Google’s AI Brain Drain, SpaceX's Huge Quarter, Airtable’s 90% Collapse, US Data Fuels China AI

It's the summer. It's August 6. We're having a hard time getting a quorum here on the podcast, but David Friedberg is here. David Friedberg is back, our sultan of science. How are you doing, brother?

Great to be with you.

It's great to be with you. Everybody loves when Brad Gerstner is here. He's your Bruce Wayne if markets are your game. He brings that namaste to your payday.

Yes.

All right. Welcome back to the program, Brad.

I love it. You're bringing the rhymes back.

I bring a little intro back. We've been trying to bring it back. Chamath is on the road right now, but we will get a field report from Chamath. I called Daniel. Somehow, Sacks is going to be here, but you know how he is. He's always late because he can get a phone call from very important people, but he will break in at some point. Oh, wait, I see it in the text here.

Oh.

Oh, there he is. He made it.

Hey, guys.

You made it.

How do you like my beautiful summer gilet?

It's incredible. It fits perfectly. You look warm.

I don't know how Chamath does this.

Chamath Palihapitiya

We'll let your winners ride.

Rain Man, David Sacks.

Chamath Palihapitiya

I'm going all in.

And instead, we open-source it to the fans, and they've just gone crazy with it.

Love you, Wes. I think Queen of Kin Mob[?].

Chamath Palihapitiya

I'm going all in.

Well, here's the report, everybody. As everybody knows, Chamath is on the road. Oh, here he is. This is a photo, Sacks. He went to check his data center progress. I think that's in Colorado or Nevada, wherever he's building a data center.

I think that's on Dune.

Oh, it's on Dune. Yes.

It is.

Dune 4. Oh, yes. There he is admiring himself with Blue. Oh, look. Here's Nat. You know when a meme has reached its peak when your wife starts dunking on you. There it is. And here we are. This was at the Christmas party, I think. Oh, Brad, you were on CNBC with Andrew Ross Sorkin. There you go.

I wasn't sure if it was my Twitter feed that was just selecting into it, but it clearly hit everyone, right? This was a viral thing.

No, no, no, no. This has hit everything. All right, listen, we have a lot to get to. Enough with the shenanigans and small talk.

1. Google Loses Its AI Brain Trust

Google had 2 major shakeups to its AI staff on Wednesday. Demis Hassabis has moved to chair of DeepMind and chief scientist at Google. Reports describe this as Demis stepping down or being kicked upstairs. We'll get into that, but Google framed it as a promotion and says he was stepping up.

Here's Axios' quote explaining the shakeup: “Google's Gemini 3.5 Pro is months behind, with some company sources telling Axios that it's in part due to low morale.” Interesting. Several top researchers, including Gemini's co-lead, have left the firm for competing AI labs.

Jeff Dean plus 3 other AI superstars are leaving Google to start a company called Discovery Loop. Dean is a legend, Friedberg, and I think you worked with him at Google—one of the world's great AI engineers. He was employee number 30, joined in 1999, and has worked there, from what I understand, continuously for 27 years. Discovery Loop is going to be focused on deep scientific breakthroughs in AI.

Google shares were down 4% on the news of Dean leaving—$200 billion in lost market cap, if you want to correlate those 2 things. Friedberg, this is your alma mater. What are your thoughts here? Is this creative destruction? Maybe these people weren't delivering and they wanted fresh blood. Or is this just the siren call of doing a startup in an age of unlimited capital for AI and unlimited opportunity being too much for the OGs at Google not to take advantage of?

2. Infrastructure Beats Model Development

Maybe it's the 3rd bucket: if you're the board and management, you're having a debate about how to best deploy capital. Google has made a commitment to deploy $200 billion in CapEx this year in AI infrastructure and data center build-out.

Because of the CapEx and accelerated depreciation, making an investment in AI compute in the United States right now is hugely tax-advantaged. Because of the extreme demand for compute, it's a pretty obvious ROIC model—return on invested capital. If you make this sort of investment, you have significant demand for that compute infrastructure. You're very good at running the compute infrastructure. That capital can deliver massive profit returns for you with very high confidence in some forecasted period.

Building the most advanced, frontier lab-driven model also takes tens of billions of dollars of capital. The question really is: can you deliver the profits from the model? In a world where open source is becoming so good and open-weights models are catching up so quickly, and all the frontier labs are catching up to each other so quickly, does it really make as much sense to deploy tens of billions of dollars against building a model?

Yeah.

I think that the scientists we're seeing transition out are the scientists who have been at the core of model development, of making these frontier models, and they were certainly first out of the gate. You can look at some of the early interviews with Jeff Dean from a couple of years ago, where they actually had a ChatGPT equivalent internally a year before ChatGPT came out from OpenAI. Google chose not to release it for fear of cannibalizing Search and so on.

That's when Sergey Brin stepped in, and there was this whole revitalization. But as time has gone on, and as everyone has competed on models, as we've talked about many times on the show, I think it's pretty obvious that it is very hard to get the same sort of return on capital invested in model development as it is in capital invested in compute infrastructure and being model-agnostic.

What Google has is probably one of the greatest installed enterprise bases in the world for compute. They have the most enterprise customers. They have the most consumers. In both cases, they don't necessarily need to have the best model to make an incredible business. They can be model-agnostic. They can work with Anthropic. They can work with OpenAI. They can work with SpaceX. They have a significant ownership stake in SpaceX and in Anthropic, and they can work with all the open-weights models. They can host them all.

Now, if you're one of the great computer scientists—you're Demis, you're Jeff Dean, you're this whole crew—and you're inside Google, and they're allocating capital not to your models, not to the things that you're most interested in, but to infrastructure and data centers and supporting the broad ecosystem of models, you start to say, “Given the fact that I can go down the road and visit Brad Gerstner and a couple of other people and raise a couple billion dollars at a multibillion-dollar pre-money with a PowerPoint deck because I'm the greatest in the world at doing this, that might be a better path for me.”

And I think that's the moment. The way I would frame it is: CapEx is high alpha, low beta in data center infrastructure, that capital. And model development theoretically could be high alpha, but it's very high beta. It's a very risky way to deploy capital.

Totally.

So if I'm the board and management, I'm deploying more capital in compute and infrastructure and less capital into model development. That's what I think is going on.

Brad, what's your take on this?

I think David nails it. The same thing's going on at Microsoft, right? Satya is out this week citing Morgan Stanley's report and saying they're seeing over a 30% return on invested capital in tokens as a service—in the infrastructure business.

I think David's exactly right. Those are such good businesses. You deploy capital, and everybody's renting it from you. But the scientists who want to be involved in superintelligence, who want to cure cancer, who want to be on the frontier of these models, are sitting there dealing with this channel conflict at Google.

Google Cloud wants all of the compute in order to rent it out to Anthropic, while those building the frontier models internally want that compute in order to compete with Anthropic. So you have this inherent channel conflict between those wanting to build the models.

I think David said it really well, and I think that's a big challenge for them. It looks like it's being resolved in favor of being more of an infrastructure company.

Telescope out for a second. SpaceX also reported this week. They also have channel conflict. They're renting out their compute to Anthropic. At the same time, they're trying to build their own model with Grok and Cursor.

You have that channel conflict at Google. You have that channel conflict at Microsoft, although I don't even really see them pushing the frontier anymore in terms of models. Meta is talking about getting into the infrastructure-as-a-service game, and then at Anthropic and OpenAI, you don't have any of that channel conflict. They say, “We're not in the infrastructure business. We're only in the model business.”

I think it's a clarifying view as we look forward that we may, in fact, not have those companies on the frontier of model development if all these people leave.

