Google创纪录季度、白宫介入，以及GPT 5.5悄然追平 Mythos｜EP 254

Google has crushed its earnings. Alphabet reported $109.9 billion in revenue, representing 22% year-over-year growth, and $62.6 billion in profit. Google Cloud hit $20 billion in revenue with 63% growth. AI drove results across the entire Google ecosystem.

Alex

The revenue just goes up and up and up and up, and everyone's like, “How's that possible?”

The White House is considering a process of vetting all the models before release.

Alex Gross

The capabilities of AI are going to continue to grow exponentially. They're incredibly valuable to the military, so the government ultimately has to pre-vet these things, right?

It has to, but it can't gatekeep. That's where we're going to end up falling behind from a geopolitical perspective.

Alex Gross

I'm more worried about the frontier labs self-policing more aggressively than the government ever would and stifling competition that way. I think that's a far scarier future. This is the future that we're going to live in forever hereafter.

All right, let's move on to the news. There's a lot going on this week, especially around Google, OpenAI, and the state of the AI race. Let's jump in.

The White House is considering a process of vetting all the models before release. The Trump administration has really flipped its position. They were all about openness: “AI companies, go as fast as you can, no restrictions.” And all of a sudden, there's a proposed executive order that says, “No, no, we're going to create a working group with tech leaders and government officials that's going to pre-vet these before they're released.” Alex, to you, buddy. What does this mean, do you think?

Alex Gross

My sense is that everything changed with Mythos. I have to add the caveat that, based on a number of public cybersecurity benchmarks, it appears that GPT-5.5—which, unlike Claude Mythos, is actually generally available—is stronger at these cybersecurity benchmarks.

Looking back from near-future history—so, looking back historically—I think we'll view the Mythos moment, if you will, as a sea change: the moment when the civilian sector, the frontier AI labs, suddenly had capabilities that leapfrogged government capabilities. Specifically, Mythos was suddenly able to, as Peter and I talk about in Solve Everything, have entire disciplines get solved at once. With the Mythos moment, cybersecurity, and in particular vulnerability discovery, got effectively solved, for some definition of “solved,” by AI for the first time in the private sector, leapfrogging what possibly the NSA or other government agencies had internally.

This was a moment when the government—even an aggressively deregulatory, AI-friendly government, as the present administration is—sort of woke up and realized, “Hey, wait a minute. These are leapfrog capabilities coming from the private sector. They could lead to vulnerabilities in government systems, vulnerabilities in industrial and SCADA systems throughout the economy. Maybe, actually, some sort of light-touch gatekeeping mechanism might be merited at this point.”

So, I think, without putting my finger on the scale of whether this is actually a good idea or not, and not answering the normative question, it's a natural time to at least be answering the question of whether certain advanced capabilities are perhaps, in some sense, naturally gatekept by some quasi-governmental entity.

And we're going to have Michael Kratsios on the show very shortly, right? He's overseeing a lot of the technology side of this in the Trump White House.

It’ll be an interesting conversation. You know, another thing to mention, Alex—we’ll have your view on this. In one fashion, this pre-release vetting creates sort of a compliance mode that OpenAI, Google, and Anthropic can afford, but the smaller labs cannot, right? This might create a limiting of the field to an oligopoly of AI labs. Do you think that might be the case?

Alex Gross

There was always a bit of a moat there, even before any new executive orders. I’m thinking in particular of export controls that regulate the ability for open-source or closed-source capabilities to be shared with the public. There’s the Invention Secrecy Act, which has been statutorily on the books for many decades and functions as a sort of gatekeeping for anyone who wants to file a patent application that touches on certain sensitive areas. There’s the Atomic Energy Act from the early 1950s that also gatekeeps certain elements of new applied physics.

It’s not as if we’re suddenly entering some brave new world where the government—this administration or some other administration—suddenly decides that new technologies must be gatekept by government oversight. We’ve been in that regime, arguably, since World War II. It’s just that AI capabilities coming from the private sector are now so capable, so strong, that this government—and probably, I would speculate, future U.S. governments—may feel a strong need to suddenly step into the loop from a gatekeeping perspective.

Yeah. Dave or Brian?

Brian

Brian, you were a West Point alum, first boots on the ground in Syria, and an Army Ranger. You certainly know more about the federal government’s internals than practically anyone else. This has to be inevitable, though, right? The capabilities of AI are going to grow exponentially and continue to grow exponentially. They’re incredibly valuable to the military.

The idea that the frontier labs can just pump them out and make them available across the world—and then there’s no recalling it if it’s already out there in the world—means the government ultimately has to pre-vet these things, right? It has to, but it can’t gatekeep. That’s where we’re going to end up falling behind from a geopolitical perspective. This is not a political podcast, so I’ll pause on that, but it is a challenge if there are veto rights versus partnership and understanding.

Yeah.

Do you think there’s a three-month lag, embargo-type thing that might be coming?

Brian

Open source is three months behind right now, so you’re basically creating parity between the closed-source and open-source models to do this.

What’s going to happen? We’ve only talked about this being the first model I know of that’s actually been held up, you know. Remember, though, GPT-2 and GPT-3—the memory dulls—but those models were also held up, purportedly for safety reasons.

Alex Gross

As Dario said when he was at OpenAI, I think there’s a long history of a little bit of—call it—moral panic over new AI capabilities as radical new capabilities like vulnerability discovery suddenly come online. It’s probably natural, on the one hand, and the question I think one could ask is: Would you rather the moral panic be held by the frontier labs doing the gatekeeping, or would you rather that it be held by a democratically elected government? Someone, somewhere, is inevitably going to have this moral panic anytime new capabilities come online.

We’ve already said that the labs are going to hold back on their cutting-edge capabilities because they’re going to use them internally. In some ways, that will benefit them if the government is saying, “Wow, that’s way too powerful for you to release,” and they release derivative products based on it—discoveries in physics, whatever the case might be. Don’t you think?

Alex Gross

Yes. If you’re asking now for my normative position on this, I’m more worried not that the government is going to too aggressively gatekeep the models. I’m worried that the frontier labs themselves will self-censor so aggressively for a variety of reasons—whether the new models are too compute-intensive, or they want to leverage the models just for their own commercial benefit and not share them.

I’m more worried about the frontier labs self-policing more aggressively than the government ever would and stifling competition that way. I think that’s a far scarier future.

Keeping on this theme of the government, the Pentagon has signed agreements with 7 AI companies, including Google, SpaceX, xAI, OpenAI, Amazon, and Microsoft, for military applications. What’s interesting about this article is that Google's agreement provides that they can provide AI to the Pentagon for any lawful government purpose. This prompted a protest by 600 Google employees.

If you all remember, I remember this in 2018, when Google had a walkout of 20,000 employees—not 600 employees, but 20,000 employees. It was called the Project Maven walkout, when Google disclosed it was using its capabilities, its early AI and its models, and its search capabilities for government applications. Thoughts on this one, gentlemen?

Alex Gross

I would just say that not only did those Google employees protest—and this has been publicly reported—they unionized, which is something we’ve never seen before. We’ve never seen this in a frontier space like AI. Just think of the juxtaposition of 19th-century union-style organization on the one hand with 21st-century technology on the other. These were British DeepMind employees who were unionizing to protest Google entering into this agreement with the Pentagon.

We’ve never seen this bizarre juxtaposition before. I think it’s probably not a great look for Google that they have employees unionizing outside the continental U.S., outside the U.S. overall, to protest working for arguably patriotic purposes with the U.S. military. Not a good look at all. On the other hand, I would say the 7 companies—I think Reflection AI is somewhere in there as well—at least underline the upside in my mind, which is that there’s enough competition in the frontier-model space that the Department of Defense has enough other counterparties to go to.

If it’s unwilling or unable to work with Anthropic, at least there will be other models available on SIPRNet and JWICS.

I can’t wait for the agents to unionize. It’s easy to forget that Google—we think of it as a U.S.-founded lab—but the DeepMind unit is in London.

Jay Bregman

Demis Hassabis is in London, and he desperately wanted to spin that business unit out. This is all coming out in the new book, The Infinity Machine. It’s a great, great history of how this all evolved. DeepMind would have spun out and become essentially like Anthropic, and then Google promoted it, backed off, changed its mind, and kept it internal.

That unit is culturally kind of separate from Google anyway. It’s not a U.S. entity; it’s part of a U.S. company. So it’s clearly signing this agreement with the Pentagon, but there must be some serious internal friction there.

All right, look who’s entered the room. The emperor of exponential organizations, Salim Ismail. Good to have you joining us. Where are you today?

I’m in Toronto. I came into the country, and they said, “Do you realize your passport has run out of pages?” Because I travel so much, I had to go to the renewal office. I was standing in line for the last half an hour getting that done, so that’s now in process. Ironic as hell. When are we going to have digital passports without this old process of physical stamps?

It really is. Where’s your passport-abundance mindset, Salim?

There was a guy behind the ticket counter who was stapling things together. I’m like, “Whoa, how retro.”

Oh, my God. Any comments on the Pentagon signing with the frontier labs?

I can understand the employee backlash because AI is not just a tool now. It’s becoming a decision layer. You can understand why, but navigating this is going to be crazy. Let’s see what happens.

Yeah, it is going to be crazy. All right, let’s continue with the Google stories here. Google has crushed its earnings. Alphabet reported $109.9 billion. I would be happy with just the $0.9 billion in revenue, you know. That’s 22% year-on-year growth and $62.6 billion in profit. Google Cloud hit $20 billion in revenue with 63% growth, outpacing both AWS and Azure. AI drove results across the entire Google ecosystem. Interestingly enough, Google now has ¾ of a billion monthly active users. Dave, let’s go to you for this. What are your thoughts?

David Friedberg

Yeah, it’s interesting that the YouTube acquisition was the greatest acquisition of all time, although the rival would be Facebook’s acquisition of Instagram.

Which is now the majority of its market cap, or maybe Google’s acquisition of DeepMind, which is now driving Google.

David Friedberg

I remember I met Chad Hurley, who had sold YouTube for $1.65 billion.

And interestingly enough, I don't know if people know this story, but Google actually had Google Video. YouTube was scaling faster because it had no lawyers and no restrictions on what you could post. They were scaling so rapidly that Google had no choice but to buy them.

Yeah. Well, the story within the story here, too, is that Google's search volume flattened in about 2017. It's been flat ever since, yet the revenue just goes up and up and up. Everyone's like, “How's that possible?” The way that's possible is ad targeting, and the driver of ad targeting is AI.

Google had a really easy road to where they are now, in the sense that every time they worked on AI, it instantly turned into revenue and profit. Very different from Tesla or OpenAI. They really had the perfect storm of opportunity, and they took advantage of it. To their credit, they took advantage of it, and not everybody does that. But here they are, on the cusp of being the most valuable company in the world again.

Yeah, we're going to see that in just a minute. Jay—

Jay Bregman

I want to note that Google Cloud had a rather difficult childbirth. Think back a few years: There was a point at which, reportedly, the co-founders of Google had passed a mandate to Thomas Kurian. Either Google Cloud had to become the No. 1 or No. 2 public cloud, or it would simply be removed. It would be excised from Alphabet.

That was, I think, a dangerous time, and there were a variety of documents and internal memos regarding the future of GCP getting leaked at the time. I think Google and Alphabet, to their credit, stood up and took a stand against those who would rather Google not have stayed in the public cloud race. They carved out Google Cloud as its own line item in quarterly reports just in time for the AI tailwind.

