企业惧怕前沿模型｜Sam Altman向Trump提供OpenAI 5%股份｜DeepSeek自研芯片

It’s like rewriting Atlas Shrugged, where John G goes to Washington and says, “Why don’t you regulate me more? Why don’t you take more? Why don’t you take us, Mr. Mouch? Grab some of my stuff.” What the fuck are these people thinking, volunteering for this stuff? It’s madness.

So, what are we discussing? Number 1, Washington lifts the 19-day Fable 5 ban. What does this mean moving forward? Number 2, OpenAI floats giving the US government a 5% stake. Number 3, DeepSeek is developing its own chip. My word. And then, number 4, Meta Compute launches a cloud business, which caused the stock price to jump 10%. Thank God, we’ve needed it, Zach.

1. Washington Lifts the Claude Fable Five Ban

We are back, boys. I’m no longer at the beach in the UK. There was one comment—did you see it on YouTube? It said, “Harry is so burnt,” and then, “He looks like a panda.” I was like, gosh, this whole holiday-vibes thing isn’t working for me.

Guest

You did look like a panda.

You look dumb, but that’s okay. Sorry.

Guest

I mean, it’s a look.

It took us about 45 seconds, Jason, but don’t worry, we’re good. Listen, we’re going to start with the news of the day, or one of the most pressing topics, which is Washington lifting the 19-day Fable 5 ban. How do we analyze Washington lifting the ban, and what does it mean moving forward for both OpenAI and Anthropic in terms of the permissions they have to get?

It’s a quagmire in the sense that you’ve now been entrapped in some kind of preapproval process. They’re talking about some kind of structured preapproval process, but that hasn’t been finalized yet. The zoom-out comment is that 6 months ago, you could ship software like a free man, and now you have to get permission from Washington before you do it. It’s a big change, right? How does it pan out? Who the hell knows?

There are some arguments, in terms of cybersecurity, for some process, but it’s definitely a big step. All other things being equal, you’d prefer not to have to get permission from any administration before you can pursue your business. I think part of the reason the US is such a dynamic economy is because we don’t have a ton of that. Europe does, and now we do. We sneered at GDPR, and here we are.

Guest

Whether more regulation is better is beyond my scope. I think people have been talking about safety in AI for a long time, which is not the same issue but a related issue. I think this is just the grown-up state of LLMs and AI. It’s going to have this level of oversight, whether we like it or not.

At the end of the day, this particular issue seems minor only because Claude 4 is going to have variable, per-token pricing in a week or two anyway. Most of us aren’t even going to use it because it’s too expensive. It’s going to be a niche model, at least until it percolates into the standard Opus and Sonnet over the coming months. So, the world impact will be minor.

The world’s changed to me. Sam Altman offering 5% of his company to the US government was much more interesting, in some ways, than whether some suboptimal but inevitable oversight is coming to the LLMs.

2. OpenAI Offers the US Government a 5% Stake

I totally agree, and that was going to be my next point. If you take that one step further in terms of government intrusion or government opinion, Sam is saying, “Hey, take 5%.” How do we think about that?

Guest

Well, if you own 100% of it, now you only own 95%, so you’re kind of a little bit pissed.

Or if you own none of it, you’re not very pissed because you’re not getting diluted at all.

Guest

Yeah, exactly. I will happily give away some of it. Again, stepping back, what problem is he trying to solve? I think it’s absurd, to be clear, but let’s try to go from first principles. What problem is he trying to solve?

By definition, it’s not any of the security issues we just talked about, which at least are vaguely credible, right? It’s not cyber. It’s some kind of macro argument: AI is going to destroy everyone’s job, so we’ve got to give back.

OpenAI produced a 9-point plan about a month ago that said we’ve got to rethink everything because of the economics of AI. They’re talking about maybe restructuring the taxation system of America to tax more on capital gains and less on income, because so many people are going to be put out of work because of AI that we want to lower the tax burden on labor and increase it on capital.

This is all part of that, and the whole thing is so delusional and so far from where we are now that I just stopped listening. You’ve got a really great, growing company with no discernible impact on employment yet. You’ve got a bunch of issues you have to sort out because you’ve been lapped by your direct competitor, and your focus is on telling Congress that they should—

Remember, for context, if they give 5% of OpenAI, it’s $50 billion. Your focus is telling Congress, which raises plus or minus $5 trillion a year, that your 1% of it for 1 year means they should restructure their entire taxation system. Sure, we’ll get right on that.

The House Ways and Means Committee will call every lobbyist and kick off a process that you won’t be able to control. I predict it is exactly—this goes back to the comment on preapproval. You start with preapproval, suddenly you end up with an ownership interest, then you end up with a board member. What the fuck are these people doing?

It’s like rewriting Atlas Shrugged, where John Galt goes to Washington and says, “Why don’t you regulate me more? Why don’t you take more? Why don’t you take us, Mr. Mouch? Grab some of my stuff.” What the fuck are these people thinking, volunteering for this stuff?

It’s funny. I completely agreed with you at first—100%. I was like, this is the weirdest kiss in the ring in a weird, corrupt administration, where our president made $2 billion off crypto and friends last year in profits. And that’s cool now. It’s cool for a president to actively trade any stock, his own meme coin, and make $2 billion in 1 year. I mean, the guy’s 80. What does he need it for?

At first, I was with Rory, but then I stepped back for a minute. Sam Altman, beyond being the CEO of OpenAI, is one of the most successful investors of our generation—of all time, really. He knows everything about how startups and scale-ups are run, and he’s seen it all. I think, in some ways, he runs OpenAI with that playbook in a way the others don’t.

It’s like a super startup: how he funds it, how he thinks about it, the relationships, the scaling. When I step back, after having the exact same view as Rory—this is crazy, kiss-in-the-ring crazy stuff—it’s not Intel dying. It’s like you give 5% of your company to Shopify, like Klaviyo did, so that they don’t destroy you. We see this all the time in our portfolio. You don’t want to give 40% of your company to your partner, right? And 5% is not immaterial.

What I’ve learned from my portfolio, and I think you guys will agree, is that it creates an unexpectedly large amount of alignment. You sell 5% of your company to a $100 billion partner—it doesn’t matter. To Rory’s point, it just doesn’t matter if they own 5% of your startup. It doesn’t matter how big you are; it’s immaterial. But I’m constantly shocked by how much that brings you into the boardroom.

So, my only point is—and I could be wrong—that giving 5% of your company to placate the federal government so that you’re the good guy now means you get back to being like Stargate 2.0, when Sam was up there with Larry and everybody. He was the good guy for a little while. I think, as an investor, I take the dilution.

No, no, no, no. I mean, I understand what you’re saying, and so I’m going to paraphrase: an ownership stake with a much larger entity aligns the large entity with the smaller entity. And that’s what this is. It’s a good thing.

Guest

Well, it’s more than I would have expected it to be, because it shouldn’t be. It’s immaterial to the federal government, and it’s immaterial to Nvidia taking stakes in most companies. It’s immaterial to the economics.

There are 2 arguments I’ll make against that. The first is a business one, and then the second is a government one.

Business first: Microsoft owns 30% of OpenAI. If an ownership stake resulted in besties, they’d be besties. They’re not besties. They’re in a stale marriage, looking for a divorce, but they can’t quite pay the tax. It hasn’t worked. That’s with 30% alignment with a profit-maximizing entity like Microsoft, which is rational.

Now apply that to the US government. Do you really think that because they own some of you, they’ll align with you? That’s just not the way politics works. Go back and look at TARP. Now, admittedly, that was when the banks had screwed up, so they came in and owned a little bit.

Rory O’Driscoll

They don't have voting control, but they tell you who you can pay and who you can't pay. The weak banks deserve that. But JPMorgan was like, “Why am I getting that?” That's the TARP.

If you think, in this case, because not only have you said, “Give me 5%,” but you've also produced a document that says the things that we're doing are so catastrophically impactful on the economy of this country, Mr. Congressman, that you govern, that you need to redo your entire taxation system, it's about an hour before Bernie Sanders says, “You know, you're right. This is really impactful. Maybe we should go for 50.”

Because if it really—if you really are impacting a $30 trillion economy, if that's your, in my view, absurd statement, but you've made it and Dario's made it, so you're entitled to own it, right? If you really are destroying labor in a $30 trillion economy, do you think the political monster is going to say, “I'll settle for 5%, that's grand. Call it $50 billion. You've destroyed $15 trillion of labor value. We'll settle for $50 billion”?

Bernie's already said he wants 50%, right? You deserve whatever happens to you. You deserve being regulated by the government. You deserve having to be inclusive. The idea is so palpable. Maybe you get something small and tactical, but it's such a mistake.

Why do they do it, then? These are not dumb people. They're really smart people.

Rory O’Driscoll

Yeah. And because they believe, rightly or wrongly, that the impact of this technology is so important that all these things need to be on the table. To be fair, that belief is what gave them the self-motivation and the confidence to raise billions of dollars, and that narrative is what it took.

Because if you walked in and said, “Hey, I need $10 billion to build some stuff, and it's going to have a minor impact on some parts of compute,” I don't think you got your $10 billion. You needed to tell a story like every great CEO. This is what the world's greatest fundraising CEOs do, right? They tell the biggest story. They told the biggest story, and that's what allowed them to get the now $160 billion.