By the way, thanks to the law passed on CapEx depreciation, if you assume a 26% corporate tax rate, every dollar you deploy in CapEx—because you get to write it off this year—you're basically getting 26% off.

That's money—

Pretty good deal.

—you get right back. Yeah.

3. The Frontier Model Duopoly

Polymarket

Which companies will have the number-one AI model by the end of this year, on December 31? Of course, the last time they did this, Anthropic won, so they're not on the list. They're the winner, but who will have it going forward? OpenAI is at 32%, Google at 20%, Alibaba at 14%, and then Moonshot AI, xAI, Meta, and ByteDance are all at about 10%. So, Sacks, your thoughts here on what's the better business? Is the better business being in the language-model, frontier-model business, or is that getting quickly commoditized? Do you really want to be in the token-sale business, or is that also going to be a commodity and you just need to be on the application layer?

Here's what I think is going on in terms of the market structure. When I saw this Google news, my reaction was, “And then there were two.” Because, like Brad was saying, we used to have 5 major companies in the hunt to be the leading frontier lab, the leading frontier model, just a year ago. Now we're really down to just Anthropic and OpenAI.

So the market for frontier intelligence has become a duopoly. Elon is still in the hunt, and I'm sure Google would say they're still in the hunt. But, like Brad is saying, they may have contradictory incentives there because they can actually do quite well just with their compute. So I think that the market for frontier intelligence has become a duopoly.

I think it's a very powerful duopoly. I don't think it's being commoditized. I think what we're evolving toward is a 2-tier market structure where there's a market for frontier intelligence and there's a market for, let's call it, commodity or lagging intelligence—whatever you want to call it—that's 6 to 12 months behind. There is a market for those tokens, those models, but the reality is you can't charge anything for the weights.

You can charge for the compute. You can charge for the inference that you're providing. You can charge for essentially consulting services to help put the whole thing together. But if you're not at the frontier, you can't charge for the model layer itself. If you are at the frontier, you can charge a premium, and that's where Anthropic and OpenAI are.

I think the proof for this is just that you look at the growth rates of these companies. The latest we heard is Anthropic is now over $80 billion of ARR. It started the year at $10 billion. It had forecast $100 billion as exit ARR for the year, and most people said that would be impossible to achieve. Now it looks like they're going to do it with a couple of months to spare. So their estimates are going up—I mean, $110 billion, $120 billion, or higher for end-of-year ARR.

OpenAI is seeing acceleration. So I think what you're seeing now is a very clear bifurcation in the market. You've got a frontier-model duopoly that can charge a premium. I think of it like Apple. Apple's competing against Android. It's open source. Android actually has more users in the world, but all the monetization goes to Apple because people are willing to pay for the premium experience.

I think, in a similar way, people are willing to pay a premium for true frontier intelligence if it's really at the leading edge. But if you're not at the leading edge, there's a huge market for that too, but it's highly commoditized. People are just willing to pay you for the compute.

Yeah.

That's what I see happening right now.

Yeah.

Jason, what do you think?

Well, if you look at Google Cloud, they posted 82% year-over-year revenue growth, which is something we've never seen in the history of these cloud providers. Elon Musk and xAI just had the SpaceX earnings. We're going to get into that, but they also had a massive uptick in their Elon Web Services, as I've dubbed it.

If you look at Google, I still think Google will be the number-one AI company because they have so many people using AI inside their products already. They have 5 products now with over 3 billion monthly users each. Android, Search, Gmail, Chrome, and YouTube all have over 3 billion monthly users.

If you've used any of these products recently, they are becoming AI-first products—YouTube especially, but obviously Chrome and Gmail. You're seeing tools pop up there for AI. Friedberg, you kind of alluded to this: They have 13 products total with over a billion users, and that now includes Gemini. In Q2, Gemini had over 950 million monthly active users, tripling year over year.

They will be the number-one AI company in terms of consumer usage by far, I think, this year. That doesn't mean that the frontier models aren't great businesses. They obviously are. But I have been using exclusively non-frontier models, and for 95% of the jobs I'm doing, Sacks, it's good enough.

I just posted about this, and Elon and I got into it a little bit. I think you referenced this in our group chat. I tweeted just the other day, “The difference between the open-source models I'm using and frontier models is negligible already.” I believe that to be a true statement for the work I'm doing, and Elon responded back to me, “It's actually a world of difference.”

If you're doing something other than making a copy of a video game, or you have incredible speed needs, the frontier models are not necessary anymore. They're just not necessary. The people using the frontier models are doing it because their company set it up, and it's too hard to implement open source right now. But it's going to get easier and easier to implement, so I'm still going with open source and Gemini being the leaders in this space.

Yeah, look, I think it's true for your use cases that the cheaper commodity intelligence, that middle of the market, is good enough.

Yeah.

Look, an Android phone would be good enough for me. I could get by on a cheap Android phone. You know what? I still pay a premium for this because I use it so much.

So if you're a business—for example, a hedge fund—in a highly competitive industry, you don't want to take the chance that you're not getting the best intelligence to power your models. There's a lot of industries like that where the competitive dynamics will drive you to pay for the best intelligence.

There are also situations—this goes back to the blog post that Decagon posted—where, if you're looking for use cases, you also want to use the true frontier. When you're dealing with immature use cases, you don't know where the value is going to be, and you're searching for an opportunity to use AI. You just want to use the best because the return on finding those use cases is going to be so much greater than the small premium you're paying at the token level.

I think there are a lot of examples like that when the use case is immature, when you're in a competitive industry, or when you're just deploying AI. You want the convenience of the full stack.

So why not go with frontier models, to summarize your argument?

Yeah, and look, again, unless your employees are doing something stupid, like you create a leaderboard and they're token-maxing, I don't think the cost is that great. The benefit that you're getting is huge. A lot of people are just like, “Give me the best. I'm willing to pay a premium for the best.”

I'll take a slightly different take. I think that it's not necessarily whether you take the best model or the open-source model. I think that there's a blend that's happening. At least that's what I see.

For example, we'll use open source, open weights for a vast majority of simple workflow applications. But when it comes to specialized applications where we really need to have high-quality model proficiency—for example, in life sciences or genomics modeling—I'm going to go for the premium model. If I'm working at a media company and I'm trying to do AI rendering of video, I'm going to use Gemini's video model. It is the best model, or Sora, or whatever the best model is for that particular application.

So I think the idea that there's a model that you pick for everything is the false assumption. On the consumer side, it's likely the case that consumers are not going to be using some open-weight model because they can pay $20 or $40 a month and get ChatGPT, Gemini, or Claude and be very happy paying $40 a month. They'll basically be able to minimize their cost to run that for consumers.

For enterprise, I think the enterprise is going to be very active in selecting a blend of models that are going to make the most sense: a very cheap open-weight model for simple workflow applications, individual employees spinning up an app, whatever; more complex models for those really key workflow tasks; and then specialized models.

And I will say it is way too early to count Gemini out on building—

For sure. 100%.

—incredible specialized models. They have the best video data. They have the best life sciences data. They've been working on this for far longer than Anthropic or OpenAI on the life sciences side. They're very well ahead on that front. Demis is still going to be running Isomorphic Labs.

So when it comes to these specialized models—verticalized, specialized models like video, life sciences, and protein folding—I think these are the things where you're really going to see Gemini shine. Then every enterprise is going to have a mixture.

But hey, if you can be the cloud service provider with that mixture of models, which is what Google Cloud Platform can now be, I'm going to sign up to work with GCP versus working just with Anthropic.

4. Token Prices Keep Falling

One of the things this has created is downward pressure on pricing. We saw OpenAI and Claude do massive price cuts for tokens, so they are reacting. They're not taking it sitting down, and the orchestration between these models is being built into a lot of harnesses inside of enterprises. So what's your take on the downward pressure on token pricing, or is this just great for consumers and enterprises because—

Listen.