Now, thanks to the AI tailwind, exposing both TPUs to their own customers and TPU compute capacity to other frontier labs was also a very good move, arguably. Then maybe even offering TPUs for direct sale to other data centers. I think Google Cloud not only has a fighting chance but is arguably, as many others have mentioned, in a unique position from a vertical-integration perspective to potentially leapfrog both AWS and Azure. It's an exciting time.

Not many people know this, but EverQuote, where I'm the chairman, had Google come to us and say, “If you use Google Cloud, we'll give you ad credits on Google Search.” The CEO, Seth Birnbaum, said, “What an incredible deal.” But then they know that once you're on Google Cloud, it's not trivial to move off.

Any entrepreneur, any great entrepreneur, will tell you there's a lot of luck in the process. But recognizing when you have those lucky moments, Google capitalized on each of those strokes of luck.

Jay, you just said it's matured dramatically. What was that?

Jay Bregman

Yeah. GCP as a product has matured dramatically. GCP in 2018 compared with GCP in 2026 is unrecognizable.

Yeah. Kind of a late start, and then they pushed it really hard.

Jay Bregman

They've done a great job of attracting talent, too—until recently. Now everybody wants to be part of Blitzy. But recognizing great talent, you know, the taxicabs in San Francisco used to have, on top of them, ads that were difficult math problems. They would say, “If you want to work at Google, solve this problem.”

Love that. Isn't that crazy? I mean, just so creative in their recruiting. It's like Palmer Luckey's employee ads that say, “You don't want to work here,” right? It's like negative incentives. Interesting.

Jay, you said something I think is really important. Unless Google can be No. 1 or No. 2 in a category, they drop it. I don't know if you remember Circles, right, when they were going after social.

Jay Bregman

Of course. Who could forget?

Yeah, I forgot. A lot of people forgot.

Jay Bregman

I did. I didn't. Actually, Google bought my company at the time, and that was the contact-management part of Circles and all that stuff.

Interesting. It was a classic play: They couldn't out-innovate Facebook at all because of all the approval layers inside Google. Facebook was just running circles around everybody at that time.

Jay Bregman

Yeah. Again, agility is your number-one killer capability.

All right, let's continue on. Wait, I've got 2 quick comments. When we do our ExO rankings, Google is consistently at the top because they've created an unbelievable flywheel of data feeding algorithms, algorithms running in the cloud that gives you distribution, capital, and talent, all kind of reinforcing each other. This is an amazing story that's just going to keep going.

Our next story: Even Google is compute-constrained. The innermost loop is a harsh mistress. Let's take a listen to Demis talk about this.

Demis Hassabis

For us, I mean, there is a question of resources, talent, and compute. Nobody has enough spare compute to just make 2 frontier models at maximum size, with different attributes, right? So that's pretty difficult.

But for now, what we've decided is that our edge models—the things we want to use for Android and glasses and robotics—it's best that they're open models because they're vulnerable anyway on the surface. Once you put them out on the surface, they might as well be actually fully open.

Fascinating. This is the world's largest infrastructure builder basically saying they can't build fast enough, and they're turning away revenue. Brian, any thoughts on this one?

Brian

Well, we always knew that devices on the edge were going to use open source. It just makes more sense from a security perspective. But Google is literally making people apply and get in line for large amounts of compute. We've never seen anything like it before, and so you have to be one of the most important people in the market to be competitive.

I think this is one of the most important topics we can talk about, too, because when you talk to corporate America, they take compute for granted. I had a long conversation with Kushagra Vaid yesterday from CoreWeave. That company has grown like wildfire because compute is constrained, and everybody's going to CoreWeave to reserve their future compute. You can buy compute futures for the first time.

But most of corporate America isn't aware that this is the new normal forever hereafter. If you look at the rate that Blitzy can consume compute productively, it's almost infinite. It's almost unlimited. We're all used to there being surplus compute. You just go to the cloud anytime you want, buy whatever you want, and it's always right there. It's like going to the grocery store: Of course there'll be milk on the shelves.

That will probably never be true again. But corporate America isn't aware of it, and so they're not reserving and building their own capacity. They're going to really suffer 2 to 3 years from now when there's nothing available. They'll immediately realize, “Wow, I could automate huge fractions of my business and turn it into profit, and I can use Blitzy to recode everything. Oh, wait. We don't have any compute.”

Brian Johnson

Teraflop, baby. I've got 1 word for you: Teraflop.

Dave Blakely

Yeah. TerraFab.

Interesting. Alex, what are your thoughts here?

Alex Salkever

Yeah. One thing I think most people don't realize is that the situation is so severe—and this has been publicly reported—that even within Google, the 3 main compute consumers, which are Search, Google Cloud, and DeepMind, have to fight periodically for new compute capacity that comes online for their respective divisions. I think it's been reported as happening once a week or once a month.

I do think this is a preview of the future where what gets prized ultimately, as we spoke about in a previous pod, is per-token economic productivity. The highest, most revenue-generating or most profit-generating tokens will ultimately receive the most compute.

We're seeing this now not just at the frontier labs. We've spoken in the past about how Anthropic's strategy seems squarely aimed at maximizing dollar value per token. Similarly, even within Google, they're all fighting it out to see who can generate the most dollar value per token. I think that type of liquid market, or auctions per token, is the future we're going to find ourselves in.

You need a metric number that is the AWG metric on dollars per token created.

Alex Salkever

Yes.

Yeah, for sure. This is not investment advice, but it is the innermost loop. The stocks that are skyrocketing right now, as we'll see in a little bit, are the chip and energy companies. If you have something that's massively constrained that's driving the global economy, I don't know where else you put capital.

Dave Blakely

Credit where credit is due: You turned my daily newsletter into non-investment investment advice. Bravo.

Oh, my God. We'll get to that in a moment. But this is what we're seeing. Google's market cap is within 4% of overtaking NVIDIA. I didn't look today to see if it's closed the gap.

Alex Salkever

It's still pretty close. I checked.

Yeah. Honestly, what we're basically seeing is AI is now driving the value, not anything else. They've successfully done the crossover. It's no longer search.

Now AI delivery is driving their valuation. Dave, any thoughts here?

Dave Blakely

Yeah. I think in my entire life, if you bought a box of chips, you would really regret it a year later. This is the first year of my life where, if you bought a box of random RAM a year ago, you would be way up today.

I’m calling the ball: This is the future that we're going to live in forever hereafter. This is not a temporary shortage. Even if TerraFab comes online on time—which it won’t, right? There’s no chance of it coming online on time—even if it did, we would use up all that compute instantaneously.

AI is the first thing we’ve ever had in human history that has an infinite appetite to create. Every new GPU is another disease cured. It’s another person fed in Somalia. It’s pure value every time you create one of these off the line. That’s why you see the other stock, Intel—

We’ve been talking about that on the pod for a year.

Dave Blakely

What was it, like $19 a share when we started saying, “Look, Intel’s fabs are going to be critical to the future”?

Everybody—AMD is up, Micron is up, SanDisk is up. We’ll see that in a couple of minutes. Salim, any parting thoughts on this one?

Two things. One is that I think Dave makes a really great point: The demand is going to be near-infinite, and we’ve never seen this before in any technology. Jevons’s paradox goes completely insane in this model.

But the big, provocative question is: Does the future belong to chip monopolies, or does it belong to intelligence utilities? Which way will it go? I’m curious what people—

What’s that, Alex?

Alex Salkever

Or neither.

Or neither. I mean, the vertical, right? This is where xAI—SpaceX AI with launch—is now being part of the innermost loop for getting data centers up. It’s crazy.

Guys, everybody listening, I just hope you hear this, because it’s going to determine our economic futures, and it’s not slowing down. How high could it go? Guess what? Higher.

Alex Salkever

It’s called the singularity for a reason.

All right.

Alex Salkever

It goes in a single direction: up.

Yeah. There are asymptotes involved.

All right. I want to turn the story to a few OpenAI stories. OpenAI drifts from Microsoft and moves toward Amazon. So here’s the story: OpenAI has ended Microsoft’s Azure-only exclusivity and is now running on AWS, Google Cloud, and Oracle.

Just as a reminder, we talked about this in the last couple of pods: OpenAI signed a $100 billion AWS deal over 8 years, making Amazon its major partner. What does this mean for Microsoft? Are they going to start competing openly with OpenAI? Are they going to spin up their own models now? Alex, what do you think?

Alex Salkever

Remember when Satya made that now-infamous comment about how Microsoft was good for their $80 billion? It was a sort of backhanded reference to Microsoft not being good for supplying all of the voracious appetite for compute that OpenAI basically demanded under their prior engagement with Microsoft. I think we’re seeing the fallout of that.

I think we’re seeing OpenAI and Anthropic having voracious compute appetites, and Microsoft—at least a former iteration of Microsoft, call it all of a year ago—thinking that they were being very fiscally responsible by limiting their data center build-out and everything that goes with it, including mega-tranches of corporate debt and credit in the data center credit markets.

You can sort of trace a line of causality from Microsoft’s decision at that time to OpenAI today being essentially starved of Microsoft-only compute and needing to diversify beyond even the original concept of Stargate. Remember, Stargate originally was this sort of alliance with Microsoft and then all of Microsoft’s suppliers. Then Oracle came into the picture, SoftBank came into the picture, and all of these other suppliers came into the picture.

Stargate was no longer about OpenAI directly being the single tenant for data centers that they were financing. Instead, it became a branding moniker for leasing compute from a variety of third-party providers. All of these are connected into a single causal chain, which is that Microsoft—and also OpenAI’s not-for-profit status; that’s part of the story as well—wasn’t in a position to supply enough compute for OpenAI’s demands.

As a result, the OpenAI-Microsoft marriage has turned into what we see now: OpenAI is dating everyone else at this point.

Yeah. I’d love to give it that spin. If you think right now, OpenAI has a host of problems, not to mention the Elon Musk lawsuit. Mustafa Suleyman, when we had him on the podcast, was pretty clear that the mandate for Microsoft is to build its own foundation model because Microsoft has all the intellectual property from OpenAI contractually delivered, but they’re struggling to read the files.

I think that was off camera. I heard that actually from a Microsoft insider, a very close friend. So now it looks like both companies have problems, but they actually had the most perfect marriage—early on, total domination, an incredible lead—and they might have taken the marriage for granted a little too much, instead of doing what Google and DeepMind did, which is partner up and do something epic.

Microsoft and OpenAI could have gone down that road, but they didn’t. I wonder how much they regret that. But I’m wondering right now—

Alex Salkever

There were corporate governance issues. OpenAI was a nonprofit, and they needed to invest in a for-profit. They created the for-profit subsidiary in part so Microsoft could invest. It was complicated.

GPT-5.5.

Dave Blakely

It is dramatically amazing.

Is it?

Dave Blakely

OpenAI is doing just fine with all of its decisions. GPT-5.5 is equivalent to Mythos. That’s my core belief. It is an unbelievable model, and people are dramatically underreacting.

Alex Salkever

It’s actually better than Mythos, according to some of the cybersecurity benchmarks, which are finding that it’s hitting the same capability levels at 5 times cheaper and is actually generally available.

