, "Here, we can build an entire operating system in real time. Let's go. I'm consuming a trillion tokens right now building," and nobody gets it, right? Nobody can relate to building an operating system.

Then Josh gets up there and says, "Here's a real human musician, and here's her trying to portray herself to her fan base." The crowd goes crazy. It just shows you the human aspect of this is so dominant, even in Google, even within the empire. The human aspect of it is so dominant in people's minds.

It doesn't come through on the videocast when 20,000 people see Josh present something and they go, "Wow! Oh my God, I can totally—" The vibe is contagious across the whole crowd, and it's just hard to capture that in a video clip. It's remarkable, and it's going to unleash so much creativity. It's such an exciting time.

I wish that was the only vibe—that everyone could just capture it and then hold it and bottle it in. But this was just a great moment.

Alex, your take: Why is Google still branding NotebookLM as NotebookLM? It should have been folded directly into Gemini, or maybe Google Workspace, or something else. Why does this still have an independent brand? I don't understand it.

Alex Hormozi

Well, look at all the others. It's like Spark and Flash and Antigravity, and all this is really fragmented all over the map, with divisional kind of branding.

Google has this reputation, for better or for worse, of launching lots of products and having a culture where product managers get promotions for launching but not maintaining products. I really love, in the spirit of more wood behind fewer arrows, seeing all of this functionality unified in a way that gets sustained.

Yeah, especially because AI is such a unifying force. You can put one voice and interface on top of all of this mess.

And Google's branding originally was so good. You know, all the other search engines—it was so colorful and humanistic and friendly and all of that. So, speaking directly to Josh and the Google team, please just unify all of these offerings and maintain them. Don't keep launching 10 different products and product names that we'll forget about a few months from now. Just please unify all of them and maintain them.

You know, the idea of a Daily Brief—I love it. Skippy gives me a daily brief. This is a beautiful integration here. Being able to know what you're doing, when you're doing it, what your intention is, and giving you updates all the time on your flight: There's a new flight that's available now; it's 5 times cheaper, or whatever the case might be. And the weather is going to be hotter than expected, so make sure to pack differently. That level of overlay intelligence is going to be magical.

Alex Hormozi

It is, but at the same time, remember Google Reader, Google's RSS reader that Google abandoned despite having a rabid user base, myself included? I know Google really wants to own the news feed. That much is obvious, but please just maintain it.

Okay, Salim.

2 thoughts. One is, I've always thought about NotebookLM, Alex's earlier comments, as this weird thing sitting out there because it incorporates presentation, learning, and interaction. What's the difference between doing this in the Gemini app and doing it in Spark? People are going to create a lot of confusion around this.

The point that Alex made, I just want to double down on, is that in any big company—we had this at Yahoo as well—you're rewarded for getting something out there, but then you've got a strategic project manager looking across resource allocation about 10 different projects. And so you get a peanut butter problem where you're very thin across all the different projects. You don't iterate very well.

One of the few companies that iterates very well is Apple. They will relentlessly iterate on their products. Most other companies don't iterate very well.

There's a whole thing written by Brad Garlinghouse called the Peanut Butter Manifesto from when I was at Yahoo that mirrors this old challenge: How do you navigate this when resources are limited? They're not run as startups, with each individual team doing startups with KPIs of their own and targets, et cetera. They're run in hierarchical structures in many cases, and you suffer a lot from that.

All right, Google's new product called audio glasses. We heard earlier about their partnership with XREAL. Here's a partnership with Samsung and a couple of different glass manufacturers. Let's take a look at how this is going to impact our lives. 2 videos to show, then we'll discuss them.

Shahram Izadi

The next big milestone for Android XR is intelligent eyewear. Today, I'm excited to announce that our first audio glasses will arrive this fall. They are designed to give you all-day help with Gemini that's spoken into your ear privately rather than shown on a display.

These glasses let you stay hands-free and heads-up for things like listening to music, taking photos, making calls, or tapping into your phone apps, all without reaching for your pocket.

Won-Joon Choi

At Samsung, our vision is to enrich people's lives and help shape how we live tomorrow. In close partnership with Google, we're introducing intelligent eyewear that empowers you to connect to the world with confidence.

Built with Samsung's precise engineering and craftsmanship, we're merging form, function, and helpful intelligence to create something you'll want to wear. In eyewear, every millimeter counts.

Today, we're thrilled to share a first look at our coming styles, co-created with our eyewear partners, Orby Parker and Gentle Monster.

All right, the elephant in the room here: It's got forward-looking cameras, but you're not seeing words or images on the screen. You're being spoken to by your AI. Interestingly enough, I think that being present in life has just been cooked as well.

Imagine you're walking around and you're not talking to your wife, your girlfriend, or your kids. You're just having the agent whisper to you all the time. Salim, what do you make of this?

A bunch of things. I was really disappointed that there's no visual on the screen. I mean, doing the audio might be, in Alex's words, very mid.

But we're moving to that point where human-computer interaction becomes continuous and becomes an ambient layer that's just ongoing. I think that is the bigger story here, because that will just continue to play out as we merge with technology.

Already, we pick up our phones 80,000 times a day. This just continues that in a very unnoticeable way, and the form factor is very workable.

Google should have owned smart glasses. Instead, Meta is running away with this space. Apple is also playing catch-up. Where was Google?

Google had it. I was one of the earliest users of Google Glass. Remember that?

Yes, the Glassholes of the world. Did you get punched?

I did not, fortunately, but they had a battery life of 5 minutes, and they self-bricked through operating system updates. It was, I think even Google would recognize, prematurely released.

Google could have kept iterating—to this earlier point about doubling down—on smart glasses from the Google Glass era, and they didn't. They basically abandoned Google Glass to enterprise and then abandoned them completely. Now this, I think, represents a complete reset, except without all of the conveniences that Google Glass had.

Meanwhile, Meta was iterating away, spending billions of dollars, sure, but iterating away at smart glasses. Now Meta has the lead in the space, not Apple and not Google.

I would love to see a very competitive smart-glass market between Meta, Google, and Apple, but I really would like to see Google XR, in particular Android XR, stepping it up a notch and shipping much more quickly. Right now, Meta's running away with the space. Dave, what do you think, pal?

Dave

This is where society is going to have a huge rift, because the punching in the face was a very real thing last time Google went down this path. They are trying to own the consumer and be a consumer-friendly brand, but if they roll out a product where half of society is walking around recording everything all day long and the other half is offended by that, then that's going to be a major, major problem, and they're stuck. They want to own the space.

Alex

They have the technology to do it. People are going to want to talk to their Skippy, their agent, all day long. They're not going to want to lose touch with it. So this is a great way to stay in touch with your kind of agentic world that's working for you behind the scenes. I'm very eager to use it. I'm also not super eager to get punched in the face.

