史上最大LBO、SPAC 2.0、开源AI模型与各州AI监管狂潮

David Sacks is in some deep negotiations for the United States of America, from the SCIF. There’s no T at the end. It’s just T.

No T.

Oh, “Skip to My Lou.” Okay, he’s in a SCIF doing something with his BlackBerry and a bunch of generals. Nobody knows what’s going on in Sacks’s life, but he’ll crack in from the SCIF any moment now.

Did you guys see that Pete Hegseth announced a PT and fitness test for the generals? Could you imagine if Sacks had to pass a PT test?

Oh my God. They should totally make it a requirement for the administration. “Sacks, we need you to do 1 push-up.”

Sacks, what do they do if you don’t pass? They remove you from your position?

You probably get a cure period.

We should do a push-up contest. That would be great. Winner takes all. How many push-ups can you do, Friedberg?

You have to adjust for people’s heights. I’m the tallest of all of you. I have a much longer limb system.

What does that mean?

So 20 for me is much harder than 20 for you, Jason.

I mean, 20 is easy for me at this point.

Yeah. You’re like Bilbo Baggins. It’ll take you like 8 seconds to put a tank, man.

Bilbo Baggins. How about Thor? I’m like Thor at this point. I’m going into my Daniel Craig era.

Cal has an edge relative to my Daniel Craig weight-to-weight ratio. That’s highly advantaged.

All right, let’s get started. Anyway, what is your arm length, JCal? Do you have a good arm length?

My wingspan technically is enough to kick your ass with one hand tied behind my back. That’s actually—let your winners ride.

Rainman David, we open-sourced it to the fans, and they’ve just gone crazy with it.

EA is being taken private in the largest take-private deal in history: $55 billion. That stacks up to, let’s see, TXU, the power company, in 2007, and HCA Healthcare at $33 billion. This is a large deal.

Investors in the take-private include Saudi Arabia’s PIF, Silver Lake, and friend of the pod Jared Kushner’s Affinity Partners. It’s $210 a share, a 25% premium to the stock. Kushner’s largest LP at Affinity, as you know, is PIF as well.

The PIF has invested over $900 billion in many things, as you know: Lucid Motors, LIV Golf, the SoftBank Vision Fund, Uber back in the day, Newcastle in the Premier League, and Electronic Arts, obviously, in the video game business. EA was founded at Sequoia’s office in 1982 in San Mateo. Shout-out to our guy Roelof Botha, who joined us for the All-In Summit.

Their headquarters are still in Redwood City. Madden NFL, The Sims—that’s why you have the background of The Sims this week—Need for Speed. Pretty insane deal here. Chamath, this is a high watermark for private equity, any way you look at it. The PIF loves games. They are the biggest shareholder in Nintendo, Savvy Games, and Scopely. I mean, they just keep buying games. What are your thoughts here on this deal happening right now?

I really like it. Let me give you the bull case, and then let me give you what the bear case would have to believe.

The thing to remember is that video games are the anchor pillar of usage across the entire internet. Last week at our poker game, we had Matt Bromberg join us just for dinner, who’s the CEO of Unity, and Alex Blum, who’s the COO of Unity. One of the stats that they shared with us at dinner was that 3 billion people play games, which is just an incredible, incredible stat.

So, in many ways, it’s much bigger than social networking and social media, or at least as big. EA is this 800-pound gorilla, but I think the problem is that they’ve always been these gatekeepers, and I think there’s a risk and a chance that these gatekeepers get eroded away. Specifically, who I’m talking about are folks like Microsoft and Xbox.

At the point that this company is going private, there are some really interesting things happening. Xbox, I think the day after the EA deal got announced, decided to hike prices 50% for their subscription service. What happened over the subsequent few days is that so many people tried to cancel that the site went down.

So, what are you seeing happening? You have distribution gatekeepers trying to raise prices and take share. Then you have the original IP owners, who have not had a well-funded way of fighting back in a category that is basically as important as, and frankly more important than, social media.

I think if you take an asset like this private, it allows you to take your time to clean up the OPEX model, figure out who does what, use the best of all these next-generation tools, and then find ways of distributing outside the scope of Xbox and PlayStation so that you can take more of your share. If you do those things, this is a multihundred-billion-dollar asset, and I think it could be just an enormous win.

So, I think it’s very smart. What’s the bear case? I think the bear case is extending a theme that I’ve talked about here a few times, which is that I think the value of patents, and by extension IP and copyrights, is going to go away. In that, there’s going to be a spectrum where certain content IP holders lose and other ones win.

I think gaming is on the winning side, to be honest, and I think content studios in general—the Disneys, the Hulus, the Netflixes—are on the losing side. But the bear case would be that these toolchains allow the number of games being built to increase by 2, 3, or 4 orders of magnitude, and that they are distributed by other places like the social media sites.

I just think that’s a pretty low probability. So, on balance, I think Jared and Egon did a killer deal. I really like it.

For people who don’t know, Unity makes the 3D software that people build games in. It’s a public company worth $16 billion, also backed by Roelof and Sequoia back in the day. Incredible company.

Friedberg, what are your thoughts on the gaming industry versus, say, social media versus traditional media? We’re seeing massive amounts of money being put into each of these, but this is about time, and for this next generation—let’s say millennials and younger—we’re seeing a big mix.

Obviously, they don’t have cable TV, so that’s been plummeting, but they do play games. They do like YouTube, TikTok, et cetera, and they do love social media. What’s the future here, as you see it?