We've got massive competition.

That's the thing. America's winning. This is exactly what you want. We have a massively competitive market. We have Chinese open source, domestic open source, frontier model labs that are doing what they're doing. We have downward pressure on pricing.

You know, David referenced a duopoly. I think it's hard to call it a duopoly when you're only a few years into this and you have giants like Amazon, Microsoft, and Google. I do think he's right. I do think they've emerged as the pure plays. Their revenues would suggest that they're gaining share of wallet.

But there are 2 points I want to make here because I think they're non-consensus views that were spoken this week. One was Elon's response to you, Jason.

Hmm.

Right? Over the last 2 weeks, everybody's been saying that the Chinese have caught up, that open-source models have caught up in intelligence, that they're much cheaper, et cetera. And Elon comes out and says, "Not so fast. We're entering the singularity, and the frontier models are way further ahead than people think."

I believe that to be true. I think for your use case, they're very similar, but I don't think that's the most sophisticated use case that people are trying to train on and trying to experience. And then Jensen came out this week and said closed models are actually cheaper if you don't have to build it for yourself, if you don't have to pay the training costs and have a lot of expertise to fine-tune, maintain, guardrail, and keep it safe.

So he's basically making the argument that not only are the frontier models further ahead, but that the cost differential between the two is not what everybody's making it out to be, which I think explains why they continue to run away with it on the revenue side of the equation.

But I think we have healthy competition. You're right, JCal. For the vast majority of use cases, I think token consumption is going up for the open-source guys, while share of economics is going up for the frontier labs. I think that's what we want to see.

Yeah, and it's just Android versus iPhone all over again. One platform makes the profit. One gets the majority of users, at least globally, in usage.

5. SpaceX Posts A Huge Quarter

All right, let's talk SpaceX here. They had their first earnings report as a public company. Shares dropped 13%, I think because people were a little concerned about the surging AI capex. It's down 30% since going public in June, but it seems to have settled in at a $1.4 trillion valuation. It went public, obviously, above $2 trillion.

Q2 results were spectacular, is the only way to put it: $7.8 billion in revenue, up 92% year over year. Let that sink in. And 67% quarter-over-quarter growth in AI revenue. xAI Web Services more than tripled quarter over quarter to $2.6 billion. That's not Cursor; that hasn't closed yet. But that's going to be one of the great purchases in history.

This is from Elon Web Services, renting out compute specifically to Anthropic and Google from Colossus, a collection of servers. But capex was up $18.4 billion in the quarter. That's 6× year over year. Obviously, you can do the math there for a run rate of about $75 billion.

I'll stop there and get your reaction, Brad, to the SpaceX IPO. I know you've been tracking this and commented on it heavily.

Yeah. Listen, I think that—first, let's start off—$1.4 trillion of value creation for this company is extraordinary. So the fact that from peak to trough it's down 40% or 50% from the IPO—we had that chart out a few weeks ago. Remember that within 6 months of the IPO, almost all these tech stocks are down 50% peak to trough. We see it again here with SpaceX.

I thought it was a really solid quarter. I thought his guidance was pretty extraordinary: $100 billion in ARR by the end of the year, and he pulled forward the $1 trillion target in ARR by a year, from 2031 to 2030. Now, to just put that in perspective, Morgan Stanley's 2030 revenue estimate is $325 billion, which is also extraordinary. Remember, this company did $18 billion in revenue last year.

So whether you're taking Morgan Stanley's numbers or Elon's numbers, clearly the market is not pricing that in. At $2 trillion, we were pricing ahead a couple of years. I think now the value reflects kind of where we are.

The market has questions about a few things. Here's what they are. Number 1: on the rental business, the rental of compute business, he rented out a huge block of compute to Anthropic. It's the question that we've been talking about here: are you going to use the compute to build your own frontier model, or are you going to rent it out? And if you rent it out, are you going to be able to find those people who have the capital to offtake that compute?

He's talking enormous numbers, 10 to 20 gigawatts, and people are wondering how they're going to be able to finance that. And remember, those businesses, the GPU rental businesses, tend to trade at very low multiples. Look at CoreWeave, et cetera.

On the frontier-model business, I think this is the sleeper. I think he said on the call that Grok tripled tokens in the month of July. That doesn't include Cursor. Cursor was already on a path to go from $3 billion to $10 billion by the end of the year. Cursor plus Grok could be at $10 billion to $20 billion by the end of the year. That would be an extraordinarily valuable asset, going to trade at a much higher multiple than the data-center business.

And then, of course, we haven't even talked about Starlink and what he's going to do. I think he's going to run the table on mobile. So this is the normal consolidation. We have funds across Silicon Valley that are distributing their shares. The stock has traded down a bit. Nothing surprising to me here. Now it's all about execution.

I think the 2 most important things to watch are, number 1, how do the Grok and Cursor revenues end the year?

Hmm.

And number 2, the traction they get on continuing to replace traditional mobile carriers with Starlink.

The distribution started, I think, today or yesterday. I got my first distribution from a fund I'm in.

I'm in a couple of funds that are in SpaceX. Seems like everybody's in that. And that will obviously create downward pressure if you are amongst the people who want to cash out and have been in it for a long time, but I'm holding these for my grandkids.

Sacks, your take on these spectacular—yeah, I guess is the only way to describe them—results coming from a vertical that wasn't part of SpaceX's business 9 months ago.

Yeah, look, I thought it was a very bullish earnings call. I was a little bit surprised that the stock went down after the earnings call because not only was it a beat and raise, but also I think Elon spoke to a lot of their plans.

The only thing I would add to what Brad said was around Starship. Elon basically said—we all saw it, right?—that the Starship test flight was successful, that Starship is floating in the ocean, and the heat shield worked. That's going to enable more flights of Starship now at a more accelerated rate. That paves the way for the V3 satellites, which enable much more bandwidth for the Starlink network, which then powers the whole direct-to-cell play.

You had that piece of it. I mean, just the whole telecom aspect seemed very on track, and they're very bullish about that, and then you've got the whole AI data-center play. Now, on the data centers, I think what they said is that they expected to go from 1.4 gigawatts of compute to about 2 by the end of the year, and Elon said that the spot price for compute is in the $30 to $50 per watt range.

So, you do the math. A gigawatt is 1 billion watts, so $30 to $50 per watt means $30 billion to $50 billion per gigawatt, and I think they're at the high end of that range right now. So when Elon says, "Look, we're going to end the year at $100 billion of ARR," all you have to believe is that they're at 2 gigawatts of compute running for $50 a watt to hit that.

That doesn't include Starlink or the launch business or the Grok-Cursor piece or any of these things. So I think that's why they're so optimistic.

And multiple ways to win is what you're saying, Sacks.

Yeah.

There's multiple ways to win with this stock.

6. Starlink Becomes A Cash Machine

Chamath Palihapitiya

I think Starlink's just an unbelievable juggernaut cash machine. If you look at the financials, their segment reports—Space, Connectivity, and AI—and on the connectivity side, the Starlink side, they generated $2.6 billion in adjusted EBITDA. You can kind of approximate that to be kind of operating cash flow. Space was kind of negative $200 million, so call it break-even, and AI was plus $1.1 billion.

But AI, to Brad's point, it's unclear whether the pricing they're getting on compute rental today is temporary and at a premium because of the lack of compute available in the market today, and people that need compute are paying Elon a premium for that compute.

Right.

Chamath Palihapitiya

So I think there's a question mark where that goes. But the connectivity piece on Starlink—$4.3 billion in the quarter and $2.6 billion in adjusted EBITDA. He's got 12 million subscribers. That's doubled year over year. $66 ARPU per month, what people are paying per month, and he grew 20% quarter over quarter.