There were 3 commencement addresses this week, including Eric Schmidt's in Arizona, where, as soon as the commencement speaker said “AI,” the whole crowd went boo. If you're not aware that's what's going on, it's very easy to live in your echo chamber, especially here in Silicon Valley or in Cambridge. But you've got to walk around Mississippi or Nebraska to really understand how big a deal this is going to be. So I think the glasses are going to be a huge forcing function in this inevitable...

You're right. Good points. Audio glasses—hello, oxymoron.

Alex

Well, audiovisual oxymoron. No, they go together very well. But the audio feedback layer—I think they want to provide a product that actually works consistently. I think probably the imaging on AR glasses is still kind of weak. The version of the Meta glasses I've tried is okay, but it's still a far way off from being able to turn on an ambient AR layer that's convincing and compelling. The brightness isn't there.

But audio—having your AI be able to say, “Oh, okay. You were shopping for that. Do you like that one? I'll order it for you right now.” Or, importantly, when I see Alex approaching me and I've forgotten his name, the glasses can say, “Oh, you know, that's Alex Wissner-Gross. He's got an IQ of 100,000.” That should be enough. I'll never forget your name, Peter, for the record.

Thank you. I appreciate that. Nor I yours, my friend.

Alex

But I think the audio interface is a smart move to make it clean, compelling, and consistent—something that you can interface with on a regular basis. I am concerned about this issue. You guys all have this, right? You're with your family or with friends, something pops up on your phone, and you focus your attention on your phone. The loss of presence in life can be really costful. Really costful.

Yeah, it really can.

Alex Wissner-Gross

And also, I think the cameras are always on. They have very low battery consumption, so you can run the cameras continuously. It's just seeing everything you see and then talking to you about what you're looking at. So that's Alex, that's Elaine.

But if you really want a display, it can talk to your phone, and you can look on your phone to see anything it wants to tell you there. So it's all integrated through Bluetooth anyway. I think the consumer would rather have the longer battery life and not have to worry about it dying every hour, like Alex was saying.

This is a good temporary stepping stone. Like you said, Peter, putting the display in front of your eyes—if you think being not present in the moment is bad with this in your ear, imagine when it's flashing between you and your wife.

It's a good product design. I know that, Alex, we're going to lose you in a moment, but I wanted to have the last 2 segments with you still with us. Let's jump into Demis's presentation on Gemini for Science.

Demis Hassabis

I'm excited to announce Gemini for Science, which brings together powerful AI tools to help accelerate research. Gemini can already assist in solving complex problems, but our new Labs prototypes streamline daily scientific tasks, whether it's staying on top of newly published papers, transforming research goals into usable code, or generating new hypotheses.

Another powerful tool for science is simulation. AI simulations are going to be critical for understanding and predicting dynamic systems that are simply too complex to model directly today. Our state-of-the-art WeatherNext models can predict hurricane paths faster and more accurately than traditional systems.

At Isomorphic Labs, we're modeling molecular interactions to massively accelerate the development of new medicines, supported by leading industry partners. We're now in the preclinical stage with multiple projects, including potential treatments for immune disorders and cancer. When we look back at this time, I think we'll realize that we were standing in the foothills of the singularity.

Standing in the foothills of the singularity. I love that line. I wonder where Demis got that line. That's such a nice line. You've said that before, Alex. It's a nice line. Thanks, Demis.

Alex Wissner-Gross

I think it's wonderful. I'm broadly supportive of what Demis—Sir Demis, excuse me—is doing for DeepMind in science. I think it represents DeepMind at its best when it's challenging what Demis calls root-node problems, like fusion or protein folding. I think it's wonderful.

I don't think Google as a business has a deeply vested interest in monetizing this. I think this is more for the public benefit from Google's perspective. But I have portfolio companies, companies that I've founded, that work very closely with Google on issues and technologies relating to this, and I'm broadly super supportive of DeepMind pushing these out to the public.

I think it's very important, and they've made so many interesting—I would call them innovations—relating to metascience. How do you produce more science at the algorithmic level? I hope to see much more from them in the future on this front.

Amazing. Alex, listen, thank you for joining us on this segment. I know you need to jump. Love you as always. Thank you for your brilliance. Appreciate you, pal.

Alex Wissner-Gross

Thank you, Peter.

Everybody, welcome to the health section of Moonshots, brought to you by Fountain Life. You know, AI is impacting every aspect of our lives, how we teach our kids, how we do our business, but one of the most important things that AI can deliver to us is health. And one of the things I think about when, you know, shooting for 100, 120, is am I going to have the cognitive health to be able to think clearly and keep my wits about me for the next 50 years. I'm joined here today by Dr. Don Musalem, the chief medical officer of Fountain Life and a member of my Fountain Life medical team. Don, a pleasure. So, Don, talk to me about brain health.

Don Musalem

Brain health—you’re right. This is the number 1 concern people coming into Fountain Life have: Will I remember the name of my child and the face of my loved one?

Forty-five percent of dementia cases are entirely preventable with lifestyle. What was really intriguing to me, Peter, is that a quarter of our members had advanced brain age. But over 13 months of us helping them live healthier lifestyles—eating healthier, moving their bodies regularly, and optimizing sleep—we showed that we were able to improve brain age in 46% of those individuals.

That's amazing. You know, one of the things I love about Fountain is we're constantly searching the world for the most advanced therapeutics and bringing them to our members. So, for me and all of you, I hope that you appreciate the fact that you can become the CEO of your own health. You can make sure that you've got the cognitive clarity for the next 50 years. Come and check it out. fountainlife.com/peter to learn more and become the CEO of your health. Now, back to the episode.

All right, 1 more story from Google I/O. This one is very personal. We just announced the $2 million Build with Gemini XPRIZE. Let's take a listen, and then I'll provide some detail about it.

Logan Kilpatrick

And it's the ultimate platform to make an impact. We are officially launching the Build with Gemini XPRIZE hackathon. This global hackathon is going to offer up $2 million in prizes for builders who create apps that solve actual, real-world challenges.

The premise is simple: pick a problem worth solving, build with Gemini, and let's all try to positively impact the lives of 1 billion people. To build at that scale, you're going to need some serious power. So, a call-out to all hackers and builders out there.

We've launched 2 XPRIZEs in the last couple of months. The Future Vision XPRIZE is asking people to create a film trailer and a film treatment for the movie you'd love to see that shows a hopeful, compelling, abundant vision of the future. That XPRIZE is meant to help shape people's view of the future and actually help shape agents' view of the future, so they're positive, supportive, and aligned with humans.

This is one we've talked about on this pod all the time: The future is one of entrepreneurship. Instead of getting a degree to go get a job, find a passion, find a problem, and build on it. I want to say thank you to Google for funding this. They put up $3 million for the prize purse and $1 million for operations.

Again, if you want to compete, go to gemini.xprize.com. You're going to look for a problem that impacts 100,000 people or more. Then you've got basically 90 days to build your product using AI. Very famously, you're going to describe what you want the product to do, the market, and how you want to market it in English. The agent's going to build your website and your interfaces.