But once you've told that story, and genuinely, once you believe it—and in the case of Anthropic in particular, once all your employees believe it—if you believe this thing is dangerous from a cyber perspective, from a jobs perspective, you just suddenly end up down this road. All these things become next-level logical if, in fact, the basic premise is correct.

If, on the other hand, you believe what little old me does—and I'm not them, I didn't invent this shit—that it's a really important technology, but it's not going to put 50% of the US labor market unemployed, then you believe these kinds of preemptive changes and conversations are a wild overreaction and wildly early. We'll find out which it is.

I mean, look, in 5 years' time, if Dario is correct and 50% of white-collar jobs have been replaced by AI—which I don't believe for a second—then you're damn right there's going to be political controversy. If you think 5% is going to feed that beast, you're delusional.

If half those nice middle-class people in middle-class jobs all across this country lose their jobs to AI, it's going to take a lot more than—I think it works out to what, $140 per head, which is what 5% of OpenAI would be worth—to keep the wolf from the door. So, if you believe these things are going to happen, that's why they do it. I just don't. So I'm like, whatever. This is a mistake.

3. Why This Could Invite Much More Than 5% Government Control

Is this purely a marketing exercise?

Rory O’Driscoll

Who are you marketing to? Congress.

Senators who you're willing to align yourself with. You're willing to play ball. You're open. You're not this wolf stealing jobs.

Rory O’Driscoll

You're not the wolf that you said you were. Again, it's like, “Hi, I'm a wolf, but I'm a good wolf.”

Yeah. Right. Is it trying to clean up the mess you created, to some extent? Yes. You spent 3 years saying everyone's going to be unemployed because of this thing and it's wildly dangerous, and now you're trying to walk that back while at the same time sucking up.

And if I'm going to ask you to regulate—or, as you said, Rory, last week, maybe tax Chinese or open-source models—maybe it'd be helpful if we had alignment beforehand. If I'm about to have a big ask, I think Sam is a very thoughtful communicator, and he puts stuff out there early to socialize it. They seem like little comments and exposition, but I think they're all very carefully thought through.

I think the issue is less about whether it's a good idea to give 5%. I say do it, like the Klaviyo–Shopify thing, if you think it's going to work. It's more that it's not 50 or 20. Sam is just anchoring this idea that, hey, 5% will align us with the American people, with the federal government, with the administration, without getting into politics. He's anchoring this at 5 rather than 50.

Because not only does he need the alignment, he's sensing the political winds. I think these things seem to come out of nowhere, but I think he's a very interesting communicator. He's a very good anchor in a way that isn't generally triggering, the way he does this.

Maybe other things are triggering, but he does a pretty good job of telegraphing where we might end up before it happens, right? I think it's just anchoring. Maybe it really doesn't matter what we think because it's already happened.

The decision has essentially already been made that the federal government will be acquiring a stake in OpenAI, and Sam is just anchoring it as the smallest possible stake to get ahead of this discussion.

Rory O’Driscoll

That hasn't happened yet, to be clear. That decision hasn't been made. Now, you are right: the US government, for the first time in a long time, and definitely absent a bailout, has already taken stakes in a bunch of tech companies like Intel. So who knows? Maybe it'll happen again.

The decision hasn't been, quote-unquote, taken. I don't think this administration has a decision-making process, but you're right: it's definitely, quote-unquote, on the table. Just being clear.

Does this change anything for Dario?

Rory O’Driscoll

To be clear, the proposal from OpenAI, just to be grounded in facts, wasn't, “We give OpenAI 5%.” It was, “Companies should.” So the implied statement is everyone should, including Anthropic. It's like, “Hey, everyone should give away 5% for the US government.”

To some extent, he's volunteering other people's capital, right? It all gets to the same thing, and we started with Fable. The US government is going to get enmeshed in AI in a whole load of different, probably contradictory ways, right?

It can range from cyber danger to wider regulatory danger, to economics, to Chinese open-source threats. You're just going to be enmeshed in politics.

It's funny because when you watched the internet take off, the whole emphasis was, “Cut us free.” If you look at 2 of the biggest deals at the start of the internet—maybe 3 big ones—I'm going to give you 3 really interesting 1990s regulatory issues that were amazing, amazing for the internet and the exact opposite of today.

One, you had the Telecommunications Act of 1996 that said, “AT&T, you've got to be broken up and everyone's got to give independent access,” which allowed broadband to take off. You had Section 230, I think, the one that said websites are not liable for third-party comments on their website, which is effectively what Facebook, Google, and everyone has relied on. So it was a huge amount of free speech.

The third one was, for a long time, no sales tax, which you could argue the justice of. But all those things basically meant Silicon Valley managed to have a 20-year run with the internet where the message to Washington was, “Leave us alone,” and we did great.

It's just super interesting that we're going with the exact opposite approach now, which is, “Hey, don't miss us; regulate us.” Putting our hand up and saying, “Pick us. We'd like to be regulated, too.”

Oil and gas must be looking at this going, “Wow, these people are crazy.” No one down in Exxon is saying, “You know, oil and gas is really important. Why don't we give Washington 5% and check in advance before we do drilling?”

You know, there's a different sense, just because this is 20VC—not that I disagree with any of that—but maybe this is too micro of a point. I think in the age of AI, massive dilution has been sort of institutionalized. Founders don't care anymore—not all founders, not all founders—but even 2 years ago, before all of this, before these massive rounds, most folks were fairly dilution-sensitive. VCs always have been to an extent.

Now, I find founders—you'll look at really hot startups in the news, and if you peel the layers back and look at the stub rounds, the up rounds, and the half rounds, they've done 16, 17, 20 venture rounds, often, right? Even if each one is 5% dilution, 20 rounds at 5% dilution—Rory, help me with the math—is a lot of dilution.

Look at Anthropic. You've got Dario Amodei at 1-point-something percent equity, right? Sam's at nominally zero. Anthropic is the most successful startup in tech of our lifetimes, but the founder owns 1-point-something percent. It's just an example, but I see it across tons—not all, right?

So 5% is like nothing, man. That's like the stub round I did last week. I did a round at 10, and then 14, and then 18, and then 22, and 29, and TechCrunch runs it up each time.

Guest

But those 5% and 6% stakes add up. They really add up, right? Every Ramp press release is so exciting, but I hope they're not too dilutive because there are just so many of them.

Rory O’Driscoll

It's an interesting point, and fundamentally, I don't think either of those CEOs is primarily money-motivated. But it's a great point, Jason. You are right. It is very different. Normally, the 2 winners in a startup—if you look at Microsoft, Bill Gates and Paul Allen owned a good slug each, and even Steve Ballmer owned enough to buy a basketball team at the end of it as one of the richest men in the world.

You're right, it is really odd where 1 of the 2 CEOs owns 1.7% and the other owns 0%. You're right. It totally takes the edge off the dilution conversation because it's someone else's money.

It's not that, as a seed investor, I sort of hate it because I've watched myself be diluted to levels I never even thought would ever happen. When I started investing, as an investor you have to internalize and realize it's basic for me: it's doubling again my entry price, right? As a seed investor, I used to think my real entry price was twice what it looked like because of dilution. Now I'm thinking it's 4 times.

So, Harry, you just talked about doing a seed deal at $60 million, right? I think you're really doing it at $240 million. Harry, is that the honest math today, right? And founders—not all founders. Look, there is absolutely a vibe of, “I'm going to go through YC, I'm going to raise $6 million at $60 million and never raise again.” Amen, right?

But so many founders today, after that first round, are not dilution-sensitive. And maybe it makes sense if it's a huge outcome. It's just different. It's just different.

The only fact-based comment I'll make is: you're right, though. The data from Carta, which is always excellent, says that dilution per round is going down. So maybe, in part, founders are willing to raise more because the dilution per dollar is lower.

Maybe some—not the founders of Anthropic or OpenAI, but in general—if the pricing goes up, you can do more rounds and end up with the same dilution. I think that's happening in a lot of cases as well.

Guest

Yeah, but I think what I'm personally seeing—and I think Dario is an example—I'm seeing both. I'm seeing smaller rounds, but so many rounds.

Rory O’Driscoll

Yeah, so many rounds that each round you kind of don't mind as an investor or board member. Great, do it. Wow, that's a great deal at 6% dilution—5%. That's not double digits. Then 4 months later, you do another one, and 4 months later.

It sounds like I'm complaining; I'm just learning. But if I'm doing that 5% to hold off any regulatory issues, man, just do it.

Well, I like the learning comment because you're right. I will admit that one of the areas where I think I may have been too rigid is—you do think about—you don't have hard ownership targets, but you want to have between 5% and 10% of an investment to matter.

You're seeing now that Spark is going to do amazing, and very deservedly. They're going to own 1% plus or minus. SpaceX, I think, found it—who did the original check when the rockets were still blowing up, for God's sake, right? Ballsiest check out there. They're sub-5%, or 3% or 4%, of SpaceX, right?

4. DeepSeek Building Its Own Chip & Anthropic Talking to Samsung

You're right, Jason. I think for these huge outcomes, the mental math and the mental model you had gets really turned on its head, which makes sense. If you have an exit 3 orders of magnitude larger, you can get away with just about anything on the dilution side. The question is—

Guest

I mean, Ramp's done 12 announced rounds, according to Claude. So I'm going to guess it's more like, with stub rounds, 24 rounds typically, right? I think it's done 24 rounds of funding, right?