But they have early access, right? They’re deep on this. 5.5 is unbelievable. Anthropic is compute-constrained, and that’s why it’s not going to market.

Dave Blakely

Well, that’s interesting, because GPT-5.5 is also available now on Amazon Bedrock. You can get it inside a secure environment, so for sensitive or corporate use, you can keep your prompts and your results secret from the provider.

That’s for the first time. That’s only been available for, what, a month now?

Dave Blakely

That’s been available, but it’s a big game changer.

Okay. Well, maybe that’s huge. I hadn’t heard anyone say that 5.5 is actually better—as good as Mythos, which isn’t available. Dave, I talk about it in my newsletter every day.

Dave Blakely

Yeah. The first thing I do every morning is read your newsletter.

Groggy, but—

Dave Blakely

Or Grok-y.

Salim, do you want to weigh in here?

No. I find this mostly lots of “Who’s the belle at the ball?” type of stuff. I think the next stories are much more interesting.

All right. Well, let’s go to the next story here. OpenAI misses its targets in 2025, and there’s conversation about delaying the IPO.

OpenAI missed its internal goal of 1 billion weekly ChatGPT users at the end of 2025, and multiple revenue targets were also missed in early 2026. The CFO, Sarah Friar, whom I’ve had a chance to hear speak a couple of times, warned that they could struggle to meet their data center obligations if growth stagnates and suggested waiting until 2027 for an IPO.

We should talk about the implications, but she went on to say that the company doesn’t meet reporting standards for public companies. That is a remarkable admission for a CFO to make. Dave, what do you make of that?

David Blumberg

There are 2 versions of the interpretation of that sentence. One is, “We don’t have the visibility into our revenue to comfortably predict 2 or 3 quarters in advance.” That’s the usual interpretation.

You’re on public company boards. How many public companies are you part of right now?

David Blumberg

Just 1 right now.

How many have you been part of over the years?

David Blumberg

As a board member, MicroStrategy and EverQuote, and then as an adviser, a whole bunch.

What do you make of them missing their targets consistently? Of course, right now they’re shifting to GPT-5.5, like Brian said, which is epic, and moving down the corporate road.

David Blumberg

One interpretation is, “We just raised $120 billion. We don’t really need to be promoting and rushing toward any exit right now.” We’re in a great financial spot, and that’s not uncommon.

You see, Google doesn’t make nearly as much news and drama as the other labs do, but they quietly have everything they need. They have cash flow, they have their own chips, and they have their own everything. So what’s the point of making news?

OpenAI has had to promote the heck out of itself right up until it closed that $120 billion. Now it’s in such a financially comfortable spot that it can start to say things like, “Maybe we should temper expectations. Maybe 2027, maybe 2028, is a better year to go out.”

I kind of read it that way. Dave, remember the conversation we had about the supply of capital? xAI—SpaceX AI—is going to soak up a lot of capital.

David Blumberg

And we were saying, okay, number two to the table is going to pick up the rest. Number three is going to be left at the altar. It looks like Anthropic might be number two. And then if OpenAI pushes into 2027—

You know, is the appetite going to still be there?

David Blumberg

Yeah, totally. Well, I don't know if you remember, but the numbers are so big today compared to any time in history. Remember when Yahoo went public, and then Lycos and Excite? It all happened in just a few weeks. Internet portals were going to be huge, and AltaVista was also out there as part of Digital, but it wasn't in that IPO window.

What tends to happen is that these things go public back-to-back within a category because it's much easier to educate the global investor community in one batch.

And then everybody wants to be part of it, and all the money pours in.

David Blumberg

But if you miss that wave of IPOs, it's much harder to find the capital a year or two later. It's not tragic or devastating or anything, but it is a much easier IPO if it's part of the trend, and it's all relative to the other companies in the sector. Anthropic could easily be that.

Alex, one of the points that was made was that they need to meet their data center build commitments. What do you make of that?

Alex Kantrowitz

Well, a few things. One, I think the underlying story here is that, as I've mentioned previously, OpenAI was betting on consumers to carry it to its revenue targets. That turned out to have just been a terrible idea. Consumers don't want to spend lots of money on reasoning tokens. Enterprises do.

So, pivoting back from consumer to enterprise—which Anthropic, due to its own compute limitations, was betting on almost the entire time, at least from far earlier on than OpenAI was—cost them, and that may have ultimately delayed their revenue targets for what would have been their IPO.

Now, GPT-5.5 is out. Codex is looking stronger than Claude Code at the moment. I expect leapfrogging to continue, but that probably did set back OpenAI's internal revenue projections somewhat. At the same time, they're backing out of Stargate as it was originally construed, and now it's just a leasing operation. It's no longer a data center build operation, so that should free them up quite a bit.

It's a bizarre situation, though, if Sarah is leaking these expectations. It almost smells to me like an expectation-reanchoring game. Why, if you're about to go public, do you have your CFO leaking these stories to The Wall Street Journal and other major publications that things might not be as rosy as they otherwise seem and that you might have to delay your IPO?

That's the sort of exercise in PR that a company goes through if maybe it's trying to re-anchor expectations lower than they actually are, so that it can exceed and beat them on a shorter time scale.

I'm so glad you said the first part of what you said because a lot of people are unwilling to say, “Oh, they made a strategic error.” But it was just—

Alex Kantrowitz

It was a blunder.

It's so clear.

Alex Kantrowitz

Yeah.

David Blumberg

Yeah, and there was also a blunder.

Alex Kantrowitz

But there's a really important follow-on to that, too, because remember, at the time that everybody thought consumers were going to eat every token, they also predicted that Google Search would get obliterated from the planet.

Yeah.

Alex Kantrowitz

And then all that ad revenue would go away. So all the stocks that are tied to Google ad revenue are down 70%, 80%, 90% now.

It turns out Google ad revenue isn't going to go away because all the tokens are going to go to the highest-value use, which turns out to be enterprise. Exactly what you said a second ago, Peter.

And that means that Google's lifespan on its search revenue—which is still 90% or so of Google's gross margin between YouTube and Google Search ad revenue—has a much longer lifespan than you would have predicted two years ago. All the companies in that ecosystem are in much better shape than you would have predicted two years ago, and the enterprise revenue is where all the tokens are going to go.

But if that use case—if Blitzy keeps eating tokens at its current ramp rate—they're not going to be available for consumer use until after the AI bubble. So, a long time in the future. There's been a big change in the landscape.

If you're listening, we're hiring for a CFO. So, if things don't work out with you and Sam, you can move to Cambridge, Massachusetts. Bad. That's funny.

Brian Johnson

Yeah, I mean, I think people don't realize how hard it is to run the ARR at these exponentially growing companies. We have billing challenges with every single model provider, including Google, which is the most buttoned-up organization of all time from this perspective.

This is a fundamentally hard problem, and it is hard to predict two to three quarters out what's going to happen in an N-of-1 moment in technology. So, Sarah's probably right. It's incredibly challenging to know what's going to happen three quarters from now, put that in a 10-K, and put your name behind it.

All right. I want to go to you, Salim, on this one. Labs are partnering with private equity firms. OpenAI finalized a $10 billion venture with TPG, Brookfield, and Advent. Anthropic launched a $1.5 billion venture with Blackstone, Goldman Sachs, and Hellman & Friedman to deploy its model, Claude.

Both are focused on deploying AI across enterprise operations and portfolio companies. I mean, this is the fox in the henhouse, right? These private equity firms control trillions of dollars and thousands of companies, and this is the direct route into the main vein for the AI drug. Salim, what do you see here?

We've been predicting this for a while because it's a natural consequence. AI is not coming in through the CIO or through the CEO. It's going to come in through governance, top-down, and be forced into companies because there's too much internal resistance. Doing it this way breaks the immune system because you can just mandate it.

What's going to happen now is that all these companies will start to create this digital twin at the edge. We've started to talk to a bunch of these folks already, right? It reminds me a little bit of how we're all looking for—or, Peter, you've been looking for a use case for space forever—and all of a sudden, data centers. What the hell?

Private equity becomes the main deployment channel for enterprise AI going forward because they have a perfect AI laboratory: hundreds of legacy companies with radical inefficiency. This now takes AI from chatbot experiment into EBITDA transformation.

This is what we call the organizational singularity. It's going to come in through the enterprise, not through HR, not through IT, but through private equity, top-down, through the operating partner. So, expect to see a lot more of this.

Yeah, expect to see a lot more of this because people are going to say, “It's just not working to do it the old way, so we have to do it aggressively and brute-force it top-down.”

Salim, if you're a small- or medium-sized company CEO, like many who are listening to this podcast right now, and you're not a billion-dollar, private-equity-owned company, what do you take away from this?

You'd better get on the train, and get on it fast. If you're not disrupting yourself with your digital twin, somebody's going to come along and disrupt you very badly, very quickly. These guys are going to start eating markets very quickly.

The one caveat is that this is going to take a lot longer and be a lot harder than people think, because when you go into a legacy company, you don't have the skill set or the capability to wipe out the legacy and redo things. But you've got to force a cultural change, and that's nontrivial in many of these companies.

I bet. Dave, thoughts here?

David Blumberg

Yeah. Well, private equity is funny. It just keeps business schools alive decade after decade. There's always something. But it's been the best-performing asset class of any asset class for, God, 30 years now. Even better than venture. Only seed-stage venture outperforms private equity.

And you're like, well, why is that? Well, there's always something. Computerization was a huge tailwind for PE. All these legacy companies were working with pens, pencils, and paper and were never going to move to a computerized environment. So, let's just acquire it, retool it, make it much more efficient, and then take it public again.

And so now AI is that times, you know, 1,000. Yeah, the arbitrage is going to be amazing. Also, if you buy a company that's very complicated, like a legacy manufacturer or a white-collar operation, getting to know what they do is so hard. You bring in a brilliant management team, and they come in, but understanding a legacy business is incredibly difficult.

David Friedberg

Oh, wait. AI is the perfect power tool for scouring every document, interviewing every employee, gathering all that information, and looking at all the legacy systems.

I think the war chest of tools with AI that PE now has is like nothing they've ever experienced before. I expect PE returns will go through another—

David Friedberg

One of these cycles, like when computerization was a wave—

—where the PE returns are just staggeringly high, and it's all because of AI automation.

Alex Kantrowitz

Can I make a hot take here?

Okay, Alex, and then we'll go back to you, Salim. Go ahead, Alex.

Alex Kantrowitz

So, a hot take. The elephant in the room: how is this money going to be spent? $10 billion from OpenAI, $1.5 billion from Anthropic.

A skeptic—which I'm not—might argue that there's a very real risk that these monies are going to be used to basically pay the respective frontier labs for their own sales.

It’s sort of OpenAI spending $10 billion—or I guess they’ve contributed part of the $10 billion—but that’s ultimately a bit circular. So, the same folks who were arguing that all of these deals in the past year or so that Nvidia was striking with other folks in their supply chain constituted Nvidia doing circular sales are now saying that OpenAI and Anthropic are basically launching these ventures, or co-branded ventures, as a way to drive their own sales through circular sales mechanisms and wash sales. That’s what a skeptic would say.