The team that's able to build it, market it, and scale revenue the most in 90 days wins this XPRIZE. Teaching people to fish instead of giving them fish—that's the goal here.

Salim, any thoughts?

Very exciting. It reminds me of a medical problem called lazy eye. There was no cure for it, and then a team built an app with a gamification system that greatly helped solve lazy eye. So I think there are all sorts of things we can do as we take obscure problems that affect a reasonable number of people and really go after them. Very, very exciting.

We've seen incredible outcomes when you put open innovation with these ecosystems throughout the history of XPRIZE. So, couldn't be prouder. Yeah, I'm super pumped. And, full disclosure, both Salim and Dave are on my board at XPRIZE Foundation. Dave.

Dave

You know, I think I'm excited. I think this is the highest calling for XPRIZE yet. Being involved with XPRIZE, I've been incredibly proud of it for a decade now. But now, with Google behind it, and with OpenAI also having a $100 billion charity, they need to put that money to work as well. All these megafunded companies are suddenly very, very interested in turning AI toward good, which is not an easy problem.

XPRIZE is really, really good at taking very hard problems and making them actionable. That's how I unleash that capital in ways that actually benefit humanity. You look at the Oil Cleanup XPRIZE and the impact that has had, and the kicking off of all of the space activity that we have via XPRIZE. So this is just the next chapter, but it's the biggest chapter by far.

And a reminder everybody watching and listening, if you want to be there at our Moonshot gathering on September 25th, go to moonshots.com. We're going to be awarding both the Gemini X Prize as well as the Future Vision X Prize on that day. We're going to have the five finalists, the five creators with the five top film trailers, and the five finalists with the Gemini X Gemini X Prize who have created the most revenue. They'll be there pitching, and if you're in the room, you're going to be helping to vote on who the winner is. So again, September 25th, go to moonshots.com to join the Moonshot mates and be there with us at the Moonshot gathering.

All right, I see Andrew Feldman has entered the room. Andrew, a pleasure to have you here.

Thank you for having me on your show. I appreciate it.

Of course. By way of introduction, Andrew is the co-founder and CEO of Cerebras, a pioneering wafer-scale computing company dedicated to accelerating AI training and inference. Cerebras just raised $5.5 billion. I should say, you just raised $5.5 billion—the biggest U.S. IPO since Uber in 2019. You're up 68%, with a market cap of $95 billion. Quite the coming-out party, Andrew.

It felt good.

Well, and you had to work for it, too.

This was super exciting for the team. We were able to bring a portion of the organization and their families to share it. My parents were there, as well as my wife and my stepdaughter. We made it into a family event, and it was really something special.

Amazing. Generational wealth for everybody in your organization. Extraordinary. We'll come to that in just a moment. A little bit more in AI news before we turn to Cerebras and chips.

A big story: Andre Karpathy joins Anthropic. Andrej is an extraordinary individual. He's joined the pre-training team at Anthropic to start a new initiative focused on using Claude to accelerate Claude's own pre-training research. In his announcement, he said the next few years at the frontier of LLMs will be especially formative, and that's where he wants to be. Karpathy has a stellar resume in AI. He was the co-founder of OpenAI. He left in 2017 to run full self-driving for Elon at Tesla. He returned to OpenAI in 2023 and left in 2024, and then he founded Eureka Labs.

Take a quick listen to Andrej on a podcast called No Priors. This is a conversation he had before the announcement. Listen up: he's doing a shout-out and call-out to which frontier lab wants to hire him.

Andrej Karpathy

And they're working on what's coming down the line. I think if you're outside of that frontier lab, your judgment fundamentally will start to drift because you're not part of what's coming down the line. So I feel like my judgment will inevitably start to drift as well, and I won't actually have an understanding of how these systems work under the hood. That's an opaque system. I won't have a good understanding of how it's going to develop, et cetera.

So I do think that, in that sense, I agree, and it's something I'm nervous about. I think it's worth basically being in touch with what's actually happening and actually being in a frontier lab. And if some of the frontier labs would have me come for some amount of time and do really good work for them and then maybe come back—

He's looking for a job. This is super exciting. [Laughter]

Andrej Karpathy

Then I think that's maybe a good setup because I feel like that's one way to actually be connected to what's happening but also not feel like you're necessarily fully controlled by those entities. So I think, honestly, in my mind, I can probably do extremely good work at OpenAI, but I also think my most impactful work could very well be outside of OpenAI.

Now, that's a call to be an independent researcher. That was Andrej on No Priors. In a quote he put out on X: “I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and remain deeply passionate about education.”

All right, Dave. Your take.

Dave Blakely

Andrew, he was calling to you. That was the moment he was saying, “Cerebras, I need your compute. Please call me.” [Laughter] He does. Everybody does, actually.

So Andrej came out with that autoresearch repo a couple of months ago, and I installed it. It's been auto-researching on my cloud for a while now, but it desperately wants much, much more compute. It doesn't need mega-models; it can run on lean, highly focused models. But I think he finally realized that every other OpenAI co-founder has either raised billions and billions of dollars or is at a frontier lab, and he's the only guy who doesn't have access to the big machine. He can't just sit there posting on GitHub anymore. He has to be part of one of these big machines because we're on the cusp of the singularity, and you can't miss it by being outside of the big game.

Salim, any thoughts?

I thought his point was well taken: if you're not in the frontier labs, you're missing out. Things are moving past you. I think it's also notable that he would have had his choice as to where to go, and I think it's very interesting that he joined Anthropic. Really interesting. It's interesting that this was announced on the same day as Google I/O—a little bit of marketing strategy, perhaps.

Andrew Feldman, what's your take on this?

Look, he is one of the most important and prolific thinkers in the space—not just, I think, as reflected in what he's built, but as reflected in how he's taught the community. I think it's interesting. His point that there are a small number of frontier labs that are sufficiently far ahead that, if you're not with them, you're not on the frontier probably applies to hardware, too.

If you're not building hardware for, or engaged with, at a fundamental level, one of the 3 labs—the 3 most important labs, Google, Anthropic, and OpenAI—you are not seeing what they're thinking. Just like your ideas will drift if you're a model maker, your hardware will drift from what they need as well. I thought that was really interesting, and I applied that to our domain. I think he has an extraordinary track record for useful stuff he's built.

Extraordinary. Yeah, yeah. He's also just a super, super ethical guy, and he's known for it. So he's a talent attractor. A lot of people here in the Valley want to work for the most ethical organization that they can navigate to. And he's got that aura all over him, as do you, Andrew, actually, and Cerebras, where we are right now.

Speaking about ethical organizations, let's jump into the conversation about OpenAI versus Elon. Okay, all right. How's that for a transition, guys?

Dave Asprey

Yeah, that's a stretch.

Yeah, you had to work for that one. I did. I did. I took advantage of Dave's point.