So if you extrapolate this out, he’s pretty close to being at a 24 million-subscriber run rate. On this multiple, I’m assuming this enterprise stuff, which is airlines and other things, scales, which they seem to be scaling with the consumer business. Starlink alone could be generating on the order of $40 billion of top-line revenue, with a huge amount of that flowing to free cash. That could be $30 billion of free cash flow within the year.

That alone provides the cash flow to fund much of what Elon’s doing. If you just put a 30× multiple on that, which I think you can, because these subscription businesses have a very high renewal rate and very low CAC, I think you could probably get a 30× just on the Starlink business. The Starlink business alone could be a trillion-dollar market cap within 18 months, let’s say. That, I think, funds all of the rest of this as kind of science projects and upside.

So I’m kind of making a bull case. It’s crazy to me how well the Starlink business performs, and you can see it in AT&T and Verizon, HughesNet, and Viasat. These companies have been decimated. I used to have a HughesNet satellite dish on my Sonoma County ranch. In order to get internet, that’s what we had to use. It was, you know, $200 a month or something.

Terrible—

Terrible service.

Because those are high orbit, right? And they take forever to—

Terrible service.

Yeah.

And that market got decimated by Starlink. If he launches the handset thing, that subscriber growth is going to go… Right now, he’s adding 2 million subscribers on the consumer side a quarter. You could see that going to 4 to 5 million a quarter. You could actually see an acceleration in the consumer subscription.

There are 400 million mobile subscribers—

Right.

400 million mobile subscribers just in the United States, just in the United States.

I think you can make the bull case on Starlink alone—

Yeah.

—and then the rest of it is like, “Hey, is Elon going to do well with investing the excess capital that’s spinning off of Starlink? How’s Elon going to do with that money?” Well, I don’t know who else I’d give it to to do what he’s doing with Starship, AI compute, and the Terafab. I don’t know if you guys—

Oh my God, this is science fiction—

If you want to talk about—

—is going to be in Grimes County, Texas.

How the—

Yeah.

—US gets off this dependency on Taiwan and China for semiconductors—

Yeah.

—if Elon takes this on his shoulders and he delivers what he’s showing as a vision here today, this is going to be the greatest semiconductor fabrication site on planet Earth.

Well, I would say something, David, to your point. How many CEOs or founders would just take that Starlink business, which is such an exceptional business—a trillion-dollar business going to $2 trillion—and not take any of these other risks? They would not do Terafab. They would not try to build out the data center. They would not try to build their own model.

Those are highly risky but highly important investments that are being made. It is heroic and important that we have this level of—I just think—unbridled enthusiasm for innovation on the frontier that Elon’s doing. I wish we saw more CEOs, more public companies willing to take this level of risk. We just got done talking about—

Yeah.

—some CEOs maybe taking less risk because the safe bet was easier to make. Elon refuses just to take the safe bet. He’s taking all the dollars from this thing where he has an extraordinary business and plowing them back into these things that are critically important to the United States.

And by the way, Brad, such a good point, because if you look at other CEOs and other management teams, they’re getting in on this. They’re starting to realize that buying back your shares and giving dividends is not as important as betting on the future.

DoorDash got taken to the woodshed because they’re investing too much in CapEx. Obviously, Google got smacked with its CapEx spend, so that keeps happening over and over again. On the headwinds that SpaceX is going to face, the arguments that I think will turn out to be wrong, but are valid to talk about here, are: Is this demand for tokens and compute going to keep up? Or does on-prem, desktops, and open-source models getting smaller and better actually mute demand at some point?

I don’t think it does. I don’t know if there’s an upper bound for on-demand intelligence. The second one, obviously, is that Starlink is for people who are in rural neighborhoods. If you’ve got Verizon Fiber to your building or Spectrum, you’re not putting, nor can you put, a Starlink on your building.

So the piece there that’s going to be explained, probably in the next year or 2, is that every single Tesla sold is going to have Starlink in it when they get that merger done. What that means is you’re going to have Wi-Fi networks connecting any phone to any Tesla—say, all those robotaxis out there. You’ll be able to connect also directly with the next generation of Starlink.

So your phone will be able to directly—if it’s got clear line of sight—connect to any Tesla on the road, of which there are many, and all future ones will have Starlink built into them. So those are super promising.

And then finally, there’s been a lot of speculation about the valuation. Brad, you brought it up. I think at liquidity, when people were asking you, and I heard you talk about it: Hey, private companies and venture capital are a voting mechanism, and then when it goes public, it becomes a weighing mechanism.

Sometimes you’ll have this moment in time where there’s hand-wringing about those valuations, and the hand-wringing peaked in the last quarter. You had a 160× price-to-sales ratio for SpaceX when it first came out. You take their 2- or 3-trillion-dollar market cap and put it against a smaller revenue number. Well, if you look at the revenue number increasing, now we’re down to a 45× price-to-sales ratio. So some kind of balance is occurring here, Brad, between these private and public markets, as well as the increase in revenue.

Yeah. Honestly, I think this is all super healthy. I think the SpaceX IPO was extraordinary. I think the consolidation here is perfectly predictable. Now you have a company at $1.4 trillion that I think, if you take a 3- or 4-year view, you can see yourself tripling your money in this business at a very reasonable valuation on the Morgan Stanley numbers, on the Elon numbers, or whatever. But that’s always been the bet.

Do you believe that Elon is the greatest innovator and a great allocator of capital? But the price of entry matters, right? When you get carried away on day 1 of an IPO and you buy this thing over $2 trillion—

Yes.

—you’ve got to know that this is going to happen. I was on CNBC the day of the IPO, and I said I would want to own this company, but I’m not sure today’s the day I would buy the company, right? And so that—

Entry price matters. I mean—

Entry price matters.

This is just fundamental.

But let me give you another one. We’ve talked about the Anthropic IPO, or a lot of people have talked about it, later this year. I hear a lot of people saying $1.5 or $2 trillion. David just talked earlier: It’s going to be at a run rate over $100 billion, maybe by the end of the year. That’s like 10 to 15 times revenue. That is not that much for a company that just grew 10× and is rumored to be profitable in Q2.

And so I look at the market—the consolidation we saw in the month of July. We put in the Leopold bottom, hopefully, in July, when a lot of semiconductor stocks—

Leopold.

—were down 30% or 40%.

This poor guy.

Hey, listen. The guy’s doing great. He’s apparently still up 80% for the year and just made another big private investment. I think he’s done an extraordinarily good job building the firm in a short period of time. But the market did panic around that as he had to cover.

I think all of that is really good. So as I look ahead, marching to these IPOs later in the year on the back of the SpaceX IPO, I think we’re in really good shape, particularly if these revenues continue apace.

You know what Elon’s really good at is just building stuff. Physical sites. That is such a core advantage in this world where everyone’s competing for data centers and fabs.

The software layer needs hardware in the physical world in order to deliver its software services, and there is no one better than Elon at actually doing that. Look at how Gigafactories have been stood up around the world. This is his core competency.

So, Brad, when you put Elon up against Dario Amodei, Sam Altman, and even Alphabet, I mean, man, Elon’s got a core advantage if this is what this world comes down to.

Right. He said something like that on the call, where he said, “Look, putting up data centers is nothing compared to the difficulty of putting up a rocket,” right? Creating data centers is not rocket science.

So they take some of that hardware expertise that they have from SpaceX, and they put it into data centers, and that's why they've been able to stand up more data centers—or bigger data centers—faster than all the competitors. A couple of points there. Is it clear why Starship is so important to Starlink?