Rory O’Driscoll

And Databricks has done rounds down in the second half of the alphabet. It was like a Series M there. I love the way they actually named it, not just “another Series M.”

Guest

These are honest founders. These are actually old-school founders living in the AI age at Databricks.

I have none.

Guest

It's not performative at Databricks. You know who I love, though—and then we'll get back to normal scheduling? Karri Saarinen at Linear. The dude is so disciplined and so focused. He's raised 2 rounds of funding. He never wants to meet VCs. He really refuses all VC intros. Never meet.

But are you sure that's the right outcome?

Guest

Yes. It will return my Fund I multiple times over, and I'm incredibly grateful to him and the team for doing so.

No, listen, I'm not being critical. It's a beloved product, right, with real traction, and it'll return your Fund I. That's great. Sometimes everyone's, to use Rory's term, talking their game a little bit, right?

When I sometimes see him say that, I agree with him as a founder, right? I love it. But sometimes I hear a little bit of Brian Armstrong in that. It's because, listen, maybe I could have done even better. As great as Linear is, maybe I could have done even better, but I chose to be capital-efficient.

And sometimes I feel like, was that the right choice in 2026 when the prize is so large? If the exit's a couple of billion, $5 billion, it's good. If the exit's $100 billion, right, then you just—I'm not literally—I just sometimes wonder when I see it. And I'm not saying that I'm right. I'm not remotely saying I'm right.

Rory O’Driscoll

I think what you're weighing off, to try and step back, is optionality versus upside, right? Look, there's no doubt that if the prize is $1 trillion—which it has been in at least 3 cases—it looks like it really doesn't matter what it takes to get there. You just have to get there. And if skimping on it reduces the probability of getting there even 10%, it's a huge mistake.

If, on the other hand, the prize is, as you say, $1 billion or $5 billion, then raising too much eliminates the optionality of taking that $1 billion exit, right? To be fair to a founder, 20% of a $1 billion exit is life-changing—life-changing, especially with QSBS.

For now, I think the truth is it's different by opportunity. Not every opportunity is an Anthropic opportunity. Linear is a hard one to place in that because you squint one way and you go, very bounded, very well-executed, it will be a great outcome no matter what you do. To your point, Jason, you can see another world where you have to become something bigger to even matter.

Guest

I think there's a related point just for founders today that's changed. This insensitivity to dilution is different. It's just different, right? Because if the outcome's massive, it doesn't matter.

The other thing that has changed, to Rory's point—and I think this is a positive, because it certainly terrified me as a founder, but it's not all positive—is no one's worried about making their last-round, high-priced investors money anymore. Literally no one is, because I believe investors have learned to accept 1× when it doesn't work out, without drama, without blocking, without threats.

I'm not saying weird private equity firms and non-standard investors—they play games all the time. I see it. I'm watching a threat through my portfolio from a non-standard VC right now that is blocking round after round after round.

But the Sequoias of the world and everyone, you're not blocking exits, right? And so I think founders go, “Oh, I raised at $4 billion, but maybe I exit at $800 million and I get an $80 million carve-out.” They're just not worried.

I was terrified as a founder that for every round I would get blocked by the douchebags. I just don't see any of that fear existing in founders anymore.

Yeah, without drama, right? It's just changed. I think this is the age of growth investing, and the fact that there's no downside because your investors won't block that billion-dollar round adds velocity—not on the investor side, but on the founder side.

I would have taken another round as a founder for sure if I thought I wasn't going to get blocked. I would have done it in a heartbeat. I didn't get it at first, Jason, but now I'm getting it.

And you're saying—and that's the argument that says if the high-priced later round has relatively low blocking rights, relatively low dilution, and it gives you upside optionality and doesn't preclude downside optionality, which is my point, then you're saying I would be wrong.

In fact, there are cases where, if you as a founder are running a company doing $100 million and there's some chance you can be a billion-dollar-revenue company, you should take the round. Take the latest-stage round. It might work if there's a 50% chance it works.

And if it doesn't, you'll have the preference stack. Don't waste the money, and you'll still be able to get out with the exit you would have had otherwise. I don't know if I buy it, but that's the argument.

Guest

I think there's a big “but.” I can't think of a founder I've invested in doing the big round who is worried about the return on that high-priced round. It just—my generation of founders, we were terrified of it.

Guest 3

We were terrified of the expectations.

I remember those terms where you had the block unless it’s a 2x sale and all that. So you were really stuck with that latest-stage money. But I agree, that’s actually a fair point. It’s freed up the risk.

One video that was going incredibly viral was Alex Karp on CNBC, where he really said 2 things that I think were standout comments. One is that there’s never been more skepticism from large enterprises toward frontier-model providers, specifically Anthropic and OpenAI. And then, second, that there’s a real question about the ROI of AI within their organizations. Anything to add? Any commentary on that?

Guest 3

Yeah, I actually watched it because all the whiny people were saying he looked deranged, and I watched it and really enjoyed it.

Right? I actually thought he wasn’t deranged.

Guest 3

I thought he looked more stable than he normally does. Yeah, it’s so funny because some of the examples, it’s clear there’s a whole lot of personal dynamics there, and his examples about his college and his examples—all that—that’s just his baggage to bring to the table. I read a biography recently: super-interesting dude, obviously with a lot of angst. So I think there’s a lot of noise in the system from that.

And then it was cute that he called Dario a world-historical figure, which is the Hegel concept—you know, the German philosopher—and Alex Karp is, of course, a doctor of German philosophy. So now we’re dealing with a big brain making big-brain references on CNBC, which perhaps isn’t the right place for it. But when you strip away all that, I think you’re right, Harry: the 2 comments he made were spot-on.

Corporate America is saying, “I’m spending all this money. Am I getting anything?” That’s the ROI comment. And then the other comment, which I hadn’t heard as much, and it’s obviously a little trite, but he said, “Corporate America is saying, ‘Am I giving them all this information? Are they training them? Are they learning my business and then going to be selling my business to everyone else? What’s my IP?’” Right?

Obviously, it was a self-serving comment because then they were like, “Well, Palantir will solve these problems for you, Mr. Corporate America.” It’s worth pointing out: people can bash them, but the stock went up 9% on the day, right? I didn’t realize that. I checked it just before I came in, so I didn’t think it was crazy at all. I think it was—yeah, I mean, stylistically, you kind of go, “Wow, that’s a crazy style,” but, oh my God, the points were spot-on.

I don’t know, Jason. What did you see in it?

Guest 3

I only saw the clips. As Harry knows, all we watch are clips now, right? We create long-form content to create clips, and that’s life.

I think, listen, anybody on the application side is going to be sensitive to token-model costs and all of that, right? It’s a thing that’s real and has blown up. And, yeah, he’s talking his game and his dependency.

The one that maybe he got slightly wrong but is the most interesting because it’s still a real issue, right, is whether OpenAI and Anthropic are really training on and slurping up all of our data, right? And that seems to be slightly exaggerated based on their terms of use and everything today. But this was the same week that HubSpot had to walk back that it was going to share all your prospecting data with other customers.

I want to tie them together. HubSpot’s an old-school B2B company, but HubSpot said a week ago, “Hey, we have a prospecting tool.” Prospecting is really important. It’s actually become much more important in the agentic world because all these hot AI GTM products are automating prospecting, right?

So we’re going to do what everyone’s tried to do for about a decade and a half: we’re going to pull all your data. We’re going to take all of Harry’s verified contacts, all of Rory’s and Jason’s. We’ll pull them so that when you do outbound, you’ll have a truly validated set of contacts. Their customers erupted: “You’re sharing my contacts with other customers.” They had to roll it back within a week.

It’d be fun to talk about in general, but I think it tees up the question of whether vendors overall are going to push the limits here. OpenAI and Anthropic kind of lied about the books, and they definitely lied about training on YouTube, and they’re going to push the envelope here to make their LLMs better. HubSpot did it, Salesforce is going to be tempted to do it, and every vendor that’s seeing massive competition or slowing growth is going to be tempted more and more to cut corners on training privacy. HubSpot got caught. At least they walked it back, right?

I think we should all be worried if we care about our data for real, and sometimes we over-worry about this, right? We’re not all anarchists or whatever. But people are going to be tempted to do more and more with our data. I think it’s a very valid worry. And if you’re Palantir, selling to the government and highly regulated industries, I think it’s a great play. It’s a great play. You can’t really trust these guys not to share your data. You can’t.

You’re right, Jason. The ROI comment was clean, and that comment on data wasn’t as obvious. But of course, the other thing that Karp mentioned correctly was that Anthropic in particular had, quote-unquote, opinions about how its AI should be used by the DoD. And he was making the point that when people are giving you millions of dollars, they don’t want your freaking opinions; they want your technology. I think he did a very good job of positioning himself on that side of the table.

I always heard the statement when I was younger: those that can, do; those that can’t, teach. I always liked to remind my teachers of this, which is probably why I was so unpopular at school.

Guest 3

Yeah, you would be.