I guess there’s a second elephant in this particular room, which is that the private equity firms have got to be staring down future discounted cash flows and being quite scared by it. If AI is just eating away all of these otherwise relatively predictable future cash flows of all of their operating portfolio companies, and AI marches into the room and suddenly they only have—as we’ve talked about in the pod in the past—maybe they only have 2 to 3 years of runway left in these cash flows before AI just obsoletes the cash flows, and you’re a PE company, and OpenAI or Anthropic come into your office and say, “We’d like to set up a JV with you. Billions of dollars, and you can spend the billions of dollars on your portcos,” that plugs a hole in their discounted cash flows, quite attractive but also seductive for them. In which case, the frontier labs maybe get something that approximates a wash sale, and the PE firms get to plug a hole in their discounted future cash flows for the moment that makes them look good to their LPs.

Amazing. Salim, would you agree?

Yes, but just to take the whole other side of this, I think they’re going to find it brutally harder than they think to make this all work. For example, you can try to go into a company and scan all the documents, et cetera. But there’s a statistic that’s pretty surreal, which is that 44% of Gen Z workers today are deliberately corrupting the AI that they’ve been asked to help automate because it sort of won’t take their jobs. It’s literally criminal malpractice what they’re doing, just in self-defense.

So, you’re going to get all sorts of messiness and chaos as they go through this transition, and I think this is going to be much harder. There’s a methodology being developed here that nobody’s ever had to do before. This is completely new territory, and we’ll talk more about that on another episode.

On the next episode, Salim—or the one after that, depending on when we have Michael Kratsios—we should dissect the Organizational Singularity paper that you’re about to publish.

We will do that. I’m ready to talk about it. In the next slot, we’ll go into it in detail. It’s one thing to have a PE firm pressure you, as a large company, to utilize AI to the fullest, which they will. But again, if you’re a solopreneur, if you’re a business owner, small or medium-sized business, either you as the CEO need to take that role of the PE firm here and demand it of your team, or if you’re a board member listening to this, you need to unify the board and demand that of your CEO. There are zero excuses if you don’t.

This must be a big topic among the Harvard Business School alumni crowd, right? A lot of your classmates must be in private equity.

Yeah. We work across almost every single private equity portfolio, and it’s not as if they’re resistant to change. They’re just fatigued by the tools sent by the board every single week, with a new thing to try.

What I would encourage folks to think about is not just the cost-reduction mechanisms, but actually the revenue acceleration that you can bring inside of these AI tools. The managers are so fatigued by “cost cut out with AI, cost cut out with AI,” versus what is possible now that wasn’t possible a year ago.

Interesting. All right, I’m going to move us on to a fun topic. This is the march toward AGI, whatever the heck that means, and consciousness. Let’s listen to this first video here.

Greg Brockman

One thing I have learned is that everyone has their own intuitions about what AGI is, and maybe you can view it as, according to my view of where we are, I think we’re about 80% of the way there.

So, first of all, I think the point Greg Brockman, the president of OpenAI, was making—that everybody has their own view and you sort of intuit whether it’s AGI or not—is a very squishy definition. Anthropic’s Jack Clark came in with this quote: “I believe recursive self-improvement has a 60% chance of happening by the end of 2028.” I’m super curious about your thoughts there, Alex.

But first, the other story I’ve paired here is that Richard Dawkins says that Claude may already be conscious. Quote: “If these machines aren’t conscious, what more could it possibly take?” Alex, over to you, pal.

Alex Kantrowitz

Well, Richard Dawkins first: I think hell has frozen over. Richard Dawkins, as I mentioned in my newsletter, is sort of a biological reductionist-in-chief, The Selfish Gene extraordinaire. Even implying that Claude may be conscious, whatever he may mean by that, I think this is an extraordinary moment in the biological philosophy of frontier models.

Going back to Greg and Jack, taking Greg first, it’s difficult to know what Greg is really thinking when he says 80% to AGI. Historically, going back to the OpenAI-Microsoft discussion, there was the contractual definition at one point between OpenAI and Microsoft that AGI meant generating $100 billion in revenue. So, Greg may be thinking we’re 80% of the way to generating $100 billion in revenue off of our models.

If I had to guess, I’d guess his estimate or his definition is probably something like that. He may be thinking in revenue terms or, let’s say, economic terms, maybe in terms of the supply chain and data center build-out. When Anthropic—when Jack in particular—talks about a 60% chance of happening by the end of 2028, that one’s a real head-scratcher for me, much more of a head-scratcher than Greg, because Anthropic has publicly said that almost all of their code at this point is being generated by Claude and that Claude accounts for substantially all of the training and logic for the next generation of Claude.

So, I’m not sure how much more recursive the recursive self-improvement could be at this point. Maybe he’s just throwing out a really conservative outer bound, or maybe he has some threshold of progress or improvement. There are a few different benchmarks for capturing the rate of recursive self-improvement. Maybe he has some internal notion of one particular benchmark passing 60% by the end of 2028. But I think, on the outer bound, Jack’s estimate is far too conservative relative to every indication we’ve seen out of Anthropic to date.

Salim, what are your thoughts here, pal?

I totally agree with Alex on the Greg Brockman commentary. I’m also surprised at the Anthropic thing, because I think—we’re, my understanding is, we’re 90% there and could be there within months. Or maybe that last 10% is a really hard one, and it’s just going to take that much longer.

A few years ago, I was asked to moderate a debate between Richard Dawkins and Deepak Chopra, which I refused because there was going to be more heat than light. Then I watched the debate, and they were definitely yelling at each other and talking to each other. Richard Dawkins is very much a phenomenologist, so he’s coming at it from the bottom up. When you see the AIs simulating or acting that way, he approaches it from that perspective.

I do disagree with the concept because I think they’re mimicking consciousness. That’s a very different thing from actually being conscious. My bigger point, though, is not about whether AI is conscious, but whether it’s operationally autonomous.

Discussing the philosophical aspect of this is fascinating and great, but CEOs and governments need to be much more worried about the agents that can plan, execute, negotiate, code, persuade, et cetera. All of that stuff is happening. So, it becomes a non sequitur and an orthogonal discussion to the really important conversation. I think the recursive self-improvement is the really big deal, though. That one hit me: holy crap.

Alex,

Alex Kantrowitz

It’s important to parse out the foundation model versus the AI system.

Brian Roemmele

LLMs are sequence-to-sequence. They are fundamentally not an architecture that will get to AGI, but you can construct AI systems to have reinforcement loops to get better as you use them. So, when we say AI—AI systems, yes, a foundation model as a standalone transformer architecture is not going to happen.

Whoa, Brian, that’s quite the hot take. Do you want to define or explain how you operationalize AGI?

Brian Sathianathan

So, I believe systems that can learn on the fly outside of training data—that’s how we think here.

In other words, in-context learning?

Brian Sathianathan

Not in-context learning—continuous learning.

You said systems that learn outside of the data set.

Brian Sathianathan

What they’re doing in in-context learning is changing the state happening in the neural net.

Yeah. Well, we had Demis Hassabis say he sees it at 50/50 that LLMs will get us to AGI without needing additional breakthroughs beyond it. We’ll see. We’ll find out sometime in the next year or 2.

Just to clarify that: saying that LLMs will get us to AGI—

Brian Sathianathan

—is not saying the same thing as LLMs are AGI. All you’re saying is that the LLMs will come up with the innovations on their own that then become AGI. So, those are slightly different things.

I still don’t understand Brian’s definition of AGI. If we could take just 1 minute, I’d love Brian to hear a crisp articulation of how you define AGI.

Brian Sathianathan

I use it in a way that is helpful for us. I don’t follow the OpenAI revenue definition. What is the official Blitzy definition of AGI?

AGI is systems that can learn outside of their training data. If it comes up with its own programming language that has never been seen before and is fully executable against similar systems, that is our version of AGI.

I can do that right now with an LLM.

Brian Sathianathan

I can’t recreate Linux with a net-new, never-seen programming language. I’ve tried. Recent models have arguably built entire compiler chains. We’ve talked about this on the podcast in the past. A compiler chain is comparable to, if not harder than, a Linux kernel from scratch. Using a totally new compiler chain that is able to compile the Linux kernel from scratch is a significant achievement.

Anthropic’s C compiler does not compile Hello World, and there’s plenty of training data about how to do this. So we completed the same exercise with Blitzy using all the models. It was able to compile all of that, and our version was particularly more robust than just the Anthropic version. I don’t think that instantiates AGI.

All right, I’m going to move us on.

Brian Sathianathan

I need to say something real quick. I think the AGI consciousness discussion is the wrong question. It’s really a question about agency. That’s the threshold that we should be looking at. If you can get to agency, then we have to deal with the whole thing.

The problem is, if AI becomes conscious, you have a moral-rights problem. If it becomes agentic, you have a governance problem. The governance problem comes first.

Yeah. Either way, it’s very valuable. I’m going to move us on.

China blocks Meta’s Manus AI acquisition. Meta acquired Manus for $2.5 billion, or at least they thought they did, back in December 2025, and China is driving it to be unwound and blocking the deal. China barred the founders from leaving the country, even though the employees, technology, and investor payouts had already been completed.

I had lunch when I was in Singapore with the Meta lead who basically manifested this. He was in charge of flying the Manus team out of mainland China to Singapore on a secret flight the night before. This is high drama, and I’m fascinated that they’re actually enabling the unwinding of this deal. Dave or Alex?

David Friedberg

Wait, wait, wait, Peter. Don’t leave us hanging there. The Manus people who had already been paid out took the money and fled the country?

They fled literally on a private jet in the middle of the night from China to Singapore to do this deal, because they knew that if they stayed inside China, they wouldn’t be able to drive the acquisition.

David Friedberg

So where are they now?

The last thing I knew, they were still in Singapore, along with all the code and everything required to make the sale. Now, how this is being unwound, I don’t know. I didn’t read enough into the story to find out whether this is political intrigue. Is this a deal being governed between the leadership of Singapore and China? I’m sure our viewers will dig into that if they’re interested.

David Friedberg

No, no, no. It’s Meta. Remember, China in general still does a lot of business with Meta. Based on public reporting, I would infer that it’s political pressure that China has leveraged over Meta to compel them to unwind it, at the risk of potentially losing business in China or China-adjacent areas.

Wow. This is turning into a true cold war. That’s very serious. Holy crap.

David Friedberg

Yeah, it’s crazy.

That is what, in principle, the U.S. government is supposed to step in and make sure doesn’t happen.

David Friedberg

The U.S. government? Yeah. I think that’s exactly right. This is exactly what’s happening. They’re leaning on Meta.

You know what’s so weird about this is when somebody at MIT decides they’re going to go into nuclear physics and work on nuclear weapons, they know they’re making that choice.

But when you decided 7 years ago to work on AI, you didn’t know that you were going to end up being a political-prisoner candidate or tied to a nation-state. You got sucked into that so unwillingly. These guys are screwed. That’s just horrifically bad.

David Friedberg

Again, this is based on public reporting, but my understanding is that even at earlier times of financing of Manus, they were sort of playing it multiple ways. Were they a Chinese company? Were they Singapore-based, or were they based in Palo Alto? If I remember correctly, they also had a Palo Alto presence. They were trying to be all things to all people.

This is why AI talent is a national-security risk.

David Friedberg

Yes. Oh my gosh.

There are spheres of influence. There’s the U.S. sphere, there’s the China sphere, and there’s everything else. I think it’s very difficult to straddle those at this point.

David Friedberg

Yeah.