One way to answer that question is to think about how people spend their time. Do you spend more minutes on social media, on traditional media, or playing games, and how is that trending? More importantly, which of those will accrue more benefit and, as a result, drive more hours spent from AI?

Is AI going to create more social media engagement? Is AI going to create more traditional media engagement, or is AI going to create more video game engagement? I think one way to think about this thesis is that AI is ultimately going to accrue to video game entertainment far more than social media entertainment or traditional media.

Why is that? Explain it to me.

Because I think you can create dynamic, more engaging experiences that will benefit from a back-and-forth sort of relationship more than you can with traditional content or with social media.

What we see now in a lot of gaming systems that didn’t exist 12 years ago is AI-driven players embedded in the games that act and feel a lot more like real human engagement. That is very hard to mimic from the traditional programming methods that were used in gaming, and so that makes a big difference.

For example, if you’re playing Fortnite—I don’t know if you guys play Fortnite or have played Fortnite—but if you’re a noob in Fortnite, an early player in Fortnite, you’re mostly playing against AI. Even though you go online and play against what are supposed to be other players, you’re mostly playing against AI because they tune the AI to be easier to beat so that you can slowly develop your skills.

What was happening early on was that they were seeing a high degree of churn in Fortnite because kids would go on and play for the first time, get paired up with kids who were better than them, never win, get frustrated, and quit the game. The churn rate was high.

So AI unlocked higher engagement and higher retention on the Fortnite platform. I think we’re seeing that in a lot of different gaming platforms now. AI can be used, for example, to maximally increase time, engagement, satisfaction, and happiness.

I think the Saudis saw this, and if they’re trying to diversify away from their oil holdings, entertainment and how people spend their free time—which, by the way, I think is a general macro bet that everyone should consider making—become increasingly important.

If you believe in AI and you believe in the improvements in productivity, people in the industrialized world will generally have more free time on their hands and be able to support themselves with the deflationary effects of AI over time. If there’s more time on people’s hands, the general market for entertainment is growing. If the general market for entertainment is growing, gaming is the future of entertainment. The future of gaming is AI.

The Saudis own 10% of this company prior to the deal. I don’t know if you guys have tracked the investments they’ve made, but they’ve been extremely aggressive with gaming. They have this investment division called Savvy Games, and within Savvy Games, they bought Scopely for $4.9 billion in 2023.

And then earlier this year, they spent $3.5 billion to buy Niantic, the company that makes Pokémon Go. They also own 4% of Nintendo, 6% of Take-Two, and a sizable percentage of Activision Blizzard. So they've put quite a bit of capital into small investments in other gaming platforms. They own a few gaming platforms.

So this is clearly a big thesis and a big investment that they see as the future of entertainment over time. Jared's firm, Affinity, is going to own about 5% of the company post-transaction. The Saudis are going to be the majority owners. I think this is going to end up being the next big platform play for them, and it allows them to make the important long-term investment in furthering the transition to AI without having to worry about quarter-to-quarter earnings, but really making a 10-year bet. They do talk a lot about this 2030 vision.

And if you look across those 3 categories we've been discussing here, about 60% of U.S. adults use video games every week. About 75% of Americans use social media every week, and streaming—traditional media, the Netflixes and Disney Pluses of the world—is still 83%. So these are the 3 buckets of people's time. Books and going to the movies are obviously the big losers.

I think the market was totally getting this wrong because the tick-tock of the deal is super interesting. When they were looking for the debt financing, it was about $36 billion of equity and $20 billion of debt. They called Jamie Dimon, and Jamie basically underwrote the $20 billion in on the same day, just because I think he could underwrite this pretty fast.

Some of the biggest deals are frankly so obvious that it just takes the courage to put them together, and then everybody's like, “Oh, this just makes so much sense.” Andrew Wilson, who's the CEO, is going to stay on. He's a great guy. Super compelling.

It's worth talking a little bit about the impact of private equity. If you spend any time in the region, I'm going to be in Saudi and Dubai in the first week of November doing my Founder University, and I've been out there twice a year, maybe, for the last 3 years. They'll tell you, whether you're in Doha, Abu Dhabi, or Riyadh, “We've got 6 or 7 industries we really care about.” Technology is at the top of the list. Private equity is at the top of the list. Live entertainment and sports are at the top of the list, and hospitality is also at the top of the list. Real estate—building new places for people to go—is at the top of the list.

If you look at private equity, pull up that chart I had there. This is just stunning, how big this industry is getting. You know, $5 trillion is what we're up to here, and it just keeps growing.

I think private equity is totally screwed. I don't think Silver Lake or Affinity or this deal are screwed, but I think private equity in general is totally over-owned.

All right. It's gotten huge just since 2015, tripling in size. So why is this, I guess, my question for the gentlemen here and for the audience: Why is private equity becoming so large, and what impact does that have on society?

If people can't put EA into their retirement account, they can't put Stripe into their retirement account. If we take all the great companies and start to privatize them—SpaceX, let's say, never goes public—what impact does that have on people's retirement accounts?

Okay, look, I think the history of this is important. There was a long-standing belief that the best way to generate the best risk-adjusted return—what does that mean? That means managing through periods where the stock markets go down and managing through periods of volatility—the best way to do that was to have what's called a 60/40 allocation: 60% to bonds and 40% to equities.

Over many years, especially when we artificially suppressed rates at 0 during the Obama administration, a lot of people started to move their allocations away from 60/40, and they started to make more and more investments further out on the risk curve. The biggest beneficiaries of that were venture capital, private equity, and hedge funds.