And then you kind of look at the ecosystem we’re in today and you say, “Those that can, do; those that can’t, open a cloud business to sell excess compute.” We saw this week that Meta launched a cloud business to sell access to AI compute and created a neocloud—a cloud business to sell access to its AI infrastructure, either as hosted or raw GPU rented by the hour, like CoreWeave or Nebius. The market reacted well: a 10% jump, its biggest single-day gain in 5 months, on this announcement. How should we think about this one?

Guest 3

My only thought was, why not earlier? Why not? If you’ve got the capacity, why not lease it? It didn’t bother SpaceX. It didn’t bother Amazon 20-some-odd years ago—Harry can do the history for us—when Amazon opened up AWS back in the day when it had excess e-commerce capacity. Why not, man? Why not?

It’s one thing if Meta still had massive cash flow it didn’t need, right? But I’m not sure exactly what their net cash is from their infrastructure spend. Maybe it’s 0. It just makes sense at this point. It just makes sense. Why not?

It’s been interesting. 2 companies have done the same thing, which is buy a load of compute to build proprietary assets, fail to build those assets, and then decide instead to sell that compute to others. Both of them have had a very positive market reception from that, right? One of them is SpaceX, obviously, and the other, obviously now, is Meta.

You ask yourself, what’s going on long term? What is the market actually thinking? Are they thinking there’s a Goldilocks scenario, which is: we, the market, believe that Meta in the short term has excess compute and therefore we’re glad they’re selling it, and in the long term we believe they have a wonderful use for this compute that we can’t quite figure out yet? Therefore, long term, it will be this AI-centric play and it’ll all be wonderful, right? That’s 1 view of the world.

You have the same kind of view of xAI, which is, oh, short term they had to rent this compute—they got an extraordinarily high price for it, for which all congratulations—but does the market really believe that, over the medium term, you’re going to be an AI-model provider using Grok as a top-to-bottom, state-of-the-art model? That’s 1 view of the world. Or is the market simply saying both of you have failed at your long-term goal, but being a cloud provider is a great business, and go team?

5. Meta Launches Cloud Business — Plan B or Masterstroke?

The interesting question was, what does it mean over the longer term? There are 2 positive scenarios. The first is that they build these standalone models, take that compute back in, and use it all. That’s great. The other positive scenario is that being a hyperscale cloud provider turns out to be a great long-term business. That’s great too.

Obviously, the bad scenario is if a whole load more companies go through the same journey Meta did, which is, “Oh, we think we need all this compute, but we don’t. We can’t build something useful enough for it.” Then you’re only left with a few buyers of compute.

Guest 3

OpenAI and Anthropic can clearly use it, and a whole lot of sellers of compute. Maybe it won't be such a good business 2 years from now. And that's the risk: it turns out that, right now, the assumption—and Zuckerberg said it—is, “Hey, we should invest because if we can't use it, we can always sell it.” And this is what you're seeing right now. Everything there is true up until the moment that compute demand isn't there at the margin.

That's not happening now, to be clear. It's never been tighter. But if that changes, then all these assumptions go out the window. Then the market will say, “No, I'm not glad that you bought this [__] and are now selling it to other people. I wish you hadn't borrowed at all. Take the hit.” But that's not where we are today. Compute demand still appears to be pretty strong.

Well, right now it's working. But it is odd to be able to get away with having a Plan A, reverse it, go with Plan B, and get a 10% lift. Do you think Zuck will execute on the strategy? Well, Elon did a masterful stroke with it. We've discussed it before. He got a great price for it.

Single customer, amazing job. It's not easy to do. Do you think Zuck will be able to pull it off?

Guest 3

I think you're phrasing the question wrong, with all due respect. You're basically saying, “Oh, is it Zuck or Elon?” And they're both wildly talented. Let's use the C-word: world-historical figures, right? Which I think is true, actually, right?

I think the real question is, is there another 5 gigs of demand out there that wants to be satiated? Does Anthropic have an open-to-buy? If the truth is this, if a customer wants to buy something, as every salesman knows, it doesn't take a freaking genius to sell it if you have it, right?

If Meta has a gig of compute lying around and Anthropic, 5 miles down the road, wants to buy that compute, I predict that sale will happen, right? If Anthropic doesn't want to buy that compute, then all bets are off. So that's the only thing it boils down to.

Listen, maybe I'm not that bright on this, but Zuck also just paid essentially $900 million to hire a head for WhatsApp, right? By investing $900 million into CRED, right? So it seems to me—I might be wrong, not trying to trigger anybody—

Guest 3

Just to provide context: if Meta invested $900 million into CRED, an Indian company with a CEO called Kunal Shah, I believe, and Kunal is now head of WhatsApp, with that $900 million investment in CRED, I believe it's probably more, right?

Really? Because that was a $900 million investment. They paid well over $1 billion to get someone to run something, too.

Guest 3

I agree. It was wild.

So it seems to me—what's happening there?

Guest 3

Well, clearly, and you can see the numbers, the core Meta apps are working well: WhatsApp, Facebook, Instagram. This is the engine that keeps going, right? I mean, you don't—it's not confidential, right?

In a way, Zuck's treading water while he figures it out, right? Did he overpay for scale? Alex would say, “Maybe.” I mean, probably, right? But he's treading water.

And listen, a lot of our founders are in this boat. The main engine's working, right? Something's working. I don't have all the answers today in the age of AI, right? My core business is still doing well. And I can either hide from it, or I can go maybe too all-in without having the answers, but at least I'm in the game.

As crazy as some of the questions are—did Llama really work out? Did Scale work out?—I don't know. But when the core is so successful, you stay in the game and then you rent out the compute. It's okay, right? So I give the same advice to founders that are doing reasonably well: stay in the game, man.

I think you're totally right, Jason. The core business is doing amazingly well. Now, one minor nuance: they say that part of the reason it's doing well is that AI is improving their targeting, and I believe that, but I don't believe it justifies the $70 billion or so they're spending.

But you're right, the core business is doing well, which means that there's no fundamental fatal-error risk in continuing to invest in this new marketplace, in AI, right? So if you were a Meta board member—not that Meta board members have any power whatsoever because Mark controls all the votes—but I also think, as a board member, one of the big-picture jobs you have, and you have very few jobs, is that if the company is doing something that could have fatal-error risk, that's when you at least record a no vote and say, “I wouldn't do this.”

If the CEO came in to me and said, “I'm doing this,” I'd have to say, “You've earned the right. You've got a $100 billion cash-flow business. I don't understand where you think the $70 billion of investment is going to get you, but you've earned the right to continue to play.”

So even if there was a meaningful board at Facebook with actual votes, if I was a board member, I'd be saying, “I might get it, but you've earned the right to play.” You've hedged, and worst case, we spend $70 billion, we're wrong, just like VR. So, yeah, I agree.

One of my big ahas is that people talk a lot about the fact that all the hyperscalers are spending almost all their capex and even starting to tap the debt markets to invest in compute. My big aha is this: spending isn't going to stop because the supply side says stop. Meta isn't going to say stop. Google isn't going to say stop. Microsoft isn't going to say stop.

Really, it boils down to the demand side. As long as the revenue growth rate from end-enterprise customers, even though it's 1/7th the size of your capex bill, is 2x-ing and 3x-ing—which is what we've seen from OpenAI and Anthropic—even at today's run rate, the spend is going to come. The demand side is going to be what shuts off the spigot, not the supply side.

And I think Zuckerberg is just the best example of that. He is going to keep playing as long as there are some hints on the demand side, and it's not a fatal error. Neither of those has been triggered.

On the supply of money and keeping that money machine rolling, NVIDIA starts financing its own demand with “compute now, pay later”—essentially letting providers access their GPUs through revenue-sharing and credit support instead of paying upfront.

6. Nvidia's "Compute Now, Pay Later" Scheme

Guest 3

I love it. Now that round-trip revenue is totally cool and not something you go to jail for, let's do it every single place we can find it, right? Let's just do it. I'm not saying there's anything literally wrong with it, but go for it, right? Capture them early.

I'm just shocked by how many folks have screwed this up over our investment histories. How many folks don't just go ultra all-in on startups? And if you want to pick YC because it's the simplest way to go all-in, just do it, right? It is such a talent magnet.

Folks have woken up to it to some extent, but every leader should be showering startups with infinite love in their first 24 months. It's the best long-term investment you can get. If there's any lock-in or anything at all, shower them with love.

Let's talk about what's going on here, because what NVIDIA has said—and the details matter—is that, for next-generation neoclouds, and one, I think, Nscale AI, which is one of the examples, they did 2 deals recently in early July. They actually did some kind of explanation of what they're doing.

They're basically, quote, selling you the chips up front, so they are going to recognize that hardware revenue up front. Then they're giving you, as the buyer—the neocloud—a backstop that, if you can't use that compute, you get put-back rights on it, right? So it's basically hedging the risk.

It wasn't clear to me from what I read when the money actually changes hands. But what was clear is they are taking the revenue up front. So it's, you know, as legal as church on Sunday: it's ASC 606. They're separating the revenue up front from the guarantee over time. So it's accounting-legit, but it is pretty aggressive.

What it's basically saying is, their push has been to diversify away from the hyperscalers, and they've achieved that. Even though, obviously, the bulk of their revenue comes from a small number of hyperscalers, they're starting to expand their customer count—the number of significant customers.

And the top 3 customers, I think—don't quote me on this—in the data-center business have gone from the 80s to the 50s or something like that. So they're trying to make all these new neoclouds work, right? And they're leaning over backwards. There's a lot of contingent liability they're taking on.