Okay, let me just explain this quickly. Basically, SpaceX has developed a new V3 satellite that has 10x the bandwidth of its V2 satellite. So currently, the Starlink network is powered by—

V2 satellites. They deploy them on the Falcon 9 rocket, and they launch about 27 satellites per launch, and that adds about 2.6 terabits per second of total network capacity. Starship deploys 60 of these V3 satellites per launch. That would add 60 terabits per second of total network capacity per launch, so over 20 times more capacity per launch. That's the power of it.

So if they get Starship working… By the way, the last test not only proved that the heat shield worked, my understanding is they actually launched—or rather, they deployed—20 V3 satellites as a test, and they were able to make a connection with those satellites and prove that it worked. So I—

They even had cameras on them that tracked.

The reason we were able—

Right.

—to see the Starship was because they were like, “YOLO, let’s put some HD cameras on them.”

Right. Now, I think those satellites were basically just a test.

They burned up.

They burned up. So I think the next big milestone here will be when they launch Starship with, let’s say, 60 of these V3 satellites, put them in the correct orbit, make a connection with them, and add the bandwidth to the network. That’s going to be a big milestone.

But you play this out to its logical conclusion, and the bandwidth available to the Starlink network goes up 10x, or eventually 100x. That’s when they can do all the interesting things, like direct-to-cell. There were some interesting hints that Gwynne Shotwell talked about with ground stations and what they could potentially do there. And I think, I think if—

At the end, they might buy T-Mobile or something like that, Sacks. It’s easily within their range of purchases.

And I think Elon mentioned something about potentially the Starlink network eventually handling roughly half of internet traffic. So, I mean, this thing could get so much bigger than just 12 million subscribers, to your point, Friedberg.

7. The Data Center Financing Test

But look, I want to actually talk about the data centers for a second, Brad. I do have a couple of questions about this. Elon mentioned that we’re going to be at 2 gigawatts by the end of the year. He said that we will be at 5 to 10 next year, closer to 10 than 5. So let’s just say 8, okay? So—

Yeah.

I’m just making that up—

Yeah.

—but it’s in their range. So let’s just say that’s an addition of 6 gigawatts, so they go from 2 to 8. To me, there are 2 questions there. One is, how do you know that the spot price is going to stay where it is? Can it stay at $50 per watt? How do we know? How do we track that? How much risk is there around that?

I got the sense on the call that Elon thinks that number is going up because the market is memory-constrained right now. I think he mentioned that we might see a 20% increase in memory production next year, but the demand is going up 200% plus. So the market is constrained by whatever the bottleneck is at that time. Right now, the bottleneck is memory. So where do you see the spot price going? How do we know? How much risk is there around that?

And then the other question I would have is, if you go from 2 to 8 gigawatts, you have a net addition of 6. We know that a 1-gigawatt power data center is $50 billion of CapEx.

Right.

So 6 incremental gigawatts of compute would be $300 billion of CapEx next year, assuming they build that, right? I mean, they have optionality around that, I’m sure. So how do you finance that? What’s the most non-dilutive way? They said their payback is a year or less.

Wow.

I’m sure that’s tied to the spot price.

Yeah.

So you only have to finance it for a year, and the question is, do you think NVIDIA gives them that financing, or how will this play out, I guess, is my question?

It’s a great framing, David. First, it’s $50 billion per gigawatt to build, minimum, okay? So you’re at $300 billion. In order to finance that, it seems to me you either have to go into the market and borrow the money, or you have to do a dilutive equity raise, neither of which they want to do, or you get NVIDIA to backstop it, which they’ve indicated they’re going to do more of.

But the problem there is NVIDIA shareholders don’t want them backstopping an unlimited amount, because the fear in the world is that that spot price, at some point, may go against you. And when it does, the payback period changes.

Now, nobody thinks that the payback period is going to be 1 year, even though the spot price is suggesting that it is today. A few months ago, people thought you would get payback over 4 years. So you basically spend 50, then you earn 10 to 15 per year. You get payback over 4 to 5 years, and then hopefully you get the 6th year, which really takes you well above 20% in terms of your returns.

Right now, the shortage is so acute, and the willingness to pay from the frontier labs is so high because they all recognize they’re on the verge of some massive breakthroughs, that they’re willing to pay 3x, 4x, 5x market pricing in order to get at-scale compute. And that’s what happened with the Anthropic deal with SpaceX. I think Anthropic would buy a lot more of that today if they could. Same with OpenAI.

You’re saying, Brad, they would be willing to overpay by a factor of up to 5x?

Well, that’s the $50 per watt that David was referencing. They would be willing to pay $30 to $50 if they could get at-scale compute that would give them a competitive advantage over the other people in the market.

And remember, there aren’t a lot of people who have the offtake revenue that can afford to buy compute at this scale, right? It wasn’t the Chinese open-source companies that were buying SpaceX’s excess compute or building the 10-gigawatt plant in Ohio. That’s OpenAI and Anthropic.

So the vast majority of the offtake commitments are coming from Anthropic, OpenAI, and NVIDIA. When you hear about the hyperscalers building all of this compute out, they’re building it out to sell to the people that we just mentioned.

So, David, net-net, if he builds 6 gigawatts next year—and, by the way, probably only Elon can actually stand up that much in that timeframe. Jensen said to me on the pod, he’s like, “Nobody comes close. Microsoft doesn’t come close. Google doesn’t come close in terms of standing it up in that timeframe.”

I think he’s going to have a challenge getting all of the componentry. I know he can stand it up, but can he get the memory? Can he get the chips? Can he get the land, power, and shell all in time? I think the offtake is there.

But to put it in perspective, this year, Anthropic and OpenAI combined, their starting total compute was 5 gigawatts. So he’s talking about incrementally adding more than they had as combined companies to start the year.

Chamath Palihapitiya

But that’s not that much of an increase when Anthropic is growing 10x year over year and OpenAI is maybe at 4x, or maybe higher now. So you add maybe—

The demand exists in the world today. I think it will exist in the world for the next 12 to 24 months, but there is a wall of worry in the market.

The reason we saw the pullback in July is Kimi scared people into thinking, “Oh my gosh, they’re going to undercut the frontier labs’ revenues.”

Chamath Palihapitiya

Right.

And if they undercut the frontier labs’ revenues, who the hell is going to pay for all this compute? That’s why you saw a 40% trade-down in the CoreWeaves of the world and all of the semiconductor stocks and semiconductor-related AI stocks.

Well, and in some ways, the fact that there’s a discussion going on that this next 10 gigawatts is going to cost, Sacks, $500 billion, and you ask the question, where does that come from? A secondary offering? Does NVIDIA put it on their books? Do they create SPVs off their books, like some people are doing?

The fact that we’re having this conversation, everybody’s aware of it, and the market has been educated on it means I think people will be able to change in real time if it doesn’t come to pass or if it slows down. I suspect this cannot keep up at this pace for more than another 2 years or so. I think we’re—

Although, as our good friend Bill Gurley likes to remind us, he’s like, “I can’t believe that we’re all just taking in stride this level of seller financing.”

Yeah.

He would call it circular revenues, right? But the market has gotten comfortable with this. And remember, as we saw in July, if there is a scare about demand, the whole sector trades down.

Yes.

Everything will trade down together, because that’s just the leverage that you’re pumping into the system. You’re effectively backstopping people’s ability to build ahead of their revenue.

So it becomes much more violent if you ever see demand slippage. Famous last words, I don't see it today over the course of the next 12 to 18 months. But you have these unknown unknown moments that certainly cause people to be fearful.

Yeah.

Credit spreads have continued to stay wide on these deals, so there is fear in the market about them.