The thing with private equity is that, because rates were 0, they had an infinite amount of borrowing capacity and very little downside. They were able to manufacture returns much faster than venture capital and hedge funds could. As a result, you had an initial group of people that were defining the asset class and making a ton of money, and then you had all these fast followers that said, “Well, if they're doing it, I can do it, too. So far, so good.”

But then, as always happens, you have this flood of laggards that just flood the zone. It's these laggards that make it very difficult to generate returns because they start overpaying for assets. They start mismanaging and under-managing the assets that they do own.

So where we are is that private equity was seen as a very consistent way of returning money to help improve that 60/40 portfolio. As a result, they got a lot of money, but that created a lot of competition. That's why you see this hockey-stick graph, Jason. When you see that kind of graph—

It doesn't matter what asset class it is: The returns go to 0.

And so we've seen this in venture capital—

We've seen this in hedge funds—

And we're now going to see this in private equity.

Too much money going in. To be clear, what you're saying means you kind of exit it, right? There's no returns. And so, again, I've said in any of these alternative asset classes, if you had to have 1 critical question, what are your distributions? Don't show me your IRR. What is your DPI?

The distributions on your paid-in capital. If the answer is 0, then it is a very challenged asset class. What I will tell you in private equity is that over the last 4 or 5 years, distributions have been few and far between.

So I think what's going to happen is that the money is going to come out of private equity and get concentrated into the few companies that know what they're doing, of which Silver Lake has generated, over the last 15 or 20 years, tens and tens of billions of dollars of distributions. They are just an exceptionally well-run organization. They've done these huge buyout deals successfully before. So we need to go through that in PE.

Where does the money go? The money's already leaked into private credit, which is the next big bubble that's building. It looks like this chart that you just showed, which is loaning businesses money.

It's super interesting because you make such a good point. What we're seeing in private equity is these continuation funds. Now continuation funds are coming, Chamath, to venture. I've been getting pitched on these continuation funds where, like, “Hey, take all your assets, sell them to a new group of people, and then reset the clock,” and then there's never an exit.

The good news, I will say, is that over the last year we've seen a lot more activity for shares of our companies that are still private. So the secondary market, Friedberg, is coming back in a major way.

But I do get worried about these continuation funds because now you're just moving an asset from one class to the other, and we need to have a functioning IPO market. How functional is the IPO market today? Would we say it's completely dysfunctional?

How dysfunctional is the IPO market? Let me say it another way: How do we correct that? And this leads into your new SPAC.

Look, there are 3 ways to go public. There's the traditional way, the IPO; there's the direct listing; and then there's the reverse merger, or the SPAC.

Up until I floated IPOA in 2018, I think the first way was really the only way. I was involved in 2 direct listings, Slack and Coinbase. In both of those, what I learned is that it has the same vagaries as the traditional IPO.

In the traditional IPO, you go to a bank, they underwrite you, they act as a gatekeeper, and they take 6%, 7%, or 8% fees as a result. Then they allocate what is essentially underpriced stock to their best customers. You see a 1-day pop, maybe a 2- or 3-day pop. All of those customers tend to unload, and then the stock tends to drift down. So the IPO is expensive, and it typically is mispriced.

With the direct listing, you have a different dynamic, which is that the first trade is always the highest trade, and then it just goes straight down. That happened with Slack, and it happened with Coinbase.

Spotify would be in that group as well.

Yeah. With Slack, I remember I was offside a billion dollars, and I was like, “Well, I'm never letting this happen again.” When I had the Coinbase thing, I sold it the first day. And I texted Brian. I said, “This is not a directional indication of your company. It's the dynamics of the direct listing, because I learned it the hard way: The time to sell is on day 1.”

So where does the SPAC come in, especially now in version 2? Version 2 is the thing that I have been tinkering and refining with and am trying to push in this new version. I think that it's creating an incredibly competitive vehicle where you can have a ton of money go into these private companies and take them public at a very, very low cost of capital. I think that should be very enticing.

So, you closed your financing. Can you just tell us what the capital raise was like as you went out and met with folks? What did you hear?

Yes. Nick, maybe you can find it. You know that image of the Raptor engines?

Yes, super complex to being elegantly simple.

Yeah, Nick, can you maybe just throw that up? What I would say is, like, SPAC 1.0, of which I was right in the front of the parade, had a bunch of misfires and it was complicated, but it worked.

There were some hot fires that worked, but then there were some clear misfires. The whole point was to prove that you could create a competitive alternative to the IPO. The thing that I'm the most proud of, quite honestly, is that, for all intents and purposes, I started a normalization of this vehicle that's now raised more than $150–$200 billion for American companies. I am very proud of that. That's an important thing for the American capital markets.

I think what we did in American Exceptionalism is Raptor 2. It's not yet perfect, but I do think it tries to improve on the things that I noticed were not working in Raptor 1. A lot of that is the compensation and incentives. When I showed that to investors, they were quite excited. I think that they want a competitive IPO market that brings many, many American businesses to the public market so that they can be owned by everybody. They like the transparency, and the fact that the incentives are such now that there's absolutely no compensation unless this thing really works.

Historically, they received warrants in the company, typically with a strike price of $11.50, so 15% above the issue price of the stock.

There were founder shares, but did you have a reaction from them saying, “Hey, we want some warrants. We need a little extra kicker here”? Was there some sort of desire for that?

No, in fact, it was the opposite. I think that the institutional investors—and my investors in this, 98.7% of the capital was allocated to these guys—are the best of the best. You know who they are. They're every single blue-chip, A-plus institutional investor.