And Jason's right.

Guest 3

Yeah, you do that, and it goes back to the same sentence over and over again: as long as the raw demand for compute and intelligence keeps going up and to the right, these deals will look wildly smart because they'll work.

And if that slows down and there's excess capacity, these deals will look horrible because you won't just not be growing quickly; you'll be debooking prior revenue. You'll be taking money back because your customer will have gone bust.

So the whole thing is a derivative bet on keeping this thing going. Not crazy, but that's what it relies on. I don't think, at this point when we record this, anyone's managing for downside. I think we're so far deep into a bull run like we've never seen before—bubble or not, I don't know—that if you're managing for downside, I think I'll check out of that board meeting. Thank you. Here's my junior associate.

You're right. Funny, I remember thinking a year ago, when I realized NVIDIA were talking about stock buybacks, I remember saying—actually, I think I said it on the pod—I said I wouldn't do that. I wouldn't do buybacks now because buybacks are a concern.

I actually think if you're going to be stupidly aggressive with your cash, this is actually a better way because it keeps the thing going. Now, I do think the time to manage for the downside is when no one is managing for the downside. So, there's a little part of me that just goes, “Oh, we're at that stage of the cycle, right?” And we remember that stage of the cycle in 1999–2000.

I want to say again: history doesn't repeat. It does rhyme, but it doesn't repeat. These are different companies, different times. But it is interesting. We've reached the point where the number of good customers who can pay cash and have a big balance sheet is tapping out. So, you've got to find more customers to keep the growth going. And to do that, you've got to subsidize them.

Speaking of dependence on customers—customers having the money—one of the biggest customers for NVIDIA is Anthropic. Anthropic opens talks with Samsung to build its own AI chip. That was on Thursday last week. And today, DeepSeek has announced that they are starting to build their own chips. Is this the natural progression of an ever-maturing industry? Will everyone build their own chips? How do we think about this?

Guest

Last week I said I thought it was mad, and I actually saw the comments from Anj Midha, who I think is just super smart. He responded to your thread, and his comment was, you know, you've got to own the—there are 2 arguments in favor of it that I didn't internalize last week when I said I think it's crazy for OpenAI to be building its own chips.

The 2 arguments were, one, his comment, which was some version of: you've got to own the compute. If you don't own the compute, you're screwed—a little like crypto: if you don't own the keys, you don't own the crypto asset. So, he was very much viscerally, “You've got to extend the whole way down.” And I just think he's been so smart about Anthropic in 2021. He's been so smart about the need for compute that that made me pause and think: am I wrong?

And then the second thing, kind of more technical, is if you build your own silicon, you can optimize the silicon for your model and probably get significantly more efficient than you might buying a general-purpose computing platform from NVIDIA and adapting it to your specific model. So, there are 2 arguments that I didn't have in my head literally a week ago in favor of this thing.

But I will admit I still find myself going: if you're at the app layer and that's where your value is, and then you have the model, and then you have the hosting provider, and then you have the chip, just needing to do that amount of vertical integration feels weird. But I don't get it, and I may not be understanding the big picture. That's my more tempered approach than last week.

Rory O’Driscoll

The only thing that makes zero sense to me is the argument that, hey, OpenAI, we need to build our own chips because we have very specialized needs that NVIDIA can't meet. Yes, I have a little bit of experience in the semiconductor industry. If you're driving that much volume to them and you need a special version of a chip, they'll build it for you. This is not true.

Okay, Frazelle will make fun of me in the comments or whatever. I admit my experience is a little dated, but for this amount of dollars, in my limited experience in the semiconductor industry, they'll do your own tape-out. They'll build you your own. It's so much money—80%, 50%, whatever the revenue.

This is just responding to the belief that the margins are so high at NVIDIA: to survive, we have to recapture that margin. I just think the idea that it's customized for us is just soft language because everyone's kind of dancing around being aggro with Zuck here, right? Everyone on either side is maintaining relationships. But it makes no sense in my experience. It just makes no sense.

7. Kling Raises $2.8B at $18B — Why Did Sora Fail Where Kling Didn't?

Kling raises $2.8 billion at an $18 billion valuation. It is the biggest AI video business in the world. It's doing $500 million in Q1 ARR-wise. It's clearly going to go public on the Hong Kong Stock Exchange soon. Interesting in the context of OpenAI shutting down Sora.

Guest

I think there are 2 interesting things, right? One is: if Kling can pull this off, why the hell couldn't Sora pull it off, right? Why couldn't you build the more cost-effective one? I've used all these models inside of Higgsfield, right? We could talk about it.

The second thing is just more interesting, that I wondered. So, Kling—you said $18 billion. That's what they're doing it at: $500 million now. Higgsfield, where Harry and I are both investors—I was one of the first 10 users or customers—just announced they're at $500 million in revenue. They're actually doing $2 million a day now in credit-card billings outside of the enterprise.

Okay, so however we define ARR in today's world—$600 million, whatever; $500 million in revenue—Kling is just one of the models they use, but it is important to their product, and they are allegedly raising at $5 billion.

So, one question that I sort of thought is: is there a Chinese valuation bubble potentially in AI, like there have been in prior rounds? It's a different market, right? And that just creates different dynamics for capital raising, right, for startups and foundations. If AI valuations are—and I don't know this to be true—going to be meaningfully higher in China than the US, it just, by its very nature, changes how the game is played.

If valuations are higher, right, because you've got 1 at $500 million doing AI video models at $18 billion, and 1, a partial layer on top of it, that's cash-flow positive at $5 billion, is that a 3x arbitrage? I don't know.

Rory O’Driscoll

Jason, with the greatest of respect, you've got DeepSeek raising at $50 billion, a gross discount compared to any Western alternatives, and you've got ByteDance—

Guest

It's a counterargument.

Rory O’Driscoll

ByteDance at $500 billion.

Guest

I just didn't get the $18 billion. It's a question more than me banging my fist on the table, right?

But the meta-learning is that I just didn't know video would be this big, this type of generation, right? You know, it's big on the consumption side; we just sit doom-scrolling all day, right? All of us, right? But it wasn't clear to me a year and a half ago, when these outputs were pretty crappy, just like a lot of it wasn't clear to me that the demand would be so insane.

But now that people are actually beginning to build films on these platforms, the amount of video you can consume is infinite, right? So maybe Sora should have figured it out, because it was pretty good when I would run all 4 together—Kling, Sora, Veo, and I forget the other one, the other big Chinese one—because you can run them all on Higgsfield. You can just run them all in videos.

It's kind of a bummer they shut it down, right? They just couldn't make it cost-effective, and it was a slightly inferior product. Maybe that's what the market wanted, right? Kling is still pretty cool.

Rory O’Driscoll

Yes. I also think there are far fewer freebies on Kling. They're very quick to charge. I mean, from recollection—it's so funny how quickly we forget—I can't remember how much Sora gave for free. But look—

Guest

Too much, probably. Right.

Rory O’Driscoll

Exactly. Because the point is that if you're OpenAI, your highest and best use is elsewhere: you have another use for that compute that's enterprise-centric, where you can make real money. So, you're probably, at the margin, cut off on a standalone basis if you were charging.

You're right. So, that's 1 comment I can see: if I have a finite number of GPUs and I'm falling behind on coding, there's more money in coding than consumer video. Separate comment: if all I have is a consumer video business and I can validate with a charging model that allows me to make money, then that's great.

I mean, I saw an estimate as well, Jason, for a 30-second video generation: it's about $1.30 to $2 in GPU generation costs. It's kind of a rough, very rough estimate. So, provided you can get some kind of money from it, there's a business there. Different business than enterprise coding, where those GPUs from OpenAI presumably ended up. But nonetheless, you're right: Kling has proven that there is a business here and people will pay for it.

Guest

But you're right. I mean, $500 million to OpenAI today is nothing. Not only did you lose your capacity, to Rory's point, which was the biggest issue, right? It's just not—and not only was it under-monetized, even if they were able to monetize it at the Kling or better level, it's a distraction. It's below the materiality line and a lot of capacity is used, but for Kling it's not a distraction.

Rory O’Driscoll

Totally. It's a wonderful business.

Yeah. I mean, it'll be interesting. This is why we get to invest in startups, because the distractions can become very large businesses. That might be a lot of great investments, right? That's just a distraction for us.

The most commercially successful AI video product on Earth is Chinese. The top 6 models as of today on OpenRouter are Chinese. Do you think China's running away with the model layer?

Guest

I don't think so. I don't know—the quick answer is you have to be more fine-grained. Well, first of all, the top social network for sharing short video was obviously Chinese. It was TikTok, and it got adopted here. They just competed in the same rough market as Instagram Reels and all the others.

So, that's kind of on the pre-GenAI video business. On the GenAI video business—

Guest 3

You're right. Kling is the top model. Sora decided they've got better things to do, right? We just had that discussion on the big market, which is obviously LLMs for compute and LLMs for coding. Look, the US is clearly running away from it in terms of frontier models, right?

And the Chinese counter-strategy has been open-source models, distilled in some part. Reasonable people might differ on how much of OpenAI and Anthropic, but, yeah, they are clearly numbers 1 to 6 in terms of the non-closed-source frontier models. So, in that market, they're running away with it.