So, the jury ruled against Elon Musk in the OpenAI lawsuit. A federal jury unanimously rejected Elon's lawsuit after just 2 hours of deliberation. Wow. The jurors ruled that Musk waited too long to sue, outside the statute of limitations for his claims. Elon's legal team, of course, is going to appeal.

I don't know what to say other than I would love for this story not to cloud the entire AI data center conversation that we're having. I'm very curious, Andrew. OpenAI, you have a very close relationship with OpenAI. Did you have a lot of skin in the game in this outcome?

It's only going to be good news for Cerebras, but was it relevant? I think this was a giant distraction, and billionaires in pissing matches interest me not at all. These are 2 of the most important thinkers of our generation. I think what Elon has built is breathtaking. I've met them at dinner together. He's a polymath and a brilliant thinker.

What Sam has built is one of the fastest-growing companies in the history of capitalism. Some of his ideas, including the invention of the SAFE at Y Combinator, were enormously meaningful in the structure of Silicon Valley.

I think everybody loses when they battle. I want Elon building cool things, I want Sam building cool things, and I don't want to waste time or read about disagreements. I just want these guys doing what they're the best in the world at, which is building stuff.

Well said, Andrew. Well said. I'm just really angry it got to this point. Surely they could have looked at it and said, "The statute of limitations has expired." You just don't even bother. This was very upsetting. How much angst and time, and endless money, have been spent on this? How much more could they have done if they were building?

Good point. The ruling they made was obvious from the beginning if they were ruling based on the timing. But this is the reason Elon is going to appeal, because he says that's not the point. I looked into this, and the appeal will very, very likely fail because it's a fact-based decision, and the courts rarely overturn those.

Well, I think to Andrew's point, though, Demis said earlier in this podcast that we're standing in the foothills of the singularity. An appeal is slow and long, and it's going to be irrelevant in the timeline that really matters, I think.

Yeah.

All right, let's jump into part of the innermost loop: chips. And here we are. Andrew, Cerebras' record IPO closes up 68%, with a market cap of $95 billion. And Dave, you're in the Cerebras offices this morning, aren't you?

Dave Asprey

I am. I love the energy. It's such a rare moment. I think all of America is driven by this moment, people building toward this event. Andrew's journey was longer and harder than a lot of others, and to actually get there is so hard.

I pushed that button at about 1% of the size of Cerebras, but I still pushed that same button on the Nasdaq. I didn't realize until that picture had been hanging on my wall for years now. It's like a lifetime achievement, like a Nobel Prize or an Olympic gold medal, where you carry it for the rest of your life. So few people get to experience that.

I couldn't resist the opportunity. I was only a few miles away anyway. I couldn't resist the opportunity to come here and just feel the aura of it. It's still settling in here, clearly, Andrew. Is that the way it feels?

Yeah, and you're always welcome to come on by. The mojo that we've got going here—we try and create an environment where exceptional people can do extraordinary work.

Dave Asprey

Andrew, you and I met at the Citibank event, I don't know, 6 months ago. I was there giving the dinner keynote on longevity, and I wish I had had a chance to get into your friends-and-family round before you had this epic release.

If you don't mind, tell all of our listeners and viewers here a little bit about the backstory, the founding story behind Cerebras. What was the moment when you said, "This needs to exist"? You took a very different route than NVIDIA and other chip manufacturers.

We did. The founders had all worked together at my previous startup, and AMD bought that in 2012. By 2015, we'd wandered off a little bit, and we started meeting. We saw AI on the horizon, and what we knew was that this new workload would eat through extraordinary amounts of compute.

We made 2 really big bets. We made a bet that, like graphics produced the GPU and like mobile compute supported the development of the ARM processor, this technology—this work—would be big enough to require dedicated silicon. The second bet we made was that the right strategy wasn't to build a derivative of the GPU. You needed to start with a clean sheet of paper and do something fundamentally different.

These were enormously contrarian bets at the time, and both proved to be dead right. From that foundation, we continued the innovative thinking and said that what AI was going to need was memory bandwidth. That's the sort of speed with which you can move data from memory to compute.

The way to innovate on that dimension is to use a different type of memory than everybody else uses. We have 2 types of memory. We have memory that can store a lot, but it's slow. We call that DRAM or HBM. And we have memory that's fast but can't store very much per square millimeter.

What we hypothesized was that if we could build a chip the size of a dinner plate—a chip 58 times larger than any chip ever built before—we could stuff it to the gills with SRAM, thereby overcoming its weakness in not being able to store very much per square millimeter and benefiting from its strength.

That proved to be a very difficult problem to tackle. But when we got it, it proved to be right. We are somewhere between 15 and 20 times faster than the GPU on any inference problem. The challenge along the way was that nobody had ever built a chip this big—not once in the 75-year history of the computer industry.

Keep on slicing them thinner and thinner and smaller and smaller.

That's right. Even those on the Mount Rushmore of our industry, people like Gene Amdahl, had failed, crashed, and burned. Interestingly, even after we solved this problem, we had people come and visit our labs and then try to build it, and they also failed.

It took years of perseverance and innovation, and all the credit goes to the engineering team: Gary, Sean, Michael, JP, and the team we had. We failed for years. In August 2019, we announced that we'd solved this problem that had been unsolved forever.

We thought everybody would rush to our door, and the world didn't care one bit. Why? The world was utterly indifferent. In the first generation, I think we sold 12 systems. In the second generation, we sold 300–350, and in the third generation, we're selling many, many, many thousands.

What happened was that we solved this problem and were way ahead of the market. It wasn't until late 2024, early 2025, that the models got fast enough and the models got smart enough that people wanted to use inference everywhere. That's what happened. There we were with the fastest inference machines on Earth by orders of magnitude, and suddenly people wanted to use AI. The way we use AI is with inference, and we were just crushed with demand.

In December 2025, we signed a deal with OpenAI for north of $20 billion over several years, one of the largest deals ever signed in Silicon Valley. In March, we signed a term sheet with AWS for deployment in their data centers. Business has been pretty good since.

Well, congratulations. Let me take a second and welcome Alex back. Alex, good to have you back.

Alex

Good to be back. Amazing to meet you, Andrew.

How are you going, Alex?

Alex

Doing well. I wanted to ask you, Andrew: I was talking to Valavan yesterday, your chief product architect, a brilliant guy in Toronto and an awesome friend. But I didn't realize that the company had a whole history as a training-side company. Inference is now, what, 80%–90% of the market? Moving a huge amount of data from SRAM through the processing massively benefits the inference side.

Did you see that coming in the initial design? Because I don't think a lot of the research people even knew that inference would be so dominant.

I think that we got many, many bets wrong. Any CEO who looks back over a decade that moved as quickly as ours and says they got it all right is probably not a guy you want at your birthday party. We got an enormous amount wrong.

But one of the things we got right was an understanding that we make AI with training and we use it with inference. If AI is going to be smart and useful, you need to have an inference business. And that bit we saw early.