All right, everybody, the fifth annual. If it's September, you know it's time for the All-In Summit. The fifth annual is happening. Yes, that's right. David Friedberg's been at work, and we have an all-star, all-star list of people joining us. Jensen Huang, founder and CEO of Nvidia. If you care about where AI is headed, you won't want to miss this conversation.

The best. The oracle.

Satya Nadella, CEO of Microsoft, fan of the pod, will be coming on for the second time. Jared Isaacman from NASA. The one, the only, Brad Gerstner and Bill Gurley, BG too, coming back. SpaceX's Gwynne Shotwell. My guy, Jake Paul. Nick Shirley. Lot of incredible people coming. Martin Shkreli maybe is even coming. He's... That's gonna be fun. Go to theallinsummit.com to apply today. Allin.com or theallinsummit.com. Any of those will get you there. And we're taking over Universal Studios again. We'll have our own private playground. Dave Friedberg, great job on the summit. Casino night too. I heard it's gonna be a big casino night.

Biggest yet, and the concert to be announced who will be performing at the concert, but it is gonna be incredible. So I'll just say, one of the things about the summit, we've had people come to the summit from over 60 countries. It's really incredible to meet all these people, entrepreneurs, investors, people that are just really interested in the topics that we talk about. We try and have the world's most important conversations, but it's really this amazing community experience. That's what brings folks back. So we try and invest more and more every year in making it an amazing experience, not just cool content on a stage, which I think is what a lot of these other shows really deliver, but it's like, how do you actually come and have an experience for a couple days? It's gonna be awesome. So we're excited.

It really is those three things that we focus on. One, you're gonna learn something, right? You got these great people on stage. You're gonna learn something from them. You're gonna meet new people. You're gonna network, and then you're gonna have these great experiences. It's the trifecta. Folks, you excited, Brad? You excited to be back?

What are the dates? What are the dates again?

Look at your calendar. You're speaking.

September 13th through 15th in LA.

This couldn't be better dates for the summit. I mean, we're gonna be within 60 days of an election, midterm election. We're gonna be within 30 days of an IPO, you know, potentially of Anthropic. I mean, it's gonna be heated.

8. Airtable Sells For A Fraction

The SaaSpocalypse—not the Sacks-pocalypse. This is the SaaSpocalypse, I guess, winding its way out. The indigestion might be clearing. Airtable just got acquired for less than it raised. It's a profitable SaaS company with a great product, $480 million in annual revenue, growing 20% a year—respectable if it were a public company—with almost $1 billion in cash, and it has been sold.

It's been sold for $1.28 billion, about 10% of its peak valuation, which was $11.7 billion in 2021. Now, it did have a bunch of cash, so if you include the cash position, the sale was $2.25 billion. It was acquired by a firm called Bending Spoons. This is an Italian, Milan-based company. They buy challenged but interesting businesses: AOL's legacy business, Evernote, Eventbrite, Vimeo, and Meetup.com. Bending Spoons went public last month, and its shares jumped 15% on the Airtable news.

Sacks, when we look at this, this was a company that had done a lot of things right and had a massive amount of cash in its war chest. But rumors were maybe the founders were a little exhausted, maybe some of the investors who bought in at a high level were exhausted. What can we take away from this transaction and Bending Spoons? Are they the buyer of last resort now?

Well, I think they're creating a great business for themselves because I think this will end up being a fairly profitable acquisition for them. Let me just add a piece to this: Airtable spun out its AI agent business, which is known as Hyperagent, into a separate, independent company prior to this acquisition. So I think what's going on here is that the founders and talent of the company said, "Look, we don't want to have to make this legacy product work. That's basically a private equity play." I'll explain what that means in a second.

"We want to focus on the new thing, the AI company. That's where the big value creation's going to be in the future, or the potential for it." So essentially, the talent is going to focus on the venture play, and then they're selling the private equity play to Bending Spoons. Now, why do I think this could be a good acquisition for Bending Spoons? I think there was a really interesting data point that I saw in the commentary on this, which is only 30% of Airtable's sales team was making quota.

They had a 30% sales attainment number, and that told me a lot about this business. What it told me is—and I'm reading between the lines here—that this was a company that had a successful PLG motion, in other words, organic growth, product-led growth, and they were growing about 20% a year. But that was not good enough for its board. These are investors, some of whom invested at an $11 billion peak valuation. So they're looking for a venture-type outcome.

So what happens? The board pressures the founders to do something that, frankly, is unnatural for them, which is they say, "Look, you should bolt on a traditional sales-led motion here to get the growth up faster." Does that work? No. They probably get a little bit of growth out of it, but they only get 30% attainment. So they've got hundreds and hundreds of sales reps here trying to push on a string, and it's not making it grow faster.

So now what's the opportunity for the acquirer here? Bending Spoons can go in here and do what Elon did at Twitter: eliminate 85% or 90% of the cost structure. Don't do the sales-led motion. Just go back to your product-led growth roots. You'll probably keep most of that 20% growth, and it'll be a very profitable company. You'll be able to—

80% profitable, probably, right?

Probably. People are saying they're only gonna generate a 30% EBITDA margin. I think, like you're saying, it could be 80% or 90%. I don't think you need to keep most of this business—

Hmm.

—or most of the cost structure associated with this business. Airtable is a company that has its fans. I think they will probably stick with it, and, I don't know, you could probably generate $300 million of EBITDA a year, or $400 million, while growing 10% to 20%.

Yeah.

So that's the play for Bending Spoons.

And the venture investors here, Sacks, they're—

Yeah.

—happy to get their money back and move on to the next thing. It's a bit of a push for them, in terms of being at the blackjack table—

Yeah.

—rather than having to go 10X just to catch up, and then they would have to go 10X again to make their LPs happy. It's not gonna happen.

I think the question is: If Bending Spoons can basically take this business that's not making money, probably generate $400 million a year of EBITDA, and pay for the acquisition in just 3 years—

Amazing.

—why isn't that something the company could do on its own? I think the structural problem is that it's very hard for both VCs who are on the board and the founders to shift into private equity mode. Why? Because they're gonna have to demolish what they've built, right? They've got all this loyalty to the team. They don't wanna think about, "How do I eliminate 80% or 90% of the cost structure?" It's just not what they do.

Of course.

What founders want to do, and the outcome that the board members are going for, is a venture-backed outcome. And I think they could have done this. They could do what Bending Spoons does, but—

They're not built for it, Sacks—

They're not built for it.

—is what you're saying. They're not built for it.

And moreover, the structure of the cap table is all wrong because they're sitting behind this giant liquidation preference. All these investors have to get paid back, including those who invested at this $11 billion valuation and all the way up as well.

The incentives are broken, Brad.

Right.

And you yourself, at your firm Altimeter, you were pretty frisky in this period. You made a lot of bets. I don't know if Airtable was one of them, but you made some SaaS bets there. Some of them were at high valuations. How are you looking back at that time period? Any lessons that you take going forward?

Multiples of revenue can compress very quickly, right? It works great when a company's growing greater than 50%, but remember, it's just a heuristic. It's just a very rough estimate used almost exclusively in Silicon Valley.

People are saying, “Oh my God, this thing sold for 2 times revenue.” But when you actually look at it on a look-through basis, it probably sold for maybe 30 times free cash flow. I don't think it's easy to get it to $400 million in EBITDA. I think if it was, the board would've done that.

We're involved in some of these companies. Once they slow down, the company morale goes to hell. Churn among your customers begins to spike. It starts to feed on itself. So I think it'd—

It sucks to go to work every day—that's basically the picture.

But Brad, what do you need to keep? What do you need to keep? Because look, think about—

It becomes very, very tricky. I don't know the core product and what's happening in terms of turnover in the core product, David, but my hunch is that the core product has started to really—

Fizzle as the advances in the core product have slowed down. You're seeing a bunch of churn out of it on the product side, and now people are saying, “Listen, it's almost impossible for a software company today to keep any decent salespeople, to keep any decent product development people, because they all want to go work on AI.”