It's just one thought I didn't fully appreciate. I just got back from 2 weeks in China and Hong Kong, which I didn't appreciate until I was on the other side of the Great Firewall, and now I think Jensen was right about this: when you're in China, OpenAI and Claude—and Anthropic—will not serve you. It's not just a question of being blocked. You cannot access it.

Now, you can get around it, right? There are ways, but they try to block VPN access. So, you kind of have to buy tokens or buy things. What do you expect China is going to do? The second-largest economy in the world—of course they're going to build things that are as competitive or better than we are, because you can't even use Claude in China.

If, for some reason, we don't like what's happening in China, we created it by not allowing China. Jensen's point was, you better let the GPUs go over there, right? Or they're going to just do it themselves. And literally, the fact that even in Hong Kong, which is much more open than China, I just couldn't use ChatGPT or Claude or the APIs—what do you expect?

And to Rory's point, probably video—they're going to go with it because they have so much strength there already, so much domain expertise. But, of course, they're going to build it all. You can't even use ours. [laughter] And, of course, they're going to be pretty good.

There's some pretty damn good engineers in China. They've been working on the internet and software and AI for a while. This is not—if we don't like what's happening in China, having just gotten back from 2 weeks there, what do you expect when you can't access the leaders? When you simply can't access them, they're going to build something as good or better if they can, and they can. They can come close. At least we know they can come close. [snorts]

Rory O’Driscoll

8. Open Source Plateau: Are Cheaper Models Actually Delivering Results?

I think, Jason, you're right, and it's well expressed. The only nuance I'd say—well, I don't know if Jason's right—is that Jensen was right that this is the consequence of us not allowing access to frontier, state-of-the-art chips and then state-of-the-art frontier models.

Now, you can decide as a country, going back to the government thing where we started, that that's an acceptable price to pay because you believe the national security issues are significant enough that you want to do that. I'm deliberately saying that I'm not saying they are not, and I'm not frankly equipped to assess that. But your actions have consequences, right?

It goes back to when we talked about that famous Jensen podcast with Dario, where they were kind of talking past each other. If you believe there's a national security concern on these models and these chips, and it's legitimate and real, and you've made that decision soberly as a government and responsibly, then you can choose to block access to these technologies.

But you're right, you can't expect the other side to say, "Okay, you caught us. We give up. We won't have this stuff. We'll build our own." And there will be a commercial consequence to that. That's what you're seeing here. You're right, Jason. They didn't say, "Okay, we can't have cool LLMs from Silicon Valley. We'll just give up." They said, "No, we'll build them all. Thank you very much." And they've done a pretty good job.

And now it's interesting, Harry, you just—I haven't even seen this. I hate talking about things I haven't seen, but Harry mentioned just as we came on the set here that there's information out from China that they're starting to say the Chinese government is saying, "Maybe we'll deny access to overseas users to some of the Chinese open-source models."

That's kind of hilarious in one respect because we're nervous about using them because we think using them is dangerous, and they're worried about letting us use them because they think letting us use them is dangerous, which is kind of just a zany thing, because both of those things arguably can't be true at the same time. But that's where we are.

It would be very significant in terms of competition—the competitive environment—if Chinese open-source models were removed as an alternative going forward. I think that would be obviously pretty excellent if you are a US frontier model. See prior conversation, Jason: you might be right. They might be getting something for their 5%. Or, B, you're a US open-source model provider like Reflection AI or Poolside; this would be the best thing that could happen. We'll see. I don't know. I haven't seen the press release—oh, not the press release, I haven't seen the news story.

I just see more and more moving towards open. Did you see today that the co-founder of DoorDash announced that they were moving? I'm getting it up now.

Guest 2

Towards open, yeah. I mean, everyone's trying to do that because of the expense.

Guest 3

If there wasn't a more cost-effective way, we wouldn't be building chips either. It's the same thing, right? The margins—it's just that we're reaching—we've now graduated from the experimentation phase, right? And now we have to deal with managing costs. That's what CIOs and companies do, weirdly but reasonably well, right? And it's just going to accelerate.

It's funny. I'm trying to build this project right now and launch it, and it's got a sufficiently complex algorithm that I can't understand it. I'm not smart enough, right? Folks can fry me in the comments. I just don't—it's an application I cannot fully understand how it works.

I'm using a mix of the models in Replit, which is Sonnet plus open source, basically what I'm using. You can use Sonnet and Opus, but I'm basically using it. I can't quite get it right, so I'm passing it to Fable and Opus, and then I'm running both side by side.

I ended up running Fable and Opus with Replit, and my point is, after spending about 10 hours in Replit, I couldn't solve this big algorithm problem. I solved it in about 20 minutes in Sonnet and Opus, right? So, there's going to be, even for me, this grade of problems where I lost so much time and money using the one-step-down model.

I lost a day in endless cycles. Forget about the money—$500, whatever. I lost a day. I've got stuff to do. I've got portfolio companies to rescue with my grand insights, right? I've got stuff to do. And by using Sonnet plus Opus—and I'm not sure which combination really did it—I was able to get to the heart of the problem in an algorithm I could not understand.

So, I'm just saying, I don't know how this all plays out over the coming years, but as the problems we solve get bigger and more complicated, I'm not sure I want to waste a day on a mediocre answer that doesn't work. [snorts]

Guest 2

I think you're right, Jason, and actually Jesse Zhang, the Decagon founder, did a nice post on that just now. It was good. Basically, he said, "Look, when you're trying new stuff or you don't know the problem or you can't bound the problem, you're going to use frontier models because they're smart and they'll figure out the unknown unknowns, right?"

The more it becomes a commoditized answer where you know the answer you want to give, the more you're going to push it to open source. It was a good paper. It totally made sense. And his comment was at the explosion of usage now, so you're seeing a lot of frontier-model usage.

It may well be in 2 years' time that you didn't need to pay that tax, but right now, if the only way to solve the problem is with the frontier model and the problem's worth solving, you're going to pay for it. Which is why the OpenRouter data—all the tokens with open source—is a little misleading, because all the tokens can be in one place, but all the dollars can be in the other place.

To your point, Jason, at the end of the day, you're glad you spent that $1,000 to get the answer in Fable.

Guest 3

It was actually cheaper because I needed 10 minutes.

Guest 2

Yeah, exactly. It wasn't just more expensive; it was cheaper in soft and hard costs: 20 minutes instead of 8 hours and $500. It was 20 minutes and actually zero because I get it in my $200 Max account, right? So, it's free. It's subsidized.

It's no different from any advice business. There's a reason. Sometimes you go to the nurse practitioner, and then sometimes you go to the heart specialist.

Guest 3

And we may end up blowing a—listen, we need help, and there are vendors that do this, right, that are on fire. But we're going to need help making sure that when we use cheaper models, it's actually worth it.

I think, even in—I’m bored of talking about the subject, but you brought up Decagon. If you really go deep on a lot of the data today and a lot of folks doing next-generation AI CX, there is some plateauing. The reason there is some plateauing is some of this pressure to contain, to have reasonable costs per resolution.

We're kind of standardizing this industry around $0.50 per resolution in CX, right? That's sort of the cost, right? It's gone down from $1 to—so how, assuming you're not just burning venture dollars...

Guest

If you can charge 50 cents for a resolution, what do your LLM costs have to be? 25 cents, maybe less, right? So everyone's going to say, “I didn't read the DeepSeek paper, but I'm sure they're doing it.” So they're all rushing to see, “Okay, I've got to push this,” right? And then Fin just got bought for $3.6 billion, right?

I've seen a lot of data. I'm seeing a lot of plateauing, and that may push people back to limited, more high-end models so that you can get to the next level, so that you can get to 95% true resolution of complex problems instead of, no matter what the internet says, 40% resolution of problems that aren't that hard to solve. Right? We'll see whether, over the next 6 months, now that we've all internalized it, we get all the benefits out of it that everyone thinks we'll get.

Guest 2

Two comments on that, though. One is, yes, you're right. But I think the point that Decagon's CEO was making is it's not just pricing; it's also latency and response time. There's a bunch of reasons, but I think the meta-point is this: I'm going to make a comment on the CX space.

Everything you said is correct. What I love about it is, if you think about the chasm concept, this is a market that's on the other side, the positive side, of the chasm, because implicit in everything you said, Jason, was a recognition that there is an ROI there and the shit works. One of the reasons I like this space is that a lot of these other apps companies are wrestling with, “How do I price per outcome?” These guys—this market—has already gotten to the point where the customer, not the vendor, says, “I get it. I can increase my resolution rate from 30% to 65%. I get it. That's worth it. I spend $3 an email to answer a query, so $1 or even 50 cents on customer support is well worth it.”

In other words, it's moving from the experimental side. There's a lot of talk about 95%, whatever shit, and AI ROI is not there. This is a category where everyone can articulate the 30% where it is there, and then, to your point, they can go, “Oh, the next 10% is going to cost more.” That's a high-class problem. Maybe what you're saying is you go from 30% resolution to 65% resolution at a buck a pop, and maybe from 65% to 75% it's 2 bucks a pop. You'll still happily pay it if you're the customer.