What they wanted was great companies. They want great companies to be public. And the reason is the thing that Friedberg, I think you mentioned this before: when a good company gets public, the amount of money that they can raise in the public markets, and then the amount of growth that they have in the public markets, far outclasses what they'll ever do as a private company. And so they want the simplest and cheapest way for great businesses to get out.

Chamath, do you think that, when you find a merger partner, the traditional SPAC will still be announced as a merger concurrent with a PIPE being done, where new investors are underwriting the valuation of the deal and saying, “We like this company at this price,” because we're now going to put money in in the form of a PIPE?

Historically, the PIPE was for common shares, so it was like, “This is a good price,” and everyone felt good about it. Number 1, do you anticipate that there will still be a PIPE done concurrent with the merger in this transaction? Number 2, do you think it'll look like a common PIPE? Because after the SPAC frenzy died down, in order to get deals done, the PIPE started to get done with convertible preferred securities. So they were senior to common, and they almost were like debt. How do you think this is going to play out?

A clean deal has not happened in quite some time where a SPAC has announced a merger and simply raised money via common in the form of a PIPE.

It's a great question. I think it comes down to the underlying asset. But there are some incredible companies that are private that, if they go public, will be able to demand common PIPE capital.

I think the future—maybe just prognosticating and guessing—what does Raptor 3 look like in this SPAC? I think Raptor 3 will look like where somebody, a sponsor like me, rolls everything up into one thing so that it's already pre-wired from the beginning, where I'll just speak to $1 billion, $2 billion, $3 billion, whatever it is, of flexible capital that can come in as common, so that it's a totally pre-baked IPO at a very fair price. I think that's what the Raptor 3 version of a SPAC will look like.

So more capital, and then they put their full trust and faith in the sponsor to run the deal.

Well, then—meaning, then there's no conversion risk. All the money comes over, right?

You set your compensation to be a bit Elon-like in terms of your compensation as the sponsor. It comes, if I read it correctly, Chamath, when it hits certain milestones in terms of share price.

Yeah. Nothing can be earned unless the stock is up 50%.

And then there's a tranche at 50%. When the stock is up 75%, there's another tranche, and when the stock is up 100%, there are no founder warrants in the deal—or are there?

Nothing.

I think this is great.

The reason why this is important is that all of those things that you guys mentioned increase the cost of capital to the founder, to the private company board, and to the employees. All that's unnecessary dilution. So now we take it all off the table.

Yeah, smart. The observation I had at the time, not just for your collection of SPACs in the 1.0 era, but for all of them in general—and I tried to explain this to our syndicate members and investors, as well as the CEOs, because a lot of my CEOs were asking, “Should we do a SPAC?”—was that one of them, Desktop Metal, did this. It felt like venture investing.

If you look at Opendoor, Virgin Galactic, Joby, which I don't think was one of yours, SoFi, and MP Materials, you have to look at it as a venture-type investment: 80% of venture goes to zero, and 20% pays up for the other 80%. I think people were looking at this like it was Netflix, and they were not thinking of these companies and the stages they were at.

Can I just ask a question? SoFi and MP Materials did extraordinary. In this class of companies you're going to be taking out, is it going to be the same early stage, or are you thinking more robust, more predictable revenue—let's call it resilient revenue, maybe rugged revenue?

I think it's the latter, but I think it's also important to note that this time around I've tried to really minimize retail exposure to this. I don't think that retail is well-suited right now to have these things.

My honest advice is: avoid maybe not all SPACs, but definitely my SPAC. Just avoid it. I think that there is more than enough liquidity on the institutional side for us to do an interesting deal, but it fits in our portfolio and our construction, which is a very different risk model. And so I would hate that people are out on the risk curve without really understanding the risks because, Jason, you can't predict the market. You don't know where these things are going to go.

Yeah. I mean, Desktop Metal, a 3D-printing, cutting-edge nanotechnology company, should have stayed private a couple more years. People investing in it need to understand that you're now acting like a venture capitalist, which means the return profile and how the portfolio management works is distinctly different from doing Netflix and NVIDIA and whatever other publicly traded companies.

I would just say, do not invest in these things. Don't. At least, you know, just—

I think you just inspire people to do it. I know that's not your intent, but when you say, “Don't do it”—stupid.

I'm being very honest. Don't do it.

No, no, I know. Don't buy SPACs unless it's less than 1% of your portfolio. That would be my advice.

Before we move on, can I just make one comment? I'd like your guys' take on the private-equity stuff, because Chamath made a comment that private equity is baked. One of the things to take note of in this take-private of EA—and we talked about it—is the theme of AI empowering EA to transform the business. Jared's brother Josh has, at Thrive, been executing a roll-up of CPA accounting firms and applying AI to reinvent that business.

Oh, is he really?

Yeah.

Oh, I should talk to him, because we have an investment in a company called TaxGPT.com that's basically like copilots with AI for accountants. It's doing spectacular.

What he's done is he's bought these traditional accounting firms at some multiple of EBITDA, and then he can transform the business with AI and really create a new opportunity. I've said that I think this is one of those few moments in history where there really is an opportunity to beat the market and make money in the public markets if you can be thoughtful and selective about the companies that stand to benefit from an AI execution strategy.

In all of these traditional markets where you have competition, everything's commoditized and the market is mature. It's very hard for any of these players to differentiate product, service, and, obviously, unit economics. But with AI, it's completely transformative and has transformative potential in nearly every industry.