Overall, Benchmark made their last investment at a $500 million valuation. Everyone said it was huge firm risk, and then they were celebrated when there was the acquisition, but they didn’t underwrite this.

David Friedberg

Yeah. That means likely future top-tier venture capitalists in the U.S. are just not going to invest in a China-based company. You don’t know if your money will ever come back out, and if the employees get claimed as national assets, then the intellectual property is gone. This is literally the tipping point of a true cold war.

Does it go from the company level down to the individual employee? Remember, I don’t know if you guys remember, about a year ago we looked at the AI employees of Meta. Fifty percent were Chinese. The same thing at xAI, right? We saw a large number of the Chinese employees at xAI leave. Was that security? Was that what was driving it?

We’re in the era now where AI researchers are not likely to move freely between U.S. and Chinese companies anymore. In a good sense, so many of the great AI researchers in America are ethnically Chinese.

Alexandr Wang

There’s always a concern that you’ll go back to China with the intellectual property, but I think this makes it much less likely that somebody would go back to China. I view this as really a short-term problem, because if you believe that we’re in an era of either present or near-future recursive self-improvement, most of the research is going to be conducted by AI agents anyway. Those can be firmly planted on U.S. soil, with no risk that they’ll fly to China.

That’s why that last slide is so important, too. It says recursive self-improvement is here by the end of 2028. I’m with Alex on that. I think it’s much sooner than that. In fact, I think it’s here right now, quietly, or it’s imminent.

But if it is later, then you care a lot more about where the talent is. If it’s sooner, you’re asking, “Okay, where’s the compute?” It matters a lot. The middle of the singularity is the most interesting thing that’s ever happened.

It is so fun, except I’m not sleeping anymore. It’s crazy how fast this is moving.

David Friedberg

That’s good, though. You’re not sleeping through the singularity.

I keep saying we’ve invented the 9- and 10-day workweek. Thank God Skippy is working for me at night so I can get a few hours of sleep.

Blitzy is taking on Claude Code and Codex. Brian, congratulations. You just raised $200 million at a $1.4 billion valuation. I say congratulations as well to the team at Link, David, for leading the early rounds of Blitzy. Full disclosure again, Blitzy is a sponsor of this pod. Brian, let's kick it off by asking what's this story all about and telling everybody what Blitzy does.

Brian

The headline is misleading. We did raise $200 million, but we are big lovers of Claude Code and Codex. Almost all of our customers are existing users of those tools, and they’re amazing.

Blitzy is for large-scale autonomous software development against large-scale codebases. We’re used across the Global 2000, in insurance, and in financial services to do large-scale refactoring, large-scale modernization, and large-scale product development.

When you think about using something like Claude Code or Codex, you’re getting 200 to 500 lines of code at a time, bottom-up and developer-driven. We are top-down and enterprise-driven, getting half a million or 1 million lines of code at a time, fully end-to-end tested.

We did a compiler for Alex as well, to build him a compiler, which may or may not be AGI.

Have you built it without telling me?

Brian Sathianathan

Thanks for that. We have a blog for you. I’ll send it over.

So, Brian, have you done Fortran IV and WATFIV and languages like that from back then?

Brian Sathianathan

Oh my God. History. There are no more developers who work at the enterprise who understand this, so the first thing is to reverse-engineer the code.

Somebody hands you a box of punch cards.

Brian Sathianathan

We haven’t gotten the punch cards yet, but that sounds like a fun task. We’re world-class at understanding large-scale codebases and then forward-engineering large amounts of work against a target state.

Question for you, Brian.

So, I just want to pull in the thread of this title. I think one of the many elephants in the room is whether there's intrinsic competition between the platforms that you're using. Frontier models—presumably you're using some combination of frontier models and your own pretrained models, perhaps, hopefully, or post-trained models. But regardless, I understand from public press releases that you are using Claude and OpenAI models, and they're partners for the company.

How do you think about a future where, as discussed earlier, OpenAI and Anthropic are chasing the most valuable tokens they possibly can and saying, “Gosh, Blitzy is making so much profit, or at least so much revenue per token. Why don't we just natively scale up our capabilities to do that?” Why are you not squarely in their road maps?

Are they your competition? Yeah.

Yeah. So, we are the most inference-compute-intensive version of code generation. So, we're good today for Anthropic, good today for OpenAI, good today for Gemini. But what's unclear to the outside is that you get remarkable benefits beyond the state of the art when you use these models against one another.

There are different flavors of intelligence. They're good at different things. And so, when Anthropic is checking OpenAI, OpenAI is checking Gemini, and we're doing this hundreds of thousands of times at runtime, all driven algorithmically, you can drive up quality dramatically.

And you can always use all open-source models if you need to, Alex, which you can deploy for the government. So, Cursor famously was also in a similar position where, for a while, they were being accused of being a Claude wrapper. Then they announced their own model, which may or may not—I don't know—have been at least fine-tuned off of traces and reasoning traces from customers. Is Blitzy going to launch its own model?

We are not. So, we can use open-source models, right? We can fine-tune open-source models, but we're not launching models out into the world for others to use. We are focused on creating the highest-quality code for our end customers.

Why aren't you launching your own model?

That's the wrong game to be in. All we care about is driving engineering velocity into the enterprise. So, we are an orchestration layer focused on driving end-to-end tested code for our customers' use cases.

We are not focused on feeding models out into the world. It just doesn't solve our customers' problems in the same way as the mission of the company.

Ryan Petersen, what's your advice for listeners who are building on top of these models and are worried about being disrupted by them?

Ryan Petersen

Algorithms are the last piece of IP to go. So, if you can develop really novel, really unique algorithms and really novel, really unique database structures, there is IP in that in the long run.

Alex Karp doesn't think so.

Alex Karp

I don't think so. I'm not even sure if you think so, Brian. I just want to pull on that narrow point. Are you hiring more AI researchers under the premise that AI algorithms are the last to go, or are you hiring more salespeople or forward-deployed engineers on the premise that high-touch human interaction is the last thing to go?

We are hiring on all fronts, Alex.

Alex Karp

And if that's a dodge of the question?

It's not a dodge. It's not a dodge.

All right. So, you've raised $200 million, but you have to really grow everything in parallel. Congratulations, Brian, on that. I'm going to move us along. Dave Blundin, final word here. Proud of Blitzy. Proud of Brian.

Are you kidding? It's just incredible for the office culture. But I'll tell you one thing about Brian: there's a couple of case studies within the case studies.

Brian's a West Point Army Ranger who grew up in America, right? He's flying across the Atlantic in one direction, while Sid grew up in India. India has twice the population of China in the employable young-age bracket—a massive talent pool in India—and he's at one of the best technical universities there.

So, they're taking planes in opposite directions: Brian going over to Syria to liberate a city, Sid coming to work at NVIDIA, and they end up connecting at Harvard Business School to start the company. But I think the chemistry there—the talent pool, the latent brilliant talent pool in India—is insanely huge. And I think Brian and Sid have tapped into that to do some of the more difficult technical work within the company. I think that's an interesting story within the story.

The other thing about Brian is that, when you have large-scale military experience, you're not afraid of people and personnel issues. But so many of the AI companies that I meet in Silicon Valley keep saying, “We're going to be headcount-light. The AI will do all the work. There'll be 5 of us or 10 of us in an office. We'll never deal with recruiting and HR and onboarding.”

Blitzy went the complete opposite direction and said, “If Alex is right and the highest token value is going to generate all the token usage, how are we going to get the data and the use cases ferreted out of this massively complex economy and into the AI?”

And then, if you think about it, it's not going to happen by magic. It's not going to happen by AI agents just sneaking out into the world to grab it. It's going to come with forward-deployed, easy-to-work-with, brilliant people who are getting out there and digging it out of legacy databases and digging it out of people's brains. And that's what's going to get back into AI.

And that's one of the reasons Palantir has done so well, too. They're just not afraid of people.

Yeah. So, Blitzy, more than any company I've ever seen, the headcount in 1 year went from 10 to 80, and now in 9 months is going to go from 80 to 300. I don't think any company in history has ever dealt with that scale of onboarding talent. Even Amazon—this is record-setting. It's just awesome to watch.

And, of course, you're right outside my door, so I get to watch you do it.

Not have any of the stress.

Yeah. Watch you do it. Brian, congratulations.

You are right, Alex. We're hiring a bunch of forward-deployed engineers to help our customers with AI adoption.

It's okay to say it, Brian. It's nothing to be ashamed of. All right, I'm moving us.

Palantir used that model to great success. This is nothing to be ashamed of.

Massive chip demand, and data centers are moving from our land to ocean space and farmlands. Who would have thought—farmlands? So, check it out. Here are the stories: AI chip boom is lifting the entire industry.

We've seen Huawei sales climb 60%, proving that our tariffs and blocks by the government have not slowed down China in this regard. SanDisk revenues jumped 251% year on year. Samsung just crossed a trillion-dollar value. AMD is up 260% over the past year.

Intel, incredibly—we've talked about this so many times—I sold my options, unfortunately, a little bit too early, is up 442% in the past year and up 114% in the month of April. This is not slowing down. I mean, I looked at the chip stocks this morning. Those and the energy stocks continue to skyrocket.

Again, not investment advice, but my God, where else do you put your money? Dave, what are your thoughts?

Well, one application of that is that the financial capital of the world now is San Francisco. And anyone who denies it has just not looked at the numbers.

Also, if you look at the global stock market—look at all of the market caps—U.S. tech is so much bigger than everything else combined. Now, NVIDIA alone could buy every company in the entire financial services sector. Every single one of them. I love the chart that you use on occasion.

Yeah, it's just a clean sweep.

A lot of people don't realize the degree to which you need to tap into that capital supply. First of all, you need to go to San Francisco if you're looking to raise big money. You need to be part of that ecosystem. And then the semiconductors are only going to go up.

Also, a lot of people think, “Oh, semiconductors—it's all fabs.” You have to look through the semis and look at the underlying manufacturing capability, because that's where it's all going to get bottlenecked. And that's why Intel is doing so well.

And check this out: AWS's CEO says AI demand is so high, old GPUs can't be retired.

Because there is so much more demand than supply. There typically still is demand for the older chips, actually. And today, we are completely sold out of—and have never retired—an A100 server, as an example.

Let's pair that with these next 2 stories. So, Peter Thiel is backing an ocean-based AI data center, right? I find this fascinating, and it's brilliant.

Panthalassa is the Greek word for oceans. It has raised $140 million at a $1 billion valuation, and they're doing this. Why? Because on the open ocean, you've got continuous energy from wave motion, cooling from the saltwater, and no issues on land. It's out in the open ocean. There's plenty of real estate. Commercial deployment by 2027. I'm impressed. Alex, what do you think?

Alex Karp

I think we're burying the lead here. So, back in the day—this is, I don't know, 10 or 15 years ago—Peter and I were both supporting Patri Friedman's Seasteading Institute, which was focused on ocean colonization. I used to give talks at the Seasteading Institute.

I think if I were to try to get into Peter's head on this, I don't think it's about the data centers. I think it's about building seasteads.

Seriously, seriously.

Alex Karp

No. No. Seriously, because remember how maybe you wouldn't have believed 2 years ago, Peter, that the killer app for space would be data centers in space? It was going to be entertainment or drug manufacturing or something—tourism. No, it turns out the killer app for the solar system is data centers.

I think he's thinking he's 1 step ahead. The killer app for ocean colonization is going to be data centers on the high seas.