The real problem between 2020 and 2024–2025 was that it wasn't smart enough to be useful. Everybody was focused, and all the labs were talking about the number of parameters. Now people don't care. The only question is: Does it write good code? Does it give me good answers? Can it do things that I want done?

We've moved into a regime, into a world in which it's useful, and that's how it's measured. We did recognize this. We are really good at training, but right now there's such demand for fast inference—such overwhelming demand—that we're allocating a lot of our attention to it.

Alex

I'm curious, Andrew. SRAM—you mentioned SRAM earlier. The largest models, really the standard models at this point that offer frontier capabilities, are in some cases up to 10 trillion parameters. How do you think about SRAM when—correct me if I'm wrong—the Wafer-Scale Engine 3 has maybe in the tens of gigabytes of SRAM?

40–50.

Alex

Something like that. But the largest models are in the trillions of parameters. How do you think about the future of SRAM, given that, as you said, you're stuffing it to the gills?

I think the following: Models that size have to be divided up, whether you're using GPUs or TPUs or us. They have to be cut up and spread over multiple chips. Remember, models that size, Alex, have a very large matrix multiply in the attention head.

And that doesn't fit on a GPU. You have to cut it up and go tensor model parallel. You don't have to do that with us, but you do have to spread it over 4, 6, or 8 chips. What you do is divide the model very carefully, such that no layer runs over 2 chips.

What you're moving is the result from 1 layer to the next. You can move it because that's a very small vector—a result vector. You can move it over 100-gigabit Ethernet. That little hop is slower, but the calculations that take up such a big portion of the time are so much faster that you pay a very small penalty for breaking it up into 4, 8, 16, or 20 chips: on the order of 2%.

Now, other SRAM solutions that are small—for example, Groq, which NVIDIA acquired—have to break it up because they have only 800 square millimeters to use. They have to break a big model up over 2,000 or 3,000 chips, and each of those hops hurts their performance. We have to do a few hops; they have to do thousands.

There's no way ever to fit everything on any size or amount of memory. But there is a very nice and simple way for us to cleave models and spread them over multiple chips. Yesterday, we announced and posted numbers on Kimiko 2, which is a trillion-parameter model in the open-source community. We were, of course, an order of magnitude faster than anybody else.

Oh, really? How many chips? How many wafers on that?

I forget. I've been busy.

I imagine.

It was about 1,000 tokens per second, where a really good GPU shop like Fireworks is running at 70.

Yeah. Yeah.

And they're a really good shop. So, 15×.

Yeah, that's pretty good. Yeah.

Dave

Peter and I were at Google I/O yesterday, and they showed a whole rack of TPUs operating together, generating 1,400 tokens per second and writing code. You see that, and you're like, “I need that. I need that tomorrow.”

Well, the trick there, Dave, sometimes is—and NVIDIA has been masterful at this sleight of hand—not telling you whether they mean tokens per second per user or aggregate throughput. The GPU is an extraordinarily good machine at generating slow tokens.

Dave

Yeah.

You can generate an NVL72 with 35 tokens per second, which is painfully slow, and it can generate millions of tokens. On the other hand, if you ask it to generate tokens at 200 tokens per second per user, it can support 1 or 2 users. That's a $4 million solution working on 1 user.

Dave

Yeah.

Right? It's really important when you dig into these: Are they telling you gross throughput? Is this a lot of customers who are unhappy with their performance, or are they able to serve individual customers—and how many of them? Because, Andrew, you said—

Go ahead.

You were being complimentary of Elon as an extraordinary builder, entrepreneur, and polymath earlier. One of the best in history, for sure. I'm curious: He steps up and announces Terafab, producing 50 times the number of chips that exist on the planet today, outstripping TSMC and everybody else. What do you think of Terafab? I'm super curious.

Look, I think Elon has proven himself on multiple dimensions. He's proven himself to be a visionary. The number of people who said, “You're an idiot to try and build cars in Fremont.” I mean, we've got the highest labor rates in the country, maybe among the highest in the world. We've got a regulatory regime that's unfriendly to business.

There were also the people who said you shouldn't build a rocket company, and the people who didn't understand that he was building a rocket company because he wanted to build a satellite company and needed the rocket company. He has been ahead of everybody for a very long time.

Okay, but you're talking about the vision side. That's the vision side.

He's also been able to execute on some of them, not all. That's what's cool: He's trying to do things that other people can't do.

Now, this particular problem I know a little about. Building fabs is very hard, and it's hard in a different way from some of the other problems he's attacked. I'm not saying he can't do it. I'm saying it will always take longer than he says, and it will cost vastly more money.

That's the challenge of building extraordinary things. It is not a 5- or 10-year project, in my humble view. I've been wrong before, but I put this at a 15- or 20-year project.

Wow. I think it's interesting. It's probably good for the US that we have domestic fabs.

But I think there is a reason why, even with the exact same equipment from ASML, Samsung and TSMC aren't at the same node. TSMC is ahead, and they're extraordinary. The amount of received wisdom and learning from the fabs they've built over generations cannot be underestimated.

If anybody can do it, Elon can do it.

What does that mean for Cerebras? Obviously, US manufacturing—and, like you said, you know a lot about this topic—but US manufacturing of chips is critical. I mean critical for everything: national defense, global security, everything.

If it's going to take 15 or 20 years, that's just Terafab. The TSMC migration to the US is going very slowly, way behind schedule. What does that mean in terms of, first of all, supply and demand—the raw ability to get things made? You must deal with this every day.

These things are hard to build, right? I mean, fabs are pyramids. They are our pyramids, and TSMC is the greatest manufacturing company on Earth. The challenge is that these things take 5 or 6 years to build and $40 billion to $50 billion, from the people who built the last one. That's true whether it's TSMC, Samsung, or any of the great builders here.

These are unbelievably difficult to build, and that's why they've been behind schedule in the US. I think they encountered some challenges that were unforeseen. I think we have political challenges. These things take long enough that they cross administration boundaries. When your projects can't be done in 4 years and have to cross 2 or 3 different administrations over a period of time, you have a problem.

When you have local ordinances that get in the way of building, as happened with Samsung's fab in Texas, they redesigned the fab because of a local fire ordinance that made no sense. These are painful problems that our system hasn't found a way to overcome.

I think we have to find a way to do better, because I think the reshoring of fabs—and not just the fab; the fab gets all the glory, but the packaging business is every bit as important—is something we lost entirely when the fabs left.

Yeah, actually, a good question. By the time you get something ready to put into 1 of your data centers or a third-party data center, how many different manufacturing partners has that wafer been through?

A fair number. I mean, it goes to ASE, which deposits RDL on the backside. It's diced, it's cleaned, and it comes to us for the first step. It's a long process.

When we stopped caring about fabs in the '90s, IBM sort of left, and GlobalFoundries' fabs—we didn't do anything to keep them—we lost this collection of surrounding expertise.