So as a public-market investor, if you can identify those opportunities and select them where the management team has the right leadership in place to execute against this, you could make real money. The problem is most of these companies are not led by folks that understand AI or are software-first.

And so I think there's an opportunity for more buyouts. They're not going to be of the $55 billion scale. It's worse than that.

In what sense?

We at 8090 have done the dance with all the big major private-equity firms. Here's how it goes: it always goes the same way. The partners love it because they're looking at minimal distributions, companies that are good but not great in many cases, and they want to see improvements to EBITDA and performance so that they can either sell them or move them out.

All of them. Yeah, all of them—with their existing portfolio companies. So the GPs are like, “This is genius. We should do it.” Then they’re like, “Here’s a handful of companies to go talk to.” And I’ll be really honest with you: what you find in most private-equity portfolios are B and C companies run by C and D folks.

Yes.

So the ability for them to go and embrace this is basically next to none. If I look at my customer distribution and concentration at 8090, we’re at a run rate into 9 figures already, working on a $300–$400 million deal. About a single dollar comes from a private-equity firm, although we spent a lot of time initially trying to sell our software factory and trying to sell work into them.

It’s really hard, and it’s what you said before, Friedberg, which is that the people incentives at these businesses are misaligned to the AI outcome, right? You can’t fire these people, and I don’t think the right answer is to fire them. So I don’t know what the right answer is. This is why I think private equity is very challenging.

Do you think there’s a power-law situation where perhaps a handful of investors in the public markets and perhaps a handful of investors in the private markets can identify and then put the right people in place and execute against these strategies, like Josh is trying to do with his—

I think Josh is smart, so I think Josh will figure it out no matter what. What I’m saying is, if I can show you 20 or 30 customers, a ton of revenue, and all these white papers that show upside, and I still can’t get it done inside one of these companies, I think it’s not us—it’s them, right?

It’s not inherent in traditional private equity to do this either, which maybe begs the question: Is there a new kind of private equity that can execute this? Maybe that’s an opportunity, like Josh is showing. He’s a venture investor who’s executing a private-equity strategy, and maybe that becomes the play.

I think if this works well, 2 of our biggest customers are individual decabillionaires who own businesses, and they’re like, “You’re doing this?”

So, to the extent that Josh looks more like that—an owner of 100% of the business, where it’s like, “You’re going to do it”—then I think it can work. I think the Saudis—I think the owner-operated model is the only way the AI transformation really works.

At the other end of the spectrum, it’s the public-market CEO who realizes that they have to do something real because they’ll otherwise lose their job or they’ll be disrupted. Those are the 2 cohorts that I feel today are on their front foot. Everybody else is sticking their head in the sand.

Just on the AI front, I forgot to ask you: Sir Demis, my Greek brother, didn’t he show a—

Didn’t he show the Genie 3 engine that would make infinite games?

Yeah. So it’s not actually a 3D engine. It’s a class of AI models that can render what the experience looks like and feels like—a 3D world—but it doesn’t have an underlying traditional object-rendering engine. It doesn’t have a traditional 3D physics engine. So it’s a new way of experiencing these kinds of world-interaction systems.

There are several startups. I think Fei-Fei Li is her name, the Stanford AI one.

And she has one of these. That’s a virtual-worlds company that has the same principle.

I asked Bram and Alex about exactly this at dinner.

What was their take?

Yeah. He said it’s just really, really hard to get these things to actually be legitimate engines at the scale of what Unity offers, for the quality of game that needs to be made for it to work. The interim step is going to be that the assets in it are created by AI. That’s what I’ve seen a lot of startups doing. So you want to make a character, and you’re dropping characters in, and they would be done in real time.

I think the whole thing is Unreal and Unity as the rendering engines, and the AI sits on top. The AI can render objects, concepts, and structure, and it can render the direction that you, as an engineer, would typically provide to the Unity or Unreal 3D engine. That’s going to unlock not just video games but also film.

You’re 100% right. Can I tell you an example? Yesterday, in our group chat, a bunch of people sent around Sora, the slop app. I downloaded it just to play with Sora yesterday, and the first video that came up was exactly this. It was an ATP tennis match where it was a guy’s face—the guy, like, imagine you—and he was playing against Federer.

Yeah.

Then I thought, what if he was playing against his friend and that was the actual video game? To your point, you get away from all this IP-licensing, gatekeeping stuff, and you can just get to good games faster and good content faster.

I think they’re adaptive in terms of the competition, so you’re not playing somebody who’s just going to dominate you. It just gets 5% better every time you play it. You’ll get 4% better, and it’ll make it perfectly challenging so you don’t quit and you’ll learn as you go. It’s really going to be interesting.

The same will exist in content. You’ll make shorts and films, and then the ones that have the most engagement—the AI prompting system will get better and better.

You’ll see that happening with Star Wars or Marvel. If all of a sudden Silver Surfer is an interesting character to you, or Ahsoka is interesting to you, it’ll sort of make that world or enhance that character and tell you more of their backstory. That can be very interesting as a—

How can you sit in your seat and make fun of me, call me a nerd, and you actually know the name of this Star Wars character? I don’t even know who you are.

Very important character. Ahsoka is Anakin Skywalker’s Padawan. She is a very important character. If you watched The Clone Wars, you would know this—the animated series that threads through the—

Actually—oh, look who dropped in. Oh, David Sacks is here. Did you get out of your—were you in a skiff or something? What’s going on, Sacks?

I was in some meetings, but actually, no, I was just buying some domain names.

Oh, you were? Did you get Mahalo.com?