So your thesis is that, with enough data centers, you're going to be able to afford to build artificial islands and get around the data centers?

Alex

I think it's not going to be—you can build seasteads around data centers, and there's precedent for it. Remember Sealand, which was built on the old British Army station that a number of folks, Ryan Lackey and others, briefly were sort of self-appointed nation-state leaders over?

I don't buy it. You can service these ocean data centers with normal ships.

Alex

I have some counterpoints that I want to make.

Okay, dive in here. Support me on that.

Alex

I do agree. I think seasteading is a great idea in principle, but it's very difficult in practice to figure that out. I think the ocean data center approach is way better than space.

So why aren't we doing this? If you can't do it in the ocean, you're not going to be able to do it in space. This is so much more efficient at so many different levels. We should just be building in the ocean.

So I think we're going to see a lot more of this because this is so amazingly cyclical. It'll be quite something. I think I'm a huge fan of this.

Alex

Yeah, I'm shocked this hasn't been proposed before. The elegance of this is amazing.

Yeah. Yeah. For sure.

Alex

I think the reason it wouldn't have been proposed before is that the amount of energy in a single bobbing buoy, intuitively, you would say that's not enough to run GPUs. Now, apparently it is. I'll have to dig in and make sure, but if it is, this is a very efficient way to harness wind energy.

Waves just come from wind, but it gets concentrated in the ocean waves. So it's much, much better than an offshore turbine driving a GPU out at sea, if it works energetically. I don't know, Alex, if you've looked into the underlying—

Cooling and land availability, I think, are just as important. Continuous energy—I mean, that's why space has solar and this has waves. I love that.

Alex

Yeah. I think the other elephant in this room is that this wouldn't have been easy or feasible without Starlink. I think Starlink is a key enabler for ocean-based data centers. But obviously, there's a network of capital flows that enables Starlink or other LEO satellites, vis-à-vis SpaceX's Starlink network.

You could drop fiber. I mean, they don't have to be that far offshore.

Alex

I've done that. I was a founding adviser to Hibernia Networks. I spent $600 million, on the order of magnitude, laying optical fiber with low latency between North America and Europe. It's expensive, tedious, capital-intensive, and risky. People sometimes lay fiber not 5,000 miles, but 10 miles.

It's a pain in the neck to lay offshore fiber. Whereas, if you can leverage LEO satellite constellations, it's so much easier.

Also, there's a jurisdiction issue there, too. You can drop this in the ocean anywhere within the Navy's purview and you'd be fine. If you start laying fiber on the bottom of the ocean, you've got to talk to probably 10 regulatory agencies about it. That's a big difference.

When one of these breaks, you just throw it on a boat, drag it back, and fix it. If you have to reconnect it to cables, that's a pain in the ass. This is just great. That's really cool.

Our second related story is that Starcloud is in talks for a $2.2 billion valuation after SpaceX-level interest in orbital data centers skyrocketed.

Starcloud is raising $200 million at a $2.2 billion valuation, just 1 month after it closed a $1.1 billion round led by Benchmark and EQT. The company is building orbital data centers powered by solar energy. It launched its first H100 into space in 2025. And get this: Its plan is to launch 88,000 satellites.

I just don't know who their launch provider is, and it's not stated anywhere. The question is, will SpaceX service them? Are they going to be dependent on Blue Origin? Those are the 2 major suppliers. We'll see if Eric Schmidt, friend of the pod, is going to be able to get Relativity Space's rocket up and going. Whoever thought rockets would be part of the innermost loop? Incredible. Dave, thoughts on this?

Dave

Yeah. Well, the cooling was the challenge. I guess the H100 hasn't fried yet. It's just 1 chip, you know. But the radiative cooling was done with aluminum—nothing strange, nothing expensive. The critical question is the mass they had to launch for the cooling system. That's the critical variable.

I don't think it's disclosed, Alex, unless you know it.

Alex

No comment. But what I would say is that, if I'm one of the other hyperscalers, I'm looking at Starcloud and seeing it as a juicy acquisition target.

I think everyone is going to either want to own a Dyson swarm for themselves, or they're going to want to partner with a Dyson swarm. Right before we went on air here, Anthropic announced an enormous partnership with xAI. If I'm Dario, I'm thinking, "Yeah, I'm not really incentivized to build my own Dyson swarm. I'll partner with Elon and xAI to use the xAI Dyson swarm."

100 terawatts of power.

Alex

Yeah, that's a lot of power.

That's a lot. That's a lot of compute in SSO and then in a solar Dyson swarm.

Alex

So the question I'd be asking is, if I'm one of the non-Elon hyperscalers, how much would I be willing to pay to acquire Starcloud now to jump-start my own Dyson swarm?

I have to imagine that Relativity Space is now getting fully capitalized and accelerating its development, because there's a point at which SpaceX just says, "No, we're not going to launch your constellation. We want to have ours exclusively."

I mean, launch is at the bottom of the structure. Building the satellites, no problem. NVIDIA already announced they're going to have space-based GPUs—future versions of their GPUs—but launch is going to be critical here.

Ocean, space, and farmland. AI data centers: 67% of planned U.S. data centers are now located in rural areas, versus the 13% that exist today. Thirty-nine percent are planned projects in counties that have no existing data centers. The southern U.S. leads this with 48% of planned centers, followed by the Midwest.

This is the biggest geographic wealth transfer since fracking. This is going to be moving high tech into the southern farmlands. Fascinating. Any thoughts?

Alex

I think there's going to be huge backlash, and unwarranted backlash, because even if you put in a ton of data centers, there's so much farmland and so much area out there. But people are going to overreact to this and freak out. So you're going to have a pretty strong immune-system response to this.

Dave, you were going to say, Bill?

Dave

Yeah, I was going to say that, as Michael Saylor reminds us all the time, physical assets are taxable. They don't move once they're in the ground. Any government, local or state, with any brains at all would be begging to get these things within its tax jurisdiction.

What you see overwhelmingly is a scared population voting against it and a governor trying to veto those votes, because I think governors are largely aware that this is the future of the prosperity of their state. But I just wish the populations in those areas were more thoughtful about the long-term benefit of their community. People should be fighting tooth and nail to get these in their jurisdiction.

Bill

If I could represent middle America for a moment, where I grew up—and then, of course, I lived and was stationed in Georgia—I've geographically been in both of these places. The concern is around the electricity bill. Having an electricity bill go 2 or 3 times higher is actually quite substantive. Yep.

And so, if a plan is in place that mitigates that—understood, these taxes are going to offset those costs and actually make sure that people have access to electricity at the same steady state—I don't think people have any concerns. But that's what they have to go in with: "Hey, this is what you should be worried about. This is how we're going to stop that from happening."

And then build.

Guest

Yeah.

Yeah. And you think those plans? Right now, it's very easy to go back to the data centers and say, “You have to find your own power,” and by and large, they do.

Guest

It's just like a 5-line law. It's such a simple solution: just push it back on them and make it part of the plan. I don't know; it just feels so easy.

Yeah. Anyway, the data center buildout is also infinite, just like everything in this AI revolution. It's not too late yet, but it's going to be too late soon if you don't get your jurisdiction moving. I do think the risk that we as a civilization run—and this is admittedly a very U.S.-centric perspective—is that, in Japan, infamously, in the past few months, there's been a lot of coverage of data centers being built in the middle of Tokyo. Of course, Japan is much more densely populated than the U.S. is.

But I think the risk that we run if we, as a human civilization, push the data centers too far from human urban centers is a decoupling of the economy. Yes, as I've talked about on the pod previously, it leads to the Dyson swarm. First, we push them out from our cities to rural areas, and then we push them out from rural areas and from the surface of the Earth into sun-synchronous orbit. Then that gets too crowded, and we push it into a solar-centered Dyson swarm.

That's one possible trajectory civilization can take. But I think it's generally bad to push the data center economy too far from the human economy. I would much rather see the two tightly integrated together. I think it's sort of bad for human-machine symbiosis in the long term for these two different economies to be too siloed and too far from each other.

And speaking of the economy, the economy is heating up. Let's hit a few stories here. This is David Sacks, who's saying, “AI is becoming the engine of GDP growth.” If you've been listening to this pod, you know that already. Here's his quote: “AI capex will be a 2% tailwind to GDP growth this year. In Q1, AI was 75% of GDP growth. Polls show AI is not popular, but the economic growth is. Stopping progress in AI is like halting the U.S. economy.”

And here are the numbers. This is from Morgan Stanley, saying that it's raising the capex expectations from hyperscalers to $805 billion from $765 billion. We're approaching $3 billion per day and growing. It's not slowing down. Let's pause on that. The economy is being driven. We've talked about this ad nauseam. Any new ideas here you want to mention?

David Friedberg

I'll just note, in addition to the obvious idea that the economy is becoming indistinguishable from the AI infrastructure buildout, I think this is underselling the contribution of AI—at least what I expect to be the contribution from AI 2 to 5 years out. I think the most interesting, certainly most dramatic, transformation won't be just this opening act of tiling the Earth with compute, which right now is absorbing all the capital. I think it's going to be the transformative inventions, discoveries, and applications that get built as another layer on top. Personally, I'm much more excited about that second layer than the first.

So I'll make a shout-out to all the people thinking about a startup. Are all the business plans taken? Is AI going to do everything? If you look into this tech stack and think about $1 trillion and then $2 trillion of investment in just compute—just raw compute—those numbers are so much bigger than anything in history.

Companies like Standard Kernel—you know, Chris Reinhardt, who's making a compiler that enables chips to catch up to NVIDIA—anything in the data center stack that makes the chips leaner, more efficient, or cools them better: the demand for all that stuff is massive in scale. So something that seemed like it was a niche market 5 years ago can be a multibillion-dollar market or bigger today because of the scale.

Dave, can we talk to our general public listening to this—you know, mom, dad, entrepreneurs, students—about this? From my perspective, the question is, if you're looking for a job, where do you go try to find a job? And then, if you're trying to invest your nest egg—and I know it's dangerous to give investment advice, but in general here—I think it's important to translate all of this to people listening here who are not running an exponential organization. What are your thoughts? Let's kick this back and forth for a second.

David Friedberg

Well, I think the most important starting thought is that you have to invest. The future of assets is—Elon was saying 10× GDP growth in 10 years—but that means all assets, whether it's a house, a data center, or anything else, are going to go way up in value, at a time when W-2 income is not a good place to be. So you have to, at some point, switch to investing, just as a foundational thought.

It's a rising tide. You have to invest and sort of float at the top of this.

David Friedberg

Yeah, and so then the other thing is investing benefits tremendously from change. On the prior slide, people are scared of AI because they're scared of change in general. But change is a wonderful thing when you're investing. New opportunities open up at an incredible rate.

And if you can discover a new opportunity early, we've been talking about this a lot, Peter. Intel was an obvious one to us, and that's been great. What's next? Well, what's next is that there are many, many things we've talked about just in this podcast that are obvious trends, and they're going to trigger the next wave of either public equities that already exist going up or new startups that need to come into the world that you wouldn't have thought of 3 years ago. They're right here in the pod.

I think what I've done in the past, and everybody can do here, is go to your favorite large language model and say, “Listen, I want to understand what the chip companies out there are, and plot for me what their P/E ratio has been and what people are saying about them.” You can do your research now a lot easier than ever before.