When a chip comes off a fab, it's a dead piece of silicon. The package is how you breathe power and life into it, how you get I/O into it, and how you get power into it. That's also an enormously challenging technology. It takes materials scientists, manufacturing engineers, process engineers, and deposition engineers.

We punted all of that by not caring about this industry, and it's all sitting in Taipei and Korea. The materials are manufactured in Japan. Kyocera is 1 of the leaders there.

We've got to get it all back, and we've got to make a decades-long commitment to this industry.

If you said that Terafab is 10-plus years out, if I look 5 years out, do you think Cerebras is able to manufacture on Intel, Samsung, and TSMC? Are there any other choices?

We've committed our 3-nanometer design to TSMC, so that will take us out a little bit. We do manufacture some components at Samsung and have a great deal of respect for Samsung's fab capabilities.

We've never used Intel. Lip-Bu Tan is an extraordinary leader and a longtime advocate for hardware in Silicon Valley. As you know, Dave, there was a period between about 2006 or 2007 and 2015 or 2016 when every VC firm was filled with somebody from VMware who didn't know anything about hardware, who thought compute was made by flea feces in the cloud.

We tried to explain for a long time that the way you make more virtual compute is to begin with real compute.

I've got to tell you, you've inspired so many people on campuses right now who are eager to be part of your mission to get that back. I'm going to route as many as I can through this building here to you.

Please do. Guys like Andy Bechtolsheim and Lip-Bu Tan, and a few others, were continuing to put money into hardware. Pierre Lamond continued to do it and support us as we were trying to raise money during that time period.

I think—well, I know Lip-Bu Tan. They've got a lot of work to do. He's done great things so far, but they've got some work to do before we could move to Intel. Alex or Salim, do you have a question you want to ask?

Alex

I have a quick one. Andrew, you've gone from raw invention, solving fundamental, big problems, to now going into production. When you want to scale these things, can you say how long it takes to create one of those chips? Then, over time, as you get better and more efficient in the manufacturing process, what do you hope it shrinks to?

Well, I say that the first one took 4 years and maybe half a billion dollars—somewhere between $400 and $500 million. That's why I take it to dinner when I go with my wife. Like a 10-year-old with a first dirt bike. It's coming to bed. It's in the bedroom. It's not outside in the garage. It is being carried around everywhere I go. I got a wafer.

I think the inventions cut across lithography, chip architecture, packaging, cooling, power delivery, and cooling. They included compiler inventions and algorithmic inventions. In fact, some of the hardest problems that we encountered were packaging, and we solved them 7 or 8 years before others encountered them.

The B100 was or the B200 was 18 months late, and it was late because they had a problem with CoWoS. CoWoS is a process step where TSMC uses a 65-nanometer chunk of silicon as a motherboard. They put NVIDIA's chips and the memory on it, and instead of putting it on a green board—a traditional motherboard—they put it on a piece of silicon. The wires are more efficient in silicon; they can be narrower.

This was a big invention, but we knew there would be a problem with the coefficient of thermal expansion. We knew that because we'd solved that problem in 2018. So there they were, in 2024 or 2025, struggling with a problem we'd solved 7 years earlier. That's what happens when you do pioneering work: You encounter problems, and you have a chance to solve them long before the rest of the industry even encounters them and knows they're a problem. That is one of the joys.

Obviously, in everything we do in engineering, there's a trade-off. The downside is that there are some low days. We had about 18 months where we were spending $8 million a month and couldn't solve the problem. When you have board meetings every 6 weeks, you come in, you're still trying to solve it, you still can't solve it, and you're $100 million more in the hole. Then you're $120 million in the hole. Then you're $140 million in the hole, and you still can't solve it. These are some low days. Then you have the IPO of the year, and it's a high day.

That's right. And I think, Peter, you're one—

It's such an archetypal story of an entrepreneur.

It is the entrepreneurial journey. Yeah, yeah. I think, Salim, one of the things I've learned along the way—this is my 5th startup—is that this will kill you if you can't modulate the highs and the lows. It will.

For every entrepreneur and every CEO, I tell them, first, that this is a pressure test on your soul. Second, the number of times you can get kicked in the gut before lunchtime and still have it be a good day as a CEO of a startup is amazing. Would you rather be doing anything else? No. This is what I know how to do. I'm a professional David in the battle with Goliath. This is what I know. I have no interest in doing other things, and I have no interest in working with people other than those who want to attack the hardest problems.

Yeah. Amazing. Alex, please.

Alex

Speaking of the hardest problems—and it's almost in the name Cerebras—you have, I think, a 4-trillion-transistor budget with your 3rd-generation wafer-scale engine. I'd love to talk a little bit about what's at the end of the rainbow, projecting out, say, 10 years, when you're on your nth-generation model. What does the future look like? Does it look like brain uploads running on WSE-8? What's the killer app? What does this look like in 10 years?

Alex, I think one of the fun things about being an infrastructure builder is that you don't have to have those ideas. No, really. I was with the team—and many of them are here—in the mid-'90s that helped drive down the cost of networking. We built some of the first and fastest Ethernet switches, and we had no idea that WhatsApp would arrive and make it possible even for the poorest members of our society to communicate home.

When I grew up in the '70s, the only thing I heard my grandmother say on the phone was, "Put your brother on. It's expensive." It was $4 a minute for my mother to call Australia, where her mother was. They spoke for 6 minutes a week. The only thing I heard my grandmother say was—I’d say, "Hello, Bubba." She'd say, "Put your brother on. It's expensive."

We, in a company called Yago, along with many others—with Juniper and others—put a small brick in the wall that made the cost of IP transport so low that somebody else could invent the technology that made it possible for every person to talk to their grandparents, no matter how poor they are, anywhere in the world. That's something that we didn't know. That's not the problem I set out to solve, and it's not the problem our company set out to solve.

We set out to solve a problem as infrastructure builders: We build roads. What you drive over those roads and how far you take them—that's other people's work. What we're trying to do is allow people to do extraordinary things on our infrastructure. When I think about what we're enabling, that's work for Sam. That's work for Ilya. That's work for others. What we're trying to do is make a compute platform on which their ideas can take flight. What we know is that you need faster calculations.

So what I think I heard you say is that you're very deliberately not having opinions as to the future shape of the workloads that will run on your infrastructure, and you're primarily, at this point, deferring to the frontier labs to steer the future architecture of workloads—today's frontier labs or new frontier labs, right?

We are making bets that the world will continue to depend on sparse linear algebra as an underpinning for all these calculations.

This episode is brought to you by Blitzify, autonomous software development with infinite code context. Blitzify uses thousands of specialized AI agents that think for hours to understand enterprise-scale codebases with millions of lines of code. Engineers start every development sprint with the Blitzify platform, bringing in their development requirements. The Blitzify platform provides a plan, then generates and precompiles code for each task. Blitzify delivers 80% or more of the development work autonomously, while providing a guide for the final 20% of human development work required to complete the sprint. Enterprises are achieving a 5x engineering velocity increase when incorporating Blitzify as their pre-IDE development tool, pairing it with their coding copilot of choice to bring an AI native SDLC into their org. Ready to 5x your engineering velocity? Visit blitzify.com to schedule a demo and start building with Blitzify today.