I got Mahalo for the bargain price of $1 million. That’s what it’s worth. Go to mahalo.com. I’m selling it for $1 million. I mean, it’s in the dictionary.

Yeah, I have some old assets. Somebody else should use them. I just have Begin.com, and I’m going to be working on that in partnership, probably with one of the large AI companies.

I might give you an equity swap for that. I’ll give you a—

Mahalo is the second most important name, the second most important word after aloha in the Hawaiian language. I’m surprised Benioff hasn’t tried to ask you for it.

I was just texting with Benioff.

Give it to him as a gift, dude. He’s a great guy. Just give it to him.

I will give him Mahalo.com if he gives me 4 weeks in one of his Hawaii resorts per year.

He would do that.

Oh, for the next 20 years? Oh my God. Imagine Jason for 80 weeks. Oh my God—as a houseguest for 80 weeks.

It doesn’t matter. I’ll give him the money so he buys it. Don’t worry—donate it to his nonprofit foundation, and then you can take a tax write-off.

Look at everybody. When I have something to sell—the guy with the lowest net worth on the program, when I’m trying to pay off my jet—you guys all have criticism. How come I can’t wet my beak?

Let me ask you a serious question. So you had investors in Mahalo, right?

Yes.

And I assume—

This is their domain. This is their domain. It will go to them.

Oh, so it will. Oh, okay.

It will go to those investors.

You’re paying off the liquidation preference, correct?

Okay. Just sitting there.

So now, instead of losing 100%, I’ll lose 99%.

Something like that.

It’s just startups are hard, folks.

But I have the Begin.com, and I’ve been talking to folks. Mahalo was originally a human-powered search engine like Wikipedia, which we’re about to get to, and my concept was to do comprehensive search like Naver.com or Daum in Korea—I had seen those services. And it turned out to be exactly like Perplexity, but at the time we tested machine learning, which is what everybody called AI back then, and it just didn’t work. So we were trying to hand-roll search results and then back them up with computer-generated ones, algorithmically generated ones, but the tech wasn’t there. Now, I want to do something again with Begin.com. I’m really excited about that domain name.

All right, listen. We brought up slop. Let’s get into it. 2 slop apps in a fortnight here—no pun intended. Zuck and Sam Altman have both released slop apps.

Sammy the Bull. What a deep pull—Sammy the Bull Gravano.

There it is. And here’s a look at Sora. It’s objectively extremely impressive. Here’s Sam Altman. People don’t know this: Early in his career, when he was starting OpenAI, he didn’t have the money from Elon. And here’s Sam Altman stealing an H100. Here’s Sam Altman—also, this is when he was storming the Capitol on January 6. Here he is when he was working at Google. Yeah, lots of— but it’s really good, and they are basically taking a ton of risk and solving some problems with IP.

As we all know, the IP output is where people think you're going to have to be really thoughtful or get a bunch of lawsuits. On this app, you can opt in and make your persona, like Sam did, available for everybody to use. So that whole concept of notable persons allowing their image to be used—you opt into that—and that's pretty clever. You can make it so your friends can basically make videos of you, but nobody else can. It's a thoughtful way of doing it.

However, very controversially, this thing had everybody's IP in it, and you have to opt out if you don't want your IP used. That's going to get him another whole collection of lawsuits to go with the New York Times and Ziff Davis ones. And there have obviously been a bunch of settlements now, including Anthropic settling their book thing for $1.5 billion. So, anybody play with these tools yet? What do you think, folks? What's the point of these? Do we think this is like a TikTok competitor? Do you think it's just a back door to training data? What do you think?

The closest thing is a TikTok competitor, but I use it. I thought it was okay. But again, the thing that I keep in mind whenever I try these apps for the first time is: today is the worst it'll ever be. It only gets better from here.

And so, if you look at the starting point, it won't take but a year—or maybe 2 years—for this thing, I think, to be legitimately excellent. It has to get the scripting right. It has to get the prompting right. It has to be a little bit easier for you to use. There were a bunch of prompts that I used that were rejected by Sora or by the IP, right?

Well, it just said, "Use me," but I couldn't validate that I was me. And so you have to take a picture of yourself. It's a little clunky, the app, right now, but you're right: it's going to get better in each version. The one by Zuckerberg is called Vibes.

I was looking at these, Sacks, and I don't know that this is intended to be the next great social media app as much as it's a data play to get folks to train data. When you see them, what are your thoughts on them other than "interesting"?

I haven't played with it yet.

Oh, sorry for me to say. Friedberg, you got any thoughts on it?

No, I don't have thoughts. I think we're kind of early innings. I do think there's new categories of media that none of us are really considering today. Traditional media, as I've mentioned in the past, is centrally produced and then broadly consumed. I think there are models of media that are going to emerge that are going to create new business categories or new business models, and also new media categories that are all about distributed production and not necessarily central production, distributed consumption. So that changes things quite a bit, and I think maybe this is going to start to open that door a bit.

One of the things—I thought about this, and I mentioned this in the past—where I'm like, everyone's going to make their own movie, their own video game, their own music. But there is this notion of shared cultural context. Everyone wants to talk about how the 49ers did this weekend, or whether you guys saw that show Adolescence. We want to have a conversation about some shared stories. That's the basis of societal interaction and memetics.

So I think there are elements of this being the beginning of the enabling tools, but I don't think we've actually seen what's going to happen, which is: how do you take one story and then create a distributed way of consuming that story where everyone experiences and consumes it differently?