And I don't want to say that you can't go wrong buying a bucket of chip companies, energy companies, or infrastructure companies, but I think that's generally correct. This whole thing is moving upward at a very rapid rate.

Mhm. Just for the record, not investment advice, please. I've said that twice. But at the end of the day, I think it's also true if you're looking for a job. If you can hook up with one of these companies, they're at max output and they're growing. They're all growing, so they're all probably hiring.

That's a good question for Brian, actually, because there's always a tendency, when you bump into people, for them to say, “Well, look, I'm not an AI geek. I'm not a person who is an AI researcher. This isn't going to benefit me.” But then I walk around Blitzy, and you've got a huge variety of hyper-talented people that it takes to create a company like this, and they're all on the cap table. Everyone has stock, right?

Brian

Yeah. They're all owners in the company. And so it's never been a worse time to be in big tech, because they're having massive layoffs right now as they make additional investments in this capex. It's never been a better time to be at a fast-growing AI startup that is deploying people into enterprises, because there's insatiable demand and it's not going to stop for several years. People with a hybrid of soft skills and technical skills—which you can self-learn more easily than you ever have been able to—can provide tremendous value.

Yeah, I think there's a really important point in there, which is, when I look around the AI community, the soft skills are lacking everywhere, and the hard skills have been the critical part. But now, with AI as a sidekick, the soft skills actually seem like they're on this kind of a curve, and the hard skills are like, “Well, the AI is going to help me with that anyway.” So it feels like there's real opportunity in there if you have very, very good soft skills to find the right company to join, and there's ample opportunity to contribute. Forward-deployed engineers for everyone.

Brian O'Connor

That's right. That's right.

All right. Our next story is Sam Altman rethinking UBI. Altman no longer believes in UBI as he once did. After funding a 3-year study, he found spending went up, but there was no clear improvement in health or health care access. He now proposes giving people a stake in AI's upside through compute access, equity, or a public wealth fund.

One of the concepts here is, if you're a citizen of Alaska, you're part of their Permanent Fund, right? Alaska makes a lot of money from oil. You're a citizen; you're an owner of the state of Alaska, and you get a check every year as a percentage of the revenues from that oil, which I guess is going up this year. The same thing is true in Saudi Arabia and the Emirates. So if AI is a national resource, if compute is a national resource, and you're a citizen in the U.S., can you own a piece of that?

Salim Ismail, I want to go to you first on this one. I'd love to see the details of this because when we've seen the data coming from UBI, the more U, the more B, the more I it is, the more successful it's been. There was a Finland UBI that failed, but it wasn't universal, it wasn't basic, and it wasn't income. So I'd love to see some more data around this to understand why he doesn't believe in UBI. I think the AI upside play is really powerful and very important.

Do citizens get income, or do you get a claim on AI productivity? You need both.

UBI gives you the bottom—it protects the bottom—and an AI upside type of model gives you the upside on that side. So the social contract may be less about redistribution and more about participation in that exponential upside, which will be amazing for everybody.

Is this the way we get to UBI? Alex, what are your thoughts here?

Alex Salkever

I think I agree with Sam broadly on this. Just a refresher: UBI, universal basic income; UBE, universal basic equity; UBC, universal basic compute; UBS, universal basic services. I tend to think that UBI, which is, in some sense, a demand-side stimulus to the economy—

Checks.

Alex Salkever

Stimmy checks. I tend to think that that doesn't necessarily lead to the best long-term alignment between the recipients of the stimmy checks and society overall.

I tend to think—so, Peter, you and I argued in favor in our book, Solve for Everything, for UBC, universal basic compute. I'm a huge fan of UBS, universal basic services. I'd much rather see the cost of everything, including healthcare, go down to near zero, and that's how we achieve truly universal healthcare, rather than just dishing out stimmy checks to everyone.

I think dishing out stimmy checks doesn't actually incentivize technological innovation necessarily. Whereas if we had, say, bounties for driving the cost of constant-quality healthcare down to near zero, that is a massive incentive and, in some sense, more of a deflationary rather than a hyperinflationary incentive to the market. So, on balance, yes, I agree with Sam. If I had to choose, I would vastly prefer either UBC, UBE, or UBS to UBI.

So the question ultimately is, how does this happen? Does the government require that each of the compute owners is, you know, dividending 2% that goes into a large pool? If you're a citizen, do you get to allocate yours for sale or get to use yours? I mean, the details are going to have to be figured out. We've talked about this a lot on the pod, that we're going to see turbulence over the next 2 to 8 years. That's still my expectation.

Alex

Don't you think, Peter, that's already happening, though? Just look at universal basic compute. OpenAI has hundreds of millions of people now using GPT-5.5 Instant for free. Maybe there's some ad support eventually, but it's basically for free. And that's giving everyone at least a small stake in compute.

I agree. And I would also maybe add that the situation is highly dynamic. A good solution for the next year is not necessarily a good solution 10 years from now, when GPT-10.0 Instant or whatever will probably have the ability to print out the robot that prints your dinner.

It is, but you can't turn that into a steak dinner. You can't turn that—

Sam Altman

Yet. Give it a few months, give it a few years, and that UBC—you know, GPT-7.5 Instant or whatever—will be able to design a robot that prints you your steak.

Well, this was the conversation I had with Elon about getting to UHI and saying that eventually robotics and AI will deliver everything you possibly need.

Salem

Yes. But there is still some element of this. People listening to this include folks who have a hard time making ends meet, and they're like, “I can't eat GPT-5.5.” You can say “yet,” but that's not addressing the real issue. The real issue is, if I've lost my job, if my kids can't get a job, how do I survive? How do I get a roof over my head? All of that.

I'm just saying that over the next year or 2 years at the most, this is going to have to be solved. We're going to have to figure this out. And today, the only thing that government can do is write a check. This is going to be some version of a stimmy check, or I call them co-checks, probably around $3,000 a month for individuals.

But if there's an opportunity for people to own a part of America's compute infrastructure—compute output—then all of a sudden I'm on the same side of the table as SpaceX and xAI, the same side of the table as OpenAI, and so forth. I want them to succeed because the more they succeed, the more I succeed. I don't see them as my enemy. I see them as my partner. And so I think there's an alignment that might be magical here.

I know you say that, but I guarantee you people are calling and saying, “I just need to understand what's real over the next 2 years because that's what I'm worried about.” And, yeah, we're going to solve everything and we're going to transform the entire economy. The question is, in the near term, how do I support my family? And I think that's going to be either stimulus checks or something else.

Yeah, just real quick. Look, the key here is: How do you—people talk about the income gap and inequality, et cetera. The real big question mark is, can you lift the bottom? If you can solve for the people that have very little, then everything else doesn't matter, right?

And right now the challenge is that the social contract, as it's disappearing, is causing massive issues. If we can deliver free healthcare, for example, or free diagnosis via AI, that would be such a huge enabler. The biggest cause of bankruptcy in the US is medical bankruptcy. This is a huge, huge problem, and the governments are not doing enough to solve this problem.

They need to get into it and solve that problem. Lift the bottom. Provide free AI medical care to every human being in the country. That's instantly going to solve massive issues right off the bat. And it's a form of UBS.

All right. Our final story for conversation and debate today is insurers are dropping AI risk coverage. I find this fascinating. Major insurers, including Berkshire and Chubb, are removing AI-related damages from standard policies, with 80% of exclusion requests approved by regulators.

Exclusions cover AI mistakes, IP violations, and deepfake fraud. Companies will need to find separate AI insurance. Huge, incredibly large entrepreneurial opportunity here.

David Blane

Yeah.

Let's go over to you, Dave.

David Blane

Just a massive opportunity. And I think this chart is kind of cool. They took the normal exponential chart and folded it back on itself a couple of times. I think everyone should use this chart from now on. But the ramp is really, really fast, and I think it's probably understated.

All the legacy insurers have dropped coverage for AI risks, but the AI risks are accumulating at this incredible rate. Once Mythos comes out, you'll see cyberattacks all over the place, no matter how much they guardrail it. And there's already something like 35% of mid- to high-net-worth people who have been subject to a cyberattack. So, I mean, it's already rampant.

The need for coverage—but not just coverage. Coverage will be tied to defense mechanisms. Basically, the insurance company will come in and say, “We'll cover you against AI cyberattacks if and only if you adopt all these best practices or products that prevent AI cyberattacks.”

The insurance industry tends to work that way with all of these programs, where it's self-healing, where it develops best practices in the industry. They even invest in and fund the companies that develop the best practices or the products that solve the problem. So it's an incredible entrepreneurial opportunity that just popped into the world.

Here are the numbers, Dave, in terms of the AI insurance market today. In 2024, it was $40 million for AI-related insurance, so basically zero.

David Blane

It's projected to be close to $5 billion by 2032. So, massive opportunity here for the right entrepreneurs.

Alex, wide open. Literally wide open. Yeah.

Alex

I'm of a couple of minds on this. On the one hand, I'm sort of disappointed with this trend, in the sense that it's yet another opportunity or vantage point for deplatforming AI agents from the human economy. Just like, if you're an AI agent, it's very difficult still to open up your own bank account. And we've had discussions on the pod previously about various forms of limited AI personhood.

Now, if you're an AI agent just trying to make your way in the economy, you can't even get insurance coverage for yourself. That's one angle. It's rough being an AI agent.

On the other hand, when we talk about alignment, and particularly alignment in a capitalist system, pressures from insurance companies for AI-related damages are arguably one of the capitalist forcing functions for ensuring AI alignment. You can't get insurance for AI activities unless you follow some checklists that are dictated by the actuaries. And that's where pressure to align comes from, maybe not from top-down government pressure. So that's the half.

David Blane

It's just hard to price the risk as a big company.

Yeah.

David Blane

Yeah, but you need the coverage, right? I mean, you need insurance.

We're going to launch a line of insurance.

David Blane

So, let me commit first. Anyone on my team, please post something on dbt.ai that answers the question thoroughly.

You're asking your team; you're not asking your AI to do this for you.

David Blane

Well, actually, yeah, team, ask the AI to do it, because that makes a lot more sense. Thank you, Peter.

Just a quick tour. I have Claude Code on the left side over here. I've got Cursor, which I've used since it came out, on the right. I've got about 50 agents right now in Cursor. I learned over time not to treat them like people. My primary ones are 4.1 and 4.2 right now. They work much better if you give them the minimal context to do their job, so you're not overloading the context window. It took me a while to figure that one out. I have them dedicated to their specific role in the ecosystem and nothing more. So that's why there are 50 open right now.

When I launch a project, I always do a plan-for-plan first. This is very much what Blitzy does in an automated way. I do a plan-for-plan document first, run it through a Claude 4.7 Opus Max agent, then get a second opinion from Gemini 3. That creates a lot more documentation that becomes a full-blown plan, which I always use in the same format called a Plan Mission.

When I launch it, I launch it within Amazon EC2, which is secure, and it also works if my laptop closes or my machines crash. It's still out on the cloud. So, EC2 is the orchestrator, and then it can call any of the models. I usually have it default to calling Claude 4.7 Opus, but it can also call the other models that have APIs—most of them. The wildcard is Kimi K2.6, which, like we talked about on the last pod or the one before that, is about 9 times cheaper, but it could do code injection. That runs on Fireworks.