One final question: orbital data centers—fiction, real, must-have? Are you going to put your chips up there?

I think, first, we have a serious advantage in space. In space, some of the most expensive work is the chip-to-chip communication. We've had chips in space for a long time. That's what a satellite is. A satellite looks like a PC motherboard with a camera stuck on it—a big telescope, right?

If you unpack what's in one of these small satellites, every computer hobbyist will say, "Holy cow, that looks like a server motherboard with a telescope stuck to it." Then it's hardened. Communicating and building a cluster is actually much more complicated because you have to do a lot of communication in this work.

Moving the data from the land to the cluster is a problem that we've solved a long time ago. They will continue to improve it. So being a big chip and having to move things off-chip less often is a huge advantage. I think this is an exciting domain.

Like many hard problems, the last 10% don't take 10%; they take 90%, right? Self-driving is one of those categories. The last 10%—we've been sitting at it for 8 or 10 years, and we're just now getting over the hump of the last 10% because it isn't really 10%. I've got orbital data centers in the 7-to-10-year category.

So, it's interesting, Andrew, just to pull on that very briefly. What I would have expected you to say to that would have been something like: With the wafer-scale engine, you had to design around faults. You had to be incredibly fault-tolerant at wafer scale, and in a space environment with lots of ionizing radiation, you also need to be fault-tolerant. Cerebras, with its experience with fault tolerance at wafer scale, is like the perfect computing platform for highly ionizing environments.

There are—I think we have lots of advantages, Alex, and you put your finger on one of them. You have to try and shield your silicon very differently in space.

Alex

You will get more flaws. There are single-bit errors, hard errors, and a whole collection of errors that you have to contemplate. Our ability to shut down a core and route around it is an enormous advantage in that environment.

I think we've got, as a community, some work to do over the next 4 or 5 years before we have the truly hard part of getting them in space: orchestrating the software and getting them to communicate. I've got it out to the better part of a decade before we have production in space. I think it's a very worthwhile project to pursue, but it's out a ways.

All right, Andrew. We close out these segments with an AMA with our incredible subscriber base, and we'd love to have you join us. We've chosen 8 questions from our comments, which we all love and read, and we'll be peppering them along. I'll put them up here.

I'm going to give you first shot. Andrew, you can look at the others and get ready for one of them. We have a second page we'll go to. Salim, why don't you pick one of these?

I'll go with the first one: “If the world becomes compute-constrained, does the 10-cent lawyer-for-everyone thesis still hold? Does AI become a luxury only the rich can afford?” And this comes from @GeoRust1.

If you've been listening to the podcast, this should be a fairly clear path here, right?

Every major technology starts scarce and very expensive, and then you see this with computing, bandwidth, DNA sequencing, and solar energy. They all look very constrained and very expensive initially, but then the learning curve kicks in, infrastructure kicks in, Jevons' paradox kicks in, Wright's law arrives, and competition arrives. The cost collapses.

AI compute is going exactly down the same path. You may have some bottlenecks, like chips, power generation, and data centers, but those become investment honeypots, and capital floods toward those bottlenecks.

The bigger insight is that AI is not consuming compute; it's helping design chips and optimize infrastructure—what Alex calls the inner loop. It's improving and compressing the models, and the system becomes recursive. As you have intelligence building more intelligence, this is why we get so excited by this future. It's going to drive the cost of everything down.

Maybe it's a $2 lawyer for us. Maybe it's 50 cents, but over time it's going to get to 10 cents. And that's down from $1,000 an hour.

That's number 1. Number 2, I think the problem with lawyers and accountants is that the structure of their business model for the future is exactly wrong.

Selling hours, yeah.

Right? Their business is to stand between ordinary people and obscure knowledge. That's what your accountant does. You don't want to figure out what the tax rules are related to depreciation on a property you bought or that was gifted to you in 2000. Who wants to know that?

Their business is the acquisition of obscure knowledge and the application of that knowledge to particular problems. That's exactly what language models do.

Don't you think, Andrew, that generalizes, though? What are you, other than a gatekeeper of obscure knowledge regarding the tax rules?

I so disagree with that. Honestly, what are we all? We're all specialists.

Andrew is something like an engineer.

They're actually—the reason they're called counsel is that they're giving good advice, not just on legal matters. When they're giving good business counsel, when common sense is challenging in a confusing environment, those are the times when they're at their best.

I think when you're drafting all the documents you need, most things have already been drafted. We don't need another lawyer reviewing another NDA. We don't need to—

Andrew, let's give you the next shot. Which question—2, 3, or 4—do you want to choose?

I think number 3 is interesting. I think there's a profound misunderstanding about how things—

Let's read the question: “Why can't money buy Elon or Zuck a lead in AI? Can't they just buy the best talent?” And that's from @novarift. Great question.

No. The answer is no. Why couldn't Intel build a cell phone processor? At the time, they had the best fabs. At the time, they had the best computer architects, and they destroyed tens of billions of shareholder dollars failing. Same with AMD.

It turns out in our industry that money and the acquisition of talent isn't enough. What is? There's something else.

MTP—what?

Purpose. Massively Transformative Purpose.

They could have said no. I mean, Intel could have said yes to Apple, though.

No, no, but they could have. The truth is, what led them to believe that they were in a position to say no to Apple? Intel was chasing margins. Intel was infatuated with its own profits. They had an ARM division that they sold off. Intel could have sold cell phones.

That's the thing we're trying to understand: what's going on?

They were fat, happy, and lazy. Maybe. Maybe. Or maybe there's something in your DNA that makes—

Big mutations.

I'm sitting here saying all day long, we'll take luck over skill. But I'll also say that extremely hardworking people with tremendous grit end up more lucky, and that both of those are true. That is life.

It is really hardworking people over long periods of time who have integrity and ethics, right? They get lucky more often. Luck is not equally distributed to those who work hard and those who don't.

But I think, Alex, the question isn't, of course, whether they could. Why didn't they? Why did they miss it? Why did AMD miss it? Why did—for example—NVIDIA fail for decades at everything that wasn't the GPU? They failed to build an ARM processor that worked. It was called Snapdragon. I think they failed at a Northbridge-Southbridge part, and they succeeded beyond anybody's expectations at a GPU.

I think the same question holds true for the Yankees, right? Why doesn't the team with the biggest budget win every year in the NFL? Why is there something in our thinking about organizations and talent that we don't do a very good job of describing?

There is something that is very hard to buy and that has to be made. We don't seem to be able to articulate it well, and buying the most talent doesn't seem to be sufficient. You have to have a lot of talent; it's necessary, but it's clearly not sufficient.