Agreed—

So a lot of—

But you don't need them for this product.

I agree. So—

I mean, the market's being efficient. This is where I think Bending Spoons has an advantage that the company's board and founders wouldn't have, which is that they already have an infrastructure, right? They have a core team at Bending Spoons that's managing now, I don't know, dozens of these properties, and so they can plug this in.

I think AI, in a way, makes their job easier, because in the past, the reason why you couldn't eliminate all of the talent and the infrastructure was because you needed the institutional memory. You needed people who knew the code base. Now AI can learn the code base instantly, right?

That's an interesting insight. Right.

And so, yeah. So the—

Maintaining is easier with AI.

I think maintenance mode becomes way easier with AI because you don't need the historical knowledge anymore. The AI can go in and—

And feature updates.

—and sort of reconstitute that historical knowledge.

Let me get you in here, Friedberg, if I may. When we look at the lessons from peak ZIRP and SaaS, and then we look at this moment in time, this surging AI market, are there any parallels that we might find here, or lessons between the two?

Between the ZIRP and AI eras?

The ZIRP SaaS era. We had a lot of very high valuations, a lot of enthusiasm, a lot of suspension of disbelief. We're here in the AI era. We've just talked about the price of compute and all these companies being at a 100× price-earnings ratio. Any parallels here or not? It's kind of a softball question for you.

No, I mean, I think this is—

—earnings ratio. Any parallels here or not? It's kind of a softball question for you.

No, this is a very different paradigm. The AI CapEx build-out and model training, which is where the predominance of the capital is flowing, is not about some high multiple on revenue, which is where capital was flowing into SaaS. It's like, “Oh, you get a 20× multiple, turn a dollar into $20. That's great. Let's do it all day long.”

This is a very different structure and strategy and capital allocation process. So I don't think that I would look at them as being linked in any way.

Yeah. It was a softball question, to be honest. I was letting you hit it out of the park.

Look, obviously, SaaS companies were overvalued during the ZIRP era for 2 reasons. One is that we had artificially low interest rates, so we had a speculative asset super bubble. But the other is that people were treating these things like guaranteed annuities—and actually growing annuities. They'd look at it and see, “Oh, 120% net dollar retention, so this thing will just grow 20% year over year forever as a base case,” right? And they were then priced that way.

But what we've seen with AI is obviously there's disruption, and to Brad's point, I'm sure they're seeing elevated churn right now. It's not an annuity. Things can change. So obviously now these things are trading at a much greater discount.

All of that being said, let me just say I don't think you can extrapolate to the entire SaaS space based on this 1 company, Airtable. I think there are some things about Airtable that make it very different from, I don't know, let's say a Salesforce or a Workday. Airtable was always a little bit of a quirky product.

I remember at the peak of the hype for this company, people were saying, “Oh, this is like a new Excel or a new Google Sheets.” It's basically a—

New Microsoft Office, yeah.

Yeah. It was basically a spreadsheet for words. That's how people were viewing it—as this new kind of spreadsheet for words as opposed to numbers—and it never achieved that kind of promise. It never achieved that kind of ubiquity. People understand how to use spreadsheets. Everyone uses them. Airtable never got to that point. Most people still don't know what Airtable is.

Again, it had its dedicated fans, but it was a hard product to explain to people. When do you use it?

It had a cult following, is what you're—

It had a cult following—

Yeah.

—but it never achieved that sort of level of acceptance. It was never self-explanatory in terms of why you should use it or what the use cases are. They were never able to get the marketing right because of that—

But—

—because of that.

And to be honest, if you look at Claude Cowork, Perplexity Computer, agents, those things are now doing what Airtable did.

It never carved out, I think, a niche where it was super clear when you were always supposed to use Airtable. Really, it was part of this hodgepodge, this grab bag, you could say, of no-code tools. No-code has to be the most impacted, the most disrupted area of SaaS right now because, I mean, what is Claude Code really good at? That's the ultimate no-code tool.

It's no-code. Lovable, Claude Code, Perplexity Computer—

Yeah. I mean, the thing with—

—all of these harnesses. Yeah.

Yeah, the thing with Airtable or Retool, things like this, is it's true you didn't need to be a coder to use them, but you had to learn how to use Airtable. You had to learn—

Scripting.

—how to use Retool, all these things. They were kind of alternative programming languages, in a way, and you just don't need to learn any of that anymore. I mean, you use Claude, and you just tell it what you want it to create.

And so, if you do want to create a new dashboard, some sort of verbal spreadsheet or whatever, you just tell Claude what you want. You don't have this learning curve. Look, all of SaaS is being impacted right now, but this has got to be the most impacted area. So I don't know that you can totally extrapolate based on what's happening to Airtable.

I don't necessarily think that you want to replace your CRM, your ERP, or your HR system with something that's been vibe-coded. You want the certainty for anything that involves compliance—

I gotta be honest. Do you use an off-the-shelf SaaS tool for managing cap tables, portfolios, and everything?

Well, we vibe-coded something, actually.

So, yeah, we just did the same thing.

Yeah.

My team just built something that is so mind-blowing that to buy it with off-the-shelf software would've been $250,000 in software and like $1 million in integration over 2 or 3 years, and we built it in a month. And now we have complete insight into the whole portfolio, the competitive set, the founders, everything going on.

Keep in mind that 1 of the reasons why Leopold got blown out, okay—I mean, it is because he bet on the SaaSpocalypse. Remember, it wasn't just that he was super long these chip stocks that had a correction.

Oh, is that right? He was short SaaS?

He was short Adobe and a whole bunch of other SaaS companies, and those trades also moved the wrong way on him. So again, I just think that it's painting with too broad a brush to say that all of SaaS is going to get obliterated here.

Yeah.

And there was a really good post about this. Let me just quote from it: “Nobody buys Microsoft because Microsoft writes the best code. They buy Microsoft because Microsoft is the rail that everything else runs on. Active Directory is where your employee identities live. Excel is where your board deck numbers come from. Teams is where the compliance-recorded conversation happens.

Azure holds a FedRAMP High authorization and Department of Defense Impact Level 5 clearance, which means a defense contractor cannot casually swap it out for something cheaper,” and so on down the line.

So there are a lot of really good compliance reasons why, if you're a large enterprise, you're not going to want to spend tens of millions of dollars ripping out something that costs you $1 million a year. It just doesn't make sense. And I noticed that Benioff just tweeted 5 minutes ago that 15 out of 15 cabinet agencies run on Salesforce.

Look, the government is not going to rip and replace Salesforce with something vibe-coded.

Not overnight, yeah. And not all SaaS is equal in this dimension is my point. I just bought some Figma. I think some of these SaaS companies with great founders who are in it for the long term, and who have passionate user bases, will make the jump to AI-first products. I’d put Figma in that bucket.

Just to wrap this section—

Yeah, please.

IGV is up 20% in the last 6 months. It’s up 20% in the last 5 years, right?

Explain IGV, please.

The high-growth software stock index, right? Snowflake is up 88% in the last 6 months, right? So—

IGV is an ETF of those—

IGV is an ETF of growth software companies, right? So, to David’s point, there was a panic about software companies. There was a big trade-out. Honestly, they performed pretty well. As he mentioned, in the month of July, they were up when a lot of the semiconductor AI stocks were down. Some of these companies—Databricks, Snowflake, ClickHouse, et cetera—are doing extraordinarily well.