I do think this notion is like, hey, everyone's making fun of Sam. Maybe there's some cultural context about Sam Altman that we all share, and then we're all engaging with Sam Altman in different ways. So I think we're very early, and we don't yet know how it's all going to play out, but I think that's really critical.

Something is lost because we used to all talk about the latest Tarantino movie or the latest Sopranos episode. We don't do it anymore. I do share stuff. We do talk about tweets and stuff, and there are other forms of groups, but it's not like it used to be, where 30–40 million people would see Raiders of the Lost Ark and it would be the discussion of the summer or whatever it is.

So I literally bought 20 tickets to the new Paul Thomas Anderson film, One Battle After Another, just so I could have a conversation with 20 friends about the new PTA. People really are longing for this shared experience.

Paul Thomas Anderson—he did The Master. That will be one of the greatest ever. He is a top-5 director of all time, but I know you don't care about culture. Is he like Michael Bay?

No, the opposite of that, actually. Michael Bay makes things that go boom. Paul Thomas Anderson makes things that make you go—

Michael Bay is super cool, fun to hang out with, fun to party with.

Right. Okay, well, way to bring it back to you. I don't know Paul Thomas Anderson, but it was a heck of a film. Sacks is actually very cultured when it comes to cinema. Did you see it yet, Sacks?

I have not seen it yet.

It's of the moment.

I heard it was anti-conservative, so it does have some left-wing take.

No, it kind of mocks the left and the right. It's kind of mocking both extremes. You'd love it. I think you would very much appreciate it.

All right, I'll check it out.

Yeah, I would check it out.

Hey, I have an idea. Why don't we find a topic that's interesting to talk about?

Yeah, okay, great. Well, if you contributed to the docket or showed up on time, maybe we could do that. So, unbelievable. Just so you know the inner workings right now, there's a little resentment in the group because one of us decides to change the time of the pod for 4 weeks in a row and then show up half an hour late. I won't say which person that is, Sacks.

Sacks, here's an interesting topic from Red Meat for you. DeepSeek, the Chinese LLM, just dropped its latest model, DeepSeek-V3.2-Exp. It's faster, it's cheaper, and it has a new feature called DSA, DeepSeek Sparse Attention, which makes it faster to do training and inference on larger tasks. The key takeaway is it can reduce API cost by up to 50%. The new model charges $0.28 per million inputs and $0.42 per million outputs.

Claude, which is a leading model from Anthropic that a lot of developers use and a lot of startups use, is like $3/$15, so 10 to 35 times more expensive. Obviously, people are cutting their prices pretty quickly. But, Sacks, this is your wheelhouse as our czar of crypto and AI for the United States of America. What are your thoughts here on the continued execution of the Chinese government with DeepSeek?

Well, I want you to hear Friedberg's thoughts on this because he was paying attention to this, weren't you?

Yeah. I think there's a total rearchitecture underway, and we're at the earlier stages of cost per token in terms of dollars and energy. My understanding is there's actually a lot of work going on with U.S. labs right now on a similar kind of track that's going to result in similar results. Maybe they're a little bit ahead of the curve, but we should really pay attention to the curve.

What do the models say in terms of energy demand and cost per token if these architectural changes really do drive down costs by 10×, 100×, 1,000×, or 10,000× over the coming months?

And this is open source, so just so everybody understands, it's available on AWS and GCP. At least 3.1 is; I don't know if 3.2 is available there now. But I'm hearing from a lot of startups—I don't know if you're hearing this in the field, Chamath—that they're testing it and playing with it, in some cases using it because it's so much cheaper. Are you seeing that?

We are a top-20 consumer of Bedrock. Let me tell you what it looks like on the ground. We redirected a ton of our workloads to Kimi K2 on Groq because it was way more performant and, frankly, a ton cheaper than OpenAI and Anthropic.

The problem is that when we use our coding tools, they route through Anthropic, which is fine because Anthropic is excellent, but it's really expensive. The difficulty that you have is that when you have all this leapfrogging, it's not easy to all of a sudden just decide to pass all of these prompts to different LLMs because they need to be fine-tuned and engineered to work in one system.

The things that we do to perfect code generation or to perfect backpropagation on Kimi or on Anthropic, you can't just hot-swap to DeepSeek. All of a sudden, it comes out and it's that much cheaper. It takes some weeks. It takes some months. So it's a complicated dance, and we're always struggling as a consumer: What do we do? Do we just make the change and go through the pain? Do we wait on the assumption that these other models will catch up?

So, are people making tools now that make it easier to switch between them?

No. And, by the way, I can't just make it easier to switch between them. This weekend, a different company with a huge model came to us and gave us a preview of their next-generation model. It's incredible, but then when I sit down on Monday morning with my team and I'm like, "Okay, what do we do?" we don't know what to do.

Do we cut it? Do we move over and say, "Great, we'll refactor all these workloads to run on this new model"? It's a really hard problem, and it's getting worse the more complicated tasks that we undertake.

Just for people who don't know, Kimi is made by Moonshot AI. That's another Chinese startup in the space. Sacks, your thoughts.

I think this is actually a really interesting topic, this topic of open source. I'm a big fan of open-source software because it's a check on the power of big tech in a way. What we've seen in the past, in the history of technology, is that these major categories end up getting dominated by 1 or 2 big tech companies, and they have all the power and control. Open source provides an alternate path, right? Because the community of open-source developers just puts things out there, and then you can take it and run it on your own hardware, and you're not dependent. It's a path to software freedom, if you will.

So far, so good. I think the thing that is now tricky about this is that all the leading open-source models are from China these days. China has made a really big push on open source. Obviously, DeepSeek is an open-source Chinese model. That was the first big one. Kimi is 1. Qwen is from Alibaba.