Guest

Probably. It's just going to be another stage in the evolution of AI, right? You can either say, “Listen, I got 20 cents of that 50 cents to provide the best resolution I can,” right? And that's a great answer today, but as your competition gets smarter about this and blows by you, it's going to create an amount of competitive pressure that will just be interesting, because we'll all have to get much better at this stuff.

9. Microsoft & Amazon Embed Engineers in Enterprise

Does it ultimately provide the value? Microsoft launches with $2.5 billion and 6,000 people to embed engineers inside enterprise clients, targeting the MIT finding that 95% of enterprise AI pilots deliver no measurable P&L impact. What a positive finding that was. Amazon made the same move 2 days earlier. Is this a continuation of the shift from a model to a services ecosystem? How do we think about this? I think it's going to fail.

Guest 2

Good take.

Guest

Yeah, I'll tell you why. We work with a lot of FDEs at SaaS companies because we have so many agents, right? We're working with—we have, like, the best FDE at Salesforce, the best FDE at all these folks. Not the best, but we have some of the best at all these companies. And they're effing great.

Literally, the FDEs we work with at these leaders are better than anyone I've ever worked with in customer success or support in my entire career, with maybe 1 or 2 exceptions. They are so good—the best FDEs at these companies. Then one leader—not sales, another leader—our FDE went on paternity leave for 3 months, and the new one told us they couldn't fix our bug for 3 months until the first guy got back. This was a leader, an FDE.

So my point is, I think this is going to fail because I don't think there is enough talent to do what we want to do in the enterprise. The idea makes sense on a spreadsheet. It makes sense. Rory's smarter than me, but my experience today is that it's going to fail with all the companies we work with because there's not enough depth to do it.

Literally, this is a public company that said, “You're going to have to wait 3 weeks to fix the fact that your AI is still talking about SaaStr 2026, which already happened. It happened in May. It's now July. We are going to have to wait until August to fix that bug until our better FDE comes back from paternity leave.” Think about this. This is not someone who was hired last week. How the hell are you going to scale this? Wait 3 months to fix the fact that you're talking about an event that already occurred 60 days ago? That's an F, isn't it? I'm going to throw 10,000 people who were terrible at customer success into solving massive enterprise problems. Good luck. Good luck with that one.

Guest 2

I disagree. I think it will work in a limited but interesting sense. Stepping back, I don't buy, for the record, that 95% of MIT's findings mean these things fail.

Guest

You don't buy the story? I literally just told you we were told by this public company leader that it would take 3 months.

Guest 2

I totally buy that story. I'm saying I don't think 95% of these things fail, but I do buy your story, Jason, which is that all these companies—even a smart company like you, and you're way more technically adept than 90% of corporate America—need assistance to make this shit work, right? And it's obviously very bad that they couldn't answer that in 3 weeks.

But the solution is not, “Don't get that support.” The solution is someone has to build a business whereby they have, say, 2 people capable of answering your questions. Right? And the big zoom-out question, because I actually didn't see this until I thought about it but I'm now clear on it, is who's going to meet that need?

Who, if corporate America is going to adopt all this stuff—and they're who they are—they're an oil and gas company. They're a banking company. And then, on the other side of the table, you have Anthropic and OpenAI, who are product companies to their core. You're going to need something in the middle: services companies to help them adopt.

And what's interesting is what's happening to Microsoft. Every technology company either goes bust or lives long enough to become the next generation's IBM. IBM was the enabler for the PC and, to some extent, the cloud, helping corporate America adopt. When you don't have an amazing product yourself, but you do have large enterprise trusted relationships, what you do is sell to those relationships the ability to adopt new technology from other people.

To some extent, that's what IBM's been doing for the last 20 or 30 years. IBM Global Services has been all about, “We don't build e-commerce. We don't build any of these cool things, but we'll help you adopt.”

Guest

Yeah. And we'll launch your product in 2030, but that doesn't work today.

Guest 2

Jason, to be clear, I'm not saying IBM is amazing, and I'm not saying Microsoft would be amazing at this. I'm saying that—and this is a harsh comment from Microsoft wrapped in a positive one—they're no longer the company with the new technology. Other people built consulting services to help adopt Microsoft 30 years ago.

Now OpenAI and Anthropic are the companies with the new, incredible product, and Microsoft is the more mature company with the enterprise relationships. It's going to build a large services business just like HP did, just like IBM did. If you went back and read those press releases from 20 years ago, you know: “HP, your trusted partner in global services.” IBM, same thing. It would read exactly like this.

And the summary is, “Mr. Corporate America, you need to adopt this new technology. Those dudes in Silicon Valley are pretty scary. You've never met them before, and they talk about crazy shit like the end of the world. We've been selling stuff to you for 20 years. You trust us, we trust you. We're going to make this work.”

To your point, Jason, you're right. They might make it work great, but it will be better than the enterprise trying to do it on its own. So, in summary, I think Microsoft will build a huge services business here if they want to, which also speaks to the fact that it won't be nearly as profitable as selling operating systems.

Guest

I actually think they're both right. Stepping back, we literally work with the top 1 or 2 or 3 FDEs at so many vendors, and I can tell you the depth—even at some of the best companies—the depth is not there. These are not old companies. There is no depth to the FDE chart.

So I know that this is going to fail. But Rory's also right: it is better than doing it yourself, which is often hopeless, right? But I am right that the depth today just doesn't exist. So a lot of board members and folks not close to it are going to say, “Let's go do this.”

A lot of VCs are trying to invest in AI-enabling businesses, old businesses, right? And I believe if you could attract the talent, this would be great. I just don't know that there are a couple hundred thousand people who want these jobs and are off-the-charts smart. You're lucky.

In some ways, we're back to the early, early days of B2B software, where you'd have a couple of folks who kind of understood how it all works, and no one else could solve the problems on your tool.

We’re back that way with a lot of these agentic products, I think.

What that means, by the way, is that for the model companies, the rate of adoption of their technology is, to some extent, a little bit outside their control, which is why they are doing these services businesses. The biggest problem with Exxon or Bank of America rolling out generative AI is not their ability to buy from Anthropic. It’s the ability to do change management and application building in the enterprise. And that’s going to be solved by large, trusted partners who deliver the services and the expertise.

For the record, I think the interesting point on this, circling back to demand, is that if Jason is right and the quality isn’t there, that means the adoption cycle will be longer. The biggest single question on all of this is: what’s the rate of diffusion of this technology? For the last 3 years, it’s been way faster than the diffusion of any other technology in history. The rate at which OpenAI and Anthropic got to $4 billion and $12 billion, respectively—other way around, sorry, $12 billion and $4 billion, respectively—in revenue was never before seen, right? If the next 10x takes 3 times longer because corporate America can’t adopt, that’s going to have consequences. And I think it’s the big question: how quickly can that spend on Anthropic go from $4.5 billion to $40 billion to $80 billion?

Guest

I don’t know how much it’ll impact the top line, but I definitely think you’re right, Mike. This type of rollout is going to be slower than folks hope, right? There’s just not enough talent to do it. Whether that really stops Anthropic in the aggregate is a different question, right?

Guest 4

Nothing stops you. It’s not a stop; it’s a slowdown. You’re right, Jason. It’s a slowdown. It’s a question of how fast.

Guest

I’ll tell you what, just as an aside—sorry to go into the details you don’t want to hear—one thing I learned that was really interesting: at our SaaStr AI Annual this year, we had a CPO panel. We had the CPO of Harvey there. You guys might know this, but I learned something: every deployment they do at Harvey has an FDE and a lawyer. Every single deployment has a lawyer, right?

To the extent Harvey can bring in—and I’m sure they do; I want to go deeper on this—to the extent they can bring in consulting firms and Microsoft to deploy and maintain that, it’ll work if they’re a 3-way team. But it’s just interesting: if you have a lawyer and a very experienced technical resource deploying Harvey, which has a high price point, right, you can afford it. That might be what you need to have a successful deployment there.

Can you roll this out to generic B-tier or C-tier people today? It may not just be successful, but that clearly works. It’s Captain Obvious, but I don’t think most of the companies we work with deploy a deep subject-matter expert and an FDE at the same time, together as a team. It’s a good point, and it makes sense because, if you think about it, when all you’re buying from the vendor is a database, all you need is a database expert.

But when you’re buying from the vendor intelligent answers about your own business—and if you’re running Exxon, you better be damn sure that those answers are grounded in oil-and-gas facts—you’re right. Actually, it’s interesting: probably everyone will be some combo of tech expert and domain expert.

Guest 4

And that’s why these services companies will be tricky to build, to your point.

Guest

Yeah, they go in, they learn about how your entire law firm’s business process works, and they map it against Harvey, right? That’s very—I would love to—that would be great if these services companies can do it. I’m skeptical, but maybe.

10. Ashton Kutcher Leaves Sound Ventures

Ashton Kutcher, one of the most successful investors of the last few years in terms of SPVs and OpenAI, Anthropic, and Sound Ventures—obviously his firm—announces he’s leaving his own VC firm, and he’s going to start a new VC firm with Morgan Beller, previously at Andreessen Horowitz and then NFX, and now starting her firm with Ashton. It’s a notable move in the world of venture, I guess. A new firm, one of the biggest AI investors of the last few years. Jason, what did you think?