All right, Alex, let's go to you next for a question. I want to get through our lightning round here.

Alex

Yes, of course. I have some pretty different answers to some of these other questions, but I think I have to answer question number 2, which looks like it might have been a response to a comment that I made in a previous podcast episode.

The question is: “Why wouldn't Sam—I think this is referring to Sam Altman—cut a deal with Bezos and Blue Origin to become the other counterweight to Elon?” And this is from Scott Ray Broomfield.

I think the answer is that's probably on the table. If I were Sam, I would be exploring a variety of potential heavy-launch partnerships to become a counterweight to Elon and SpaceX's AI Dyson swarm. I think heavy launch is already arguably a critical element of the stack for the future of compute in space.

Space, but not as much.

Alexandr Wang

Yep. As you know, New Glenn is more akin to Falcon 9 and doesn't hold a candle to Starship, which, by the way, will be making a launch attempt probably by the time this is out. Good luck to Elon on that launch attempt. I'm super excited.

But Starship is coming in at a factor of probably 100 times cheaper.

I'm not sure that matters. Go ahead, Andrew.

No, I think, Alex, all your points are right. And I think you underestimate Sam to your tremendous cost. What Sam has done again and again in our industry is see around corners that other people missed.

He was trying to lock down data center capacity and space last year and the year before, when all the other foundation labs didn't see it.

Really?

To lock down memory.

Oh, yeah, for sure. I didn't know that.

His ability to look at an exponential and not be afraid of what it says in 2 or 3 years, while everybody else is afraid, saying, “We're not going to need that much,” is extraordinary. And his reach is extraordinary.

With 100% certainty, I will tell you that he is exploring deals through every possible avenue to get access to compute and data center capacity. I can say that having watched from a distance. I have no inside information, but I've been dazzled by that ability of his.

I think you underestimate that guy.

You would think it would be enough to build the fastest-growing company in the history of capitalism to get a lot of respect, right? You think that might be a sufficient feather in your cap. But I think he will certainly be in conversations to get compute, whether it's in space, under the sea, using falling water, or using geothermal. He will be in those conversations, and his team will be there every single day. It's so cool to have another friend who's on the big stage. That inside perspective is awesome.

Alex's one point here is that, again, Elon's Dyson swarm is 500,000 satellites to 1,000,000 satellites. It would be like a launch every couple of minutes of a Starship. You don't get that with New Glenn, so that vehicle isn't designed for the frequency of launch that we're talking about here. Could Sam put up a mini-constellation with Bezos? Sure. Could he put something up to really compete with what Elon has proposed? Not without new launch capability. That capability is unique on the planet if it pans out as expected.

Yeah, a few thoughts. There are lots of options. Yes, I agree with the contention that SpaceX is completely dominating mass to orbit—no question about it, including dominating historic mass to orbit. So, if I'm Sam, I would be exploring probably a multichannel strategy. First, I'd be exploring a deal with Elon and SpaceX to leverage SpaceX's launch capability for my own Dyson swarm. I would be exploring alternative launch capabilities.

And then, if you really believe—and I think this is the elephant in the space room—that we're on this singularity-esque exponential, the fabs don't need to be on the ground. We can build fabs in space, and we may not be addicted to heavy launch 5 to 10 years from now. If you're Sam and you're playing the long game, then you're looking for ways to build fabs on the Moon and in LEO that don't need SpaceX's near-monopoly.

Mhm. Amazing. Dave, do you want to take us to question 4? Andrew, did you have another thought on that?

David Friedberg

No, I think building a fab on land is hard enough for me.

Great segue to question 4, which is directly related. China is building massive compute capacity. Could they sell tokens to U.S. users at very low prices and disrupt providers?

David Friedberg

It's the same answer you just gave, Andrew: no. China is not building massive compute capacity when you're looking at tokens per second, the driver of AI. They don't have that; that's why they're so desperate to import from the U.S. They're building as quickly as they possibly can, but it's not 5-nanometer, 4-nanometer, or 3-nanometer technology. It's all bottlenecked by the ASML machines, fab construction, and everything Andrew has been talking about. If they had the ability, they would love to do that, but they just don't have the compute.

Guest 3

Yeah, the dimension in which they've chosen to invest so far is power infrastructure. At that, they're just playing better than us right now. They have upgraded their grid; they have tremendous power infrastructure, and we've made bad decisions there. We are stuck with a grid that's built in the '50s. It's designed not for what we'd like it to do today, and we have trouble politically, at the local, municipal, state, and federal levels, doing projects like infrastructure.

What they have done is make a tremendous amount of investment there. They are obviously starved of compute, but they're going to try to build on what they have, which is an absurd amount of power infrastructure.

Guest 4

Also, when enterprises are going to be doing most of the token purchasing, you're not just buying the token. You're buying trust, governance, reliability, compliance, et cetera.

Guest 5

I would perhaps just add to this: it's worth noting that in the past week or so, there's been quite a bit of public reporting about how China is operating proxy services that are selling American tokens to Chinese users at incredibly low prices. They're offering 10× discounts in order to siphon the reasoning traces for training their own models. Anthropic is pursuing that.

Andrew, we want to thank you for being on. We close out every episode with user-generated content. This is sort of our our outro music. And so, let's enjoy this one. It's called We Are As Gods, a I guess comment to my new book. Today Andrew is as a god, so we have to like bow down to I think being a CEO is sufficient, but I don't want the responsibility. I congratulate Andrew on all of that. Congratulations to Andrew on an epic IPO. Amazing. All right, let's listen to We Are As Gods by Mussad Zamani. All right, enjoy. Everyone's a founder now. The power's in your hands. Everyone's a founder now. Abundance is the lever. AI is the lever. The future is forever. The mission is your soul. TAKE THE FULL CONTROL. OH, Peter sees the vision and Alex writes the code. While Dave defines the logic, and Salim leads the road. The agency of one. See the power rise. The exponential life beneath the digital skies. All right. A good one. Another good one. That was a good one. Very cool.

Again, thank you for joining us. Gentlemen, always a pleasure. I think we could have kept both conversations going for a couple more hours.

Yeah, easily. Easily. Thank you for having me. I really appreciate it. Be well. Thank you.

Thanks, Andrew.

If you made it to the end of this episode, which you obviously did, I consider you a moonshot mate. Every week my moonshot mates and I spend a lot of energy and time to really deliver you the news that matters. If you're a subscriber, thank you. If you're not a subscriber yet, please consider subscribing so you get the news as it comes out. I also want to invite you to join me on my weekly newsletter called Metatrends. I have a research team. You may not know this, but we spend the entire week looking at the metatrends that are impacting your family, your company, your industry, your nation. And I put this into a 2-minute read every week. If you'd like to get access to the Metatrends newsletter every week, go to diamandis.com/metatrends. That's diamandis.com/metatrends. Thank you again for joining us today. It's a blast for us to put this together every week.