As I just mentioned, Snowflake is up 90% in the last 6 months, which puts it in the same category as the semiconductor AI stocks. So, to David’s point, you can’t throw them all in the same bucket. But I do think that for these no-code application software companies, they’re realizing the game is up: sell the company and get what you can get.

Importantly, in the Airtable story, all the late-stage investors—we passed on this in the last 3 funding rounds, which I think were at $2 billion, $5 billion, and $11 billion—but all those late-stage investors, which were the most venerable of growth firms, got their money back, and the early-stage investors ended up making a lot. So, if this is a failure, this is a pretty good failure for Silicon Valley.

This is one of the points that was made at that time, which is, “Hey, this is a strong enough company and team and revenue base that if we just get our money back with the optionality, hey, maybe this would be a good investment.” You could say the same thing about some AI bets right now.

Well, this is one of those cases where the liquidation preference actually mattered.

Normally it doesn’t matter, but it did matter in this case.

Absolutely.

I think they got straight money here. My understanding is this wasn’t like they had a 7% interest rate, or they didn’t have a participating preferred where you get 2 times your money back, and then they do the trade. Does anybody know? I looked deeply into this, and I couldn’t find any—

David’s just saying, I think that net of cash, they may have come in a little bit less than the total cash raised. But it seemed like everybody got made whole.

Yeah. We lived through moments in time where companies had to guarantee a 1x in a sale.

Well, no. A 1x liquidation preference is standard. It just means you get—

Right—

—your money back—

Cash back.

—before other people start to profit, which is appropriate, before the common profits.

But also, the interest rates were taken out of these deals, right? I think during peak ZIRP.

Well, no. The standard terms, what’s known as clean terms, are just a simple 1x liquidation preference. The preferred just gets their money back before the common starts to participate in a successful sale of the company. That just makes sense, right?

Yeah.

Participating preferred is the double dip, right?

But participating preferred is the double dip, right?

Yeah. And look, we’ve never done that. We believe in clean terms. No one’s trying to be punitive toward founders. It just doesn’t make sense for some people on the cap table to be making money while other people are losing money.

Yeah.

It just doesn’t make sense, right?

It screws up the alignment, yeah.

Well, that’s just a transfer of value from some people on the cap table to other people on the cap table. So, the standard thing you do is make sure that the investors get paid back, and then everybody participates in the upside.

Okay. Fourth story here: China is training on US data from US providers. Forbes published an investigation called “These American Startups Are Making China’s AI Smarter,” and I think this relates to a lot of your work in the early part of the administration, Sacks.

They claim US data-labeling startups are selling valuable training data to Chinese labs, which, in turn, is helping them catch up with the US frontier labs. Two startups, Surge AI and Mercor, are both valued at over $20 billion. They sell training datasets to people like OpenAI, Anthropic, and federal agencies. They all sell the same datasets to top Chinese AI companies, according to this report, including Tencent, ByteDance, Alibaba, Moonshot, and others.

The top 6 AI labs in China, according to this report, are spending $500 million a year buying what Forbes calls “secret sauce”: PhD-written content, reinforcement learning, knowledge pipelines, and all that kind of great stuff. I have investments in a couple of these companies, including MicroOne. The founder of Micro1 didn’t participate in selling to China. He made that decision.

Sacks, what do you think about this new wrinkle? Really, the secret sauce behind a lot of these models is the data. We’ve run out of open data on the web, obviously. We talked last week about the books having their spines taken off and being scanned in. People are looking for data. Mercor, Micro1, and all these companies are providing it. Should they be providing the same data and selling it to Chinese open-source companies or not?

Well, I think we’ve got to decide what our objective is here. Are we trying to just get into a full-blown economic war with China? Are we just trying to prevent all of our companies from doing business over there? If that’s our objective, then you can take that position.

Historically, the rules have been that you want to be careful about the transfer of technology that has a dual use—technology that has a military application. My sense of data is that it’s largely a commodity. Data labeling certainly is. If you basically tell them that they can’t use data labeling, I guarantee you there’s no shortage of labor in China that they can use to do the data labeling. In fact, they probably are.

What I’m saying is there are a lot of ways to get this data. If we basically ban these companies from selling to China, we should expect reciprocal actions taken by China to ban companies over there from selling to us, maybe rare earths. These 2 countries are not completely independent of each other.

By the way, I want us to be as independent and sovereign as possible. I don’t want to have any dependencies.

No dependencies.

But we still, at this moment in time, do have some dependencies. So, I think you have to ask the question: Is this data really proprietary?

It is.

Does it have a dual use?

Yeah. It’s not data labeling.

Does it have a military application?

Yeah.

I don’t think it has military application. It’s definitely not data labeling. This is like hiring PhDs, hiring superprofessionals to create unique datasets. So, it’s science.

China can do that, too, and I guarantee you they are. I don’t think this is going to give us a decisive advantage in the AI race. It’s going to create annoyance and friction, and how bad do you want our relationship with them to be? Do you want to risk starting another trade war?

Look, I’m not against restrictions when I think they’re going to pack a punch. For example, I’m really glad that the first Trump administration limited the export of EUV lithography machines to China. That was all the way back, I think, in 2019. That was a really important decision.

So, I think targeted strategic controls make sense. I would just make sure that this one actually meets that bar.

First, I’m in absolute agreement with David that we want maximum competition. As we sit here today, the US is winning. We talked about it at the start: our frontier labs are winning, our open source is winning, and we have fairly limited regulations.

She’s coming here in September in a bilateral meeting to meet with the president. We’re advancing relations on a variety of fronts. So, I think everything looks good, and you want to continue down that path.

With that said, I will tell you that this will irritate people in Washington who feel that this, along with distillation and other things and the export of chips—all of which, at a certain level, make sense—causes people to wonder whether or not we’re making it too easy for the Chinese labs to catch up with American labs in the race to frontier intelligence.

So it's the type of story, Jason, that I think will continue to muddy the waters, that will continue to be monitored. The reason I don't think it will cause us to change our stance with respect to China is because we're winning.

But if the president asks his advisors one of these days, 6 months down the line, “Are we winning against China?” and all of a sudden he gets a response, “No, we're no longer winning. They've caught up. They've passed us,” et cetera, then these things will get a lot more scrutiny than they're getting today. I think the only reason they pass muster today is because we're still leading the race.

I gotta say, using Kimi and Qwen and GLM-4.5 for the last 60 days, my lord, these things are good, and I don't think it's very patriotic to be giving them an advantage. I wouldn't do it.

Sorry, what's the advantage? What's the data set that you're worried about that's so proprietary?

Any of these data sets are created by experts here in America who are given the queries that have errors in them. So when you give a thumbs down to a query that's highly technical, it could be code, it could be biology and science.

Chamath Palihapitiya

And they can.

No, no. If they were to do it at this scale, they would need to hire the best and brightest scientists and experts in the West. So basically, all the knowledge of the West is being put into packages for our LLMs to get better.

They're sending those same packages and reselling them to Chinese companies, which means they catch up just as quick. I think it's a big part of why they're catching up. In line with distillation, you know? It's a very similar process.

Look, if there's something truly proprietary here, I don't want us to sell our secret sauce to China, so I'd have to look into that and see: Is there some real secret sauce here? But this idea that it would seriously disadvantage China—they're graduating more math and science graduates every year than the rest of the world combined. They don't have a shortage of smart people.

Yeah, and we're graduating them and kicking them out of the country. That's the other problem. We gotta get that fixed.

Well, it's a lot of different issues here. I don't know how many of them you want to conflate.

All related. Yeah.

But this idea that they can't recreate those data sets—look, if there's something truly proprietary here, if it has a dual use, if it's military-related, I don't know that that's what this is.

Well, they're all proprietary by design, but I don't know about the dual use because I don't have the data sets here.