Guest

You know the gossip. I need to know what really happened. I mean, on its surface, it’s just crazy to leave your own firm, right, like this. It’s one thing if you’re managed out or something like that. That can’t be the case here, right? I mean, this is the guy from That ’70s Show. We need him in the fund, right? Maybe he was managed out, but I find that unbelievable, right? Unbelievable to leave it behind like this.

It’s interesting, right? In a way, it kind of reminded me of Jack Altman raising a massive amount for a solo GP fund and joining Benchmark. These are things that make sense today, but when we almost even started this podcast, they wouldn’t have made sense. It’s like, “What, Jack? Why?” And I love Jack, but why would you raise $500 million and have LPs dying to fund you and go join Benchmark, right? Because it makes sense in 2026.

And, Jason, Sound Ventures has raised a lot of money.

Guest

A billion dollars. They’re in some good names, right?

I’m being facetious.

Guest

No, no. I mean, yes, they’re in some excellent names, and, frankly, we’ve co-invested with them in some deals. They’ve been wonderful to deal with. I actually think it’s simpler than this. I could be wrong, and I’m usually a cynic in venture, but I actually don’t think there’s a deep, dark story here. I don’t think any of that applies.

I think this guy is so successful. Why do venture firms hang together and paper over the story? Because the asset is the firm and the name, and even if you hate each other, you want to manage the process well so you can keep the thing going, because the firm has a brand and a reference. I know many situations where, effectively, partners look at the other guy and say, “I’m mad at you. You’re mad at me, but we’re going to hold this thing together.”

None of that applies here. He’s a Kutcher. He doesn’t need to. I mean, I knew who Ashton Kutcher was before I knew Sound Ventures, right? If he wants to go do something else, it’s just cleaner to say, “Now I’m Ashton Kutcher doing this. I’m doing, I think, very much seed, pre-seed, deep tech. It’s a new thing.”

I think very few people are in the position whereby the name is such that they don’t have to worry about the firm brand. They just say, “I’m a famous person who’s”—for the record, I’m sure most people, when he started investing, would have had a little sneer, and he’s killed it. He’s a famous person who’s now been a brilliant investor, and now he’s a famous person doing deep seed.

So, I actually get the impression, from the vibe from the folks I’ve talked to at the firm, that there’s much less of this angst than you think, and just 2 people wanting to do different things. Because, look, a lot of the brilliant OpenAI and Anthropic investments were late—obviously, multibillion-dollar pre-money valuations—which is very different than deep tech.

I think the beauty about being a famous, rich person in America is you can pretty much do whatever you want. And if you’re TV-famous and movie-famous, you don’t have to worry about your brand or the firm’s brand name. There are very few investors where I know the name of the investor, and it took me months later before I figured out the name of the firm. This is one of the few investors on the planet where it’s the investor name.

To this day, if you said to a bunch of people, “Who’s on your cap table?” and Ashton Kutcher was an investor, they’d probably say, “Ashton Kutcher’s on my cap table,” right? So, the name doesn’t matter.

Yeah, but there are only, like, 6 brands in venture anyway. That’s not exaggerating how many brands are more famous than all of us. So, move on. It’s like, you know, I’m willing to bet there are more. If you check Google Trends, he gets more searches than Sequoia without even blinking, because 330 million people have some sense of who he is and maybe 3 million know who Sequoia is.

Guest

I’m with you. Listen, we can move on. Even for me—and listen, I’m a solo GP who would not deal with any of this crap today—I would, if I had a CFO that was working, if I had investor relations working, if I could stand my partner, if I liked my partners, if I liked coming to work. I would stick with my entity, even if I had to get some of them out. I would stick with my entity if I liked all the stuff around it.

It’s not that you can’t rebuild everything, right? It’s not that there’s no equity in the brand, but if the engine is working, I’d rather just stay. I’d rather just stay.

Boys, you can choose one more topic. What topic should we discuss?

11. ElevenLabs at $22B

Guest 4

Well, look, I think you hit a lot of good stuff. The one that maybe we’ve discussed before, but I still think is a topic that resonates, right, is the ElevenLabs secondary. I added this one: the ElevenLabs secondary at $22 billion, right?

Guest

I don’t think it’s a high valuation. Maybe that’s interesting, but I think the growth in today’s world is consistent with other rounds, right? I don’t think the price is actually that interesting. I do think, even though it’s not a new topic, the one I said is interesting. It’s like, as an employee today, why would you join something that you don’t believe will have secondary options? I really think this is a big issue.

Guest 3

One issue is, why would I join you rather than Anthropic, right, where I can make so much money at OpenAI? But there are plenty of reasons not to join Anthropic and OpenAI. We could talk about that, right? They’re pretty big companies, so your role is going to be very narrow, right? It may not be the job you want.

ElevenLabs is probably more agile than Anthropic or OpenAI, right? Your job is probably a little bit more interesting for some folks. But Jesus, if I were a hyper-talented employee, I would not want to go somewhere without liquidity. It just doesn’t seem worth it today. And so it’s a question: do you have to create this as founders? What do you do if you’re close to this level? Because the liquidity is thin. There are only so many companies like ElevenLabs that can pull off a tender offer at $22 billion, right? They’re there, and then—but if you’re not quite at that level, they go away. Well, do they? You see, Clay—I know, I know it’s much smaller. It’s $5 billion, but they did a tender offer at $5 billion. And so—

Guest 5

Yeah, that’s the minimum. I’m not saying there’s some line where you can pull it off, but maybe—and I’m not saying this literally with Clay—sometimes with something like Clay, even next year, you might not be able to pull it off, right? What if it’s a little bit soft, right? There’s only a handful of companies that can always pull it off like clockwork, right? There are only so many Databricks and OpenAIs.

But that’s a good question: why would I join anything sub-Clay? Because even if the nominal valuation is $3 billion instead of $5 billion or $2 billion, if there’s no regular liquidity program, why would I join it? Why would I join the startup? Life’s too short, man.

No, [snorts] I disagree because it’s incorrect framing, Jason. The point is, if you join something that’s already doing tender offers, right, then you’ll get an equity grant reflective of the fact that we’re already doing tender offers. So it’ll be slower. If you join something that’s never done a tender offer ever, then you’ve lost. The whole trick for employees, just like it is for VCs, is to join something that isn’t doing a tender offer today, get a healthy grant, and join a company that, within a year or 2, when you’ve vested 50–60% of your thing, starts doing tender offers.

So I’m just saying it’s a slight nuance. You said, “Don’t join anything that isn’t doing a tender offer.”

Guest 3

But I just misspoke. I meant to make the exact point you’re making. Why would you join anything where you don’t have high certainty that they’re not just going to be a unicorn? It’s not good enough that they’re going to have tender offers in the next 24 months.

How can you know, for people listening, if something’s going to have tender offers?

Guest 5

I mean, 2 years is hard, but I think in the end—look, it doesn’t actually change things all that much. It’s the same as whenever you join a startup, right? You’ve got to join startups that have the potential for big upside. And look, 10 or 20 years ago, it would be “go public.” Now that window takes 12 years in some cases. So you’ve got to have something else, and tender offers are the proxy for going public.

Anyone who joins a startup does it for 2 reasons, and I think you have to start with mission. The second one definitely is the chance of a payday, right? So I tell everyone—and it’s funny, I say this to them—I tell everyone that I operate on the operational side, “Hey, you’re a single-shot VC. You’ve got to pick only 1 deal and get it right.”

And I always say to them, “Look, when you come to choose a couple companies, and if you want any random VC input, feel free to ring me and maybe I can give you perspective.” Very few people do. It’s just funny that way, right?

I think actually one of the things I often look at is how operators make decisions. And there are a lot of things that get fed into it. Maybe they’re perfectly good other reasons: you like the people you’re working with, you like the market, it has a mission. But from a pure stock-picking perspective, Jason is right. The mission—the job at hand—is to pick a company that within 1 to 3 years will be a unicorn, will be tender-worthy. And then you make out like a bandit.

And it’s a hard thing to do. We get 20 shots. And I mean, I feel guilty, almost. We get 20 shots in goal, and they get 1.

Guest 3

Well, if you’re leaving every year, you might get 20, depending. That’s true. I’m not quite sure. It’s just—they’re sequential. Employees are sequential venture capitalists, right? They’re just sequential rather than parallel.

Damn those vesting schedules, fucker.

Guest 3

Well, now that there aren’t even vesting schedules at OpenAI and Anthropic, those issues have been solved, right? A lot of startups don’t have vesting schedules for top employees, right? It doesn’t mean you vest into all your stock, right? Sorry, they don’t have cliffs. They have vesting. You don’t have a cliff, right? That problem’s been solved by eliminating cliffs, right?

Boys, this has been fantastic. I’ve so enjoyed this. It’s so nice to be back in the studio. I was not enjoying the holiday setup. I like to be in the studio for this. But you’ve been awesome, so thank you so much for joining me. And Roy, we’ve got to let Jason go back and deliver insight to his portfolio.

Guest 5

Yeah, they need those profound insights…

…that they can’t get on X.

Guest 3

Have you guys looked at open source?

Guest 5

Have you thought about managing your token spend a little bit better?

Can we increase sales? Is that it?

Guest 3

I’ve long since internalized this, and I tell my CEO I’m actually not here to give you insights. I’m simply here to say that if we’re driving the thing off the cliff, I scream, “Stop.” That’s probably the only value add. Other than that, you guys are going to figure it out. That’s the way it works.