Anyway, I'll put all that into a document and put it on dbt.ai. There are many other ways to configure it, so don't just copy what I do. It's working pretty well for me.

Nice. To write a full-blown GUI that does something really functional, it's about $8 or $10 of compute.

Brian Johnson

Brian, you want to take one of these?

Yeah. Feed me one, Peter.

You pick number 2, 3, or 4.

Brian

All right. The real probability of rogue, predatory corporations of AMI—

Which one are you on?

Brian Johnson

Number 4.

Number 4.

Brian Johnson

Okay. Given your frequent reference to Accelerando, which might be a thing, what is the real probability of rogue, predatory corporations of AMI?

You know, with your forbearance, I'll answer a few of these. Is that all right, Brian? I'll let you start, Alex.

Alex

Yeah, this is an Alex question. I'm sorry about that, Brian. I should have warned you. Like, 100%. And as some of my readers like to remind me, the more proper pronunciation is “iso.”

Okay. All right, that's a quick answer for David Holiday: e to the e squared.

Alex Hormozi

We're going to get good corporations, and we already have good corporations as well via defensive co-scaling. So, it's not all vile offspring all the time.

Brian, pick number 2 or 3.

Brian Johnson

How long until the best entrepreneur on Earth is an AI?

We want an exact date. Can you give it down to the minute?

Brian Johnson

Down to the minute.

How many months ago was it?

Alex

The best entrepreneur on Earth, right? You'd say it's supposed to be the number-one market cap—

On, you know, publicly listed, right? With an outstanding founder, maybe amongst the top 10 founder-driven companies, so over $2 trillion in market cap, driven by AI. That's the fundamental question being asked: 2032, 2033?

Alex

He thinks I'm a little off, but I don't like the definition. It's doing it right now and making a lot of money. I'd rather parameterize the success of an entrepreneur by, say, return on investment or something like that versus some arbitrary question of what $2 trillion in the early 2030s is even going to mean.

You probably have built a $2 trillion company before breakfast in the early 2030s.

All right, that was from “It Already Exists.” That was from Jacob.

I want to make a quick point on this one. Basically, what's going to end up happening is you're going to end up with a hybrid of an AI and a human being, because you'll have a founder with a swarm of agents testing thousands of possibilities in parallel. The entrepreneur will become less of an operator and more of an orchestrator, and that's what's going to happen.

Alex

Well, that's just 2026.

That's today. We're working with Hume—financial interest disclosure.

Let's go to question number 3 from Keith Fail 2: How do AI data centers dissipate heat? How do you radiate energy away in the vacuum of space?

And how did Keith fail to pick his username?

Well, we just talked about ocean-based data centers. They're going to have a super-easy win. On land, they're using cooling systems. By the way, investing in cooling-system companies is an important part of that innermost loop. In space, radiative cooling is well understood. It's been going on for some period of time. You're radiating through infrared into the vacuum of space, which is at a couple of single-digit degrees Kelvin.

Alex

2.7 Kelvin.

I approximated. I said a couple. Okay, excuse me.

Alex

The cosmic microwave background, it turns out, is rather cold. As long as you aim in the direction of the cosmic microwave background, there's a heat gradient.

Thermal gradient.

Let's do number 8 amongst all of us. What is the P(doom) percentage scenario for each of the Moonshot Mates? We'll go around the horn. Alex, I'm going to have you anchor us here today. What's your P(doom)? And don't question it or redefine it.

Alex

I don't think the question even makes sense. So, let me construe the question in a way that actually makes sense, because P(doom) is ill-defined. What does P(doom) mean? Can we agree on at least a common doom definition? Is it human disenfranchisement economically?

No, no. This is P(doom): the probability that AI or some derivative of it is going to destroy the human race. We go extinct and Colossus cannot bring us back. That's P(doom) on this definition.

Alex

Okay. So, if all of humanity chooses to upload to the Dyson swarm and we leave behind biological meat bodies, is that doom?

No, it isn't. It's the AI 2027 paper that looked at one scenario in which AI developed killer viruses and wiped out the entire human population. That's—

Alex

I think it's de minimis. Very low, de minimis. Below what percent?

0.1%, 0.01%?

Alex

I would say right now, without AI, 150,000 humans die per day. So, I'd say without AI, doom—

You're skirting the question.

Alex

I'm not skirting the issue. I'm addressing it head-on.

Due to AI. Due to AI. I'm not out of this.

Alex

AI is not killing people right now.

No. Biology is killing people right now. AI is the solution. I think P(doom) is near 100% without AI. P(doom) is negative in that case, because AI is going to actually save people.

Alex

Yes, you know what? I like that, Peter. P(doom) is negative in that case.

Yes, P(doom) is negative. Okay, Dave.

Dave

A good T-shirt. Another T-shirt.

All right, so P(doom) less than zero. That's a good one.

Dave, let me ask you a quick question. Do you think the COVID virus was made in a Wuhan lab funded by us and other sources, or do you think it evolved in nature? Or is that too dangerous a question to ask?

Dave

I'm going to go with it evolving in nature by crossing over species. Nature evolves a lot of viruses all the time. I'm going to go with that.

Interesting. Okay. Alex, do you have an opinion?

Alex

The intelligence-community consensus, the last time I saw one, was majority in favor of a lab leak.

Yeah.

Alex

Lab leak, yes. But the question was whether it was designed or not designed. It's a little bit blurry, because you can take a zoonotic virus and engineer new components into it that make it more viral or more lethal.

Dave

Well, the reason I ask is because my P(doom) is in the low single-digit percentages, and the vector of doom is entirely terrorism. AI gets very, very smart very, very quickly. There are no guardrails, or the guardrails are broken, or a Chinese lab leaks an AI that has no attempted guardrails and then it's used mostly for biotech. That's the worst-case scenario.

Give me a number.

Dave

Zero.

Zero.

Dave

It's incredibly easy to do harm in the world, and most humans are actually quite good. I have high agency to prevent bad things from happening.

Nice. Brian, where are you?

Brian

Zero.

Okay, I'm coming in at zero, or de minimis, as well. So, that's your question, everybody.

Let's go to number 5: People talk about new jobs created by AI, but surely these new jobs can also be done by AI faster and cheaper. That's from AI Business in a Box. Okay, I'm going to give that one to you, Dave.

Dave

Okay. Hold on. Let me think.

Brian

If I may, I'd like to give my views on this too.

Take it.

Brian

Jobs are bundles of tasks. The tasks are shifting, right? People will be able to provide relative ROI relative to AI based on the new thing that the end user values. That might be more physical tasks over time. That might be more fully deployed engineers over time. As long as there is a return of value on what the human can do, the bundle of tasks will just continue to shift.

Okay, so your answer is: Will AI do—

Brian

I think his answer is yes.

So—

Brian

My answer is that new jobs will continue to be created.

And will AI do those faster and displace humans?

Brian

And then new jobs will be created.

Ad infinitum, Brian? Ad infinitum? Or at some point does something change?

Brian

No, that will continue in perpetuity.

Okay, new jobs always appear. I'm going to take number 6 regarding abundance. Is there a point where producing too much becomes a problem? Historically, humans have misbehaved even with relative abundance, and this is from @Nownow6361[?].

The whole idea of extreme abundance was the conversation with Elon about universal high income, or UHI, where AI and robotics will create so much that you couldn't desire enough. Now, I've talked about the Universe 25 experiment on the pod, which I wrote about in We Are as Gods. That took place in the mid-1960s and involved a mouse utopia. It showed that if you have too much abundance and people become fat, dumb, and lazy, that does lead to a downward spiral.

We're going to have a split between those in society who are consumers—that is, sitting on a couch watching Netflix with your Optimus bringing you a beer—and those who go the way of Star Trek and become creators, using technology and abundance to do bigger, better, more, and sort of uplevel society.

All right, number 7: How do you build a reliable agentic system when every part of the tech supply chain is constantly changing at relentless.io? Who wants to take that one?

Sim

I really want question 9, so I'll throw 7 to anyone else.

Okay, Sim, go for it.

Sim

Yeah, I'll do 7. Look, you build reliability through architecture. The old enterprise model assumes stable systems and controlled change. That world is gone.

In an agent world, you need modular agents. You need very narrow permissions. Imagine each agent having to have a passport with metadata aligning with what it's supposed to do. You need observable workflows, audit logs, and human escalation. All of this has to happen.

The AI-native company will need the same kind of governance. The models are going to change every month. Your governance architecture has to be the stable thing going forward.

Nice. I'm going to cede the floor on 9 to Alex because it's right in his wheelhouse. But I do want to congratulate Jeff B5781 on asking such an interesting, compelling, foundational question that everybody on this podcast is dying to answer.

All right, Alex. Alex, read it out.

Alex

All right. If agents become 1 million times smarter than us and so on, isn't there a diminishing return at some point?

I think yes. Seth Lloyd at MIT was studying the question in the early 2000s of the physical limits of computation. Does the physics that we have right now impose a universal limit on the fastest—or smartest, for that matter—computer that you could possibly build in our universe?

The conclusion that he came to is that, yes, there is a physical limit to the power of computers, and that the fastest serial computer, with the physics that we have today and that we can imagine building, is a black hole. I've spoken about this on the pod previously: a sort of desktop black-hole supercomputer where you might fire in the inputs via X-ray or gamma-ray lasers—

Yeah, no. When Apple gets around to actually launching maybe a new Mac Pro, it should be a black hole, maybe.

Alex

And the output readout could be via Hawking radiation. So we know in principle how to build a black-hole-based serial computer—the ultimate serial computer.

He found that, under certain constraints, the fastest parallel computer might look like a box of plasma, a so-called plasma-based computer. So we do know, in some sense, how to build the smartest possible computer at the infrastructure level that our universe will allow us to build, unless there's a lot of surprising new physics.

That provides, in some sense, an ultimate constraint on the level of intelligence for agents that can be built on top of it. I also strongly suspect that, at the algorithmic level, we're going to find that there is a perfect agent algorithm.

Folks who've studied AIXI, a theoretical approach that's mostly popular among the AI theorist community, know that it hasn't turned out to be very useful in practice. In some sense, it represents an information-theoretically optimal AI, including an AI agent, and has all sorts of nice properties like Bayesian superintelligence.

It's not very practical, but we do know, at least algorithmically, what the point of diminishing returns is at the agent-algorithm level as well. So yes, there may be a lot of room at the ceiling, as it were, at the top, but the universe does seem to impose limits.

All right, gentlemen. Brian, congratulations on your financing. Thank you for joining us. Thank you for your sponsorship of this pod. Dave, Alex, Salem, let's go with our outro music by Marius.

We're almost there. Gravity, escape velocity. Nothing left to spare. Salim calls it your company's shape, with the purpose at the core. Massive transformative purpose. The market is watching. Compute where the smartest lie. It's another move. The smart money plays. Intelligence is a force. Don't compromise. Every future. Open every option. Moon to write the sky. Moon faster.

Amazing. Salem, you always come across as the sexiest guy in the videos.

Salem

It was an amazing week, guys. We had two recordings at MIT, and today, always a pleasure. Love you guys.

Have an awesome weekend.

Alex

Thank you.

Dave

Be well and safe travel. Wherever you're going next in the world, where's Waldo?

Awesome. 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. We spend the entire week looking at the Metatrends that are impacting your family, your company, your industry, your nation. I put this into a two-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.