I think that if you want the US to win the AI race, then we're all kind of of 2 minds about this. On the 1 hand, it's good that there are open-source alternatives to the closed-source proprietary models. On the other hand, they're all coming from China.

There were some American efforts that have been important. Meta, most notably, has invested billions and billions of dollars in Llama. But the release of Llama 4, I think, was considered disappointing by a lot of people. Now there are statements by Meta that they might be backing away from open source and just going proprietary. OpenAI released an open-source model, but it's nowhere near their frontier.

There are some startups that are trying. There's 1 called Reflection AI that looks promising and is developing an open-source American model. So far, this is maybe the 1 area in AI where the US is behind China: open-source models. I'd say every other part of the stack—closed models, chip design, chip manufacturing, semiconductor manufacturing equipment, even data centers—we're ahead. But this 1 area of open source is a little bit concerning.

Interestingly, Sacks, the 2 things of note are OpenAI. OpenAI was originally supposed to do open source, so that's kind of hilarious. But the second is that Apple, which is the furthest behind of everybody, has a really interesting open-source model. So when you're behind, like Apple is or the Chinese were, you're open—you do open source—and when you're ahead, like OpenAI became with ChatGPT, you close it down.

OpenELM—efficient language models from Apple. Keep an eye on that 1.

Can I tell you what's going to make this open-source, closed-source battle even worse? Effectively, what this is is the US versus China. The US is closed and China is open, at least at the scaled models that work.

But that doesn't have to be the case, right? We could release open models too.

No, no, you're right. I'm just saying that today, if you look at the conditions on the field, the closed-source, highly performant models are American. The open-source, highly performant models are Chinese.

You would say, okay, well, what is the next downstream thing? It's what Friedberg mentioned, which is the energy and the cost of generating these output tokens. I talked to somebody yesterday who runs a huge energy business, and I have to tell you, it's not in a good place.

You saw, I think, this week, where the residents of Indianapolis were able to get their city to reject a billion-dollar data center that Google was going to build near Indianapolis, largely because of concerns about price inflation around electricity. What this energy CEO told me is, look, the next 5 years are baked. If we don't find some compelling solves—and I'll tell you what the 2 ideas were—but if we don't find some compelling solves, electricity rates will double in the next 5 years.

If you think about how consumers will view the use of AI, and then if you think about companies like us and others trying to use the cheapest version so that we are minimally impacting the downstream cost of these things, it will become an energy problem. This is a very complicated thing.

Now, his idea comes with a huge PR crisis, because if you want to take big tech, which is already viewed negatively, and make their perception even worse, you start to finger-point to them and say, “These guys are the reason my electricity costs have doubled in the last 5 years.” That's no bueno for them, and they need to find an offramp ASAP.

It's a bad look. Doubling this could take your jobs, right?

Yeah, it's terrible. Whether you believe that's true or not, that is the perception.

There are 2 offramps that he suggested, which I think are worth considering.

Offramp number 1 is what's called a cross-subsidy, which is essentially to say that they pay a rate card—which they can absorb with all their free cash flow—materially higher than what other ratepayers would pay in that geographic area. So the homeowner, his or her electricity costs, stay flat to down. The data center costs are higher, and it's the Metas, the Googles, the Apples, and the Amazons, which have hundreds of billions of dollars in free cash, that absorb it.

That was idea number 1. Idea number 2 is to start to set up some mechanism so that they can install things like batteries at every single home in and around these data centers, to allow those homes to have a better chance of actually absorbing some of this inflation without having to pay it.

That's a really good idea, and this is playing out, Sacks, in Virginia in a major way, because that's where Data Center Alley is. Forty percent of the energy in Virginia is now going to data centers. This is becoming acute. What are your thoughts here, Sacks?

Chris Wright spoke to this pretty well at the All-In Summit in terms of what we have to do. There's no question that AI is going to create a huge need for power over the next 5 or 10 years.

I think on a 5- to 10-year time frame, the answer is probably nuclear, or at least that's a big part of it. But nuclear takes at least 5 years. Within the next 5 years, it's probably gas—natural gas. The issue there is that there's a huge backlog for gas turbines, basically the engines that burn the gas to create power, and there's a 2- to 3-year backlog to spin those up.

The question is, what do you do in the next few years? I think Chris Wright talked about this, and I've heard this from other energy executives: We just need to squeeze more out of the grid. If we were to shed just 40 hours a year of peak demand to backup generators, diesels, things like that, you could get an extra 80 gigawatts out of the grid. This is what 1 energy executive told me.

The reason is because they build the grid and regulate it based on the peak, which is basically the coldest day in winter or the hottest day in summer. The same way that you build your church for Easter Sunday and the rest of the year it runs at 50%, the same thing happens with the grid. If they could just reduce the peak for 40 hours, if they could shed that load to backup generators, diesel, things like that, then they could run the grid to squeeze an extra 80 gigawatts out of it.

I think that's the bridge over the next few years. Then we need to get a lot more gas and eventually some nuclear as well. But unless you want to keep talking about electricity, I think there's some other things to talk about on open source, because I think it's a pretty interesting topic. Can we just go back there?

I was just trying to paint the case that my economic model for going to open source is better, because I can't pay $3 per million output tokens and then also pay for all this.

Actually, I want to ask you: When you're running something like Kimi or something like that, I think it would be good just to explain to the audience how this works, because I think there's a lot of confusion about what it means to be an open-source model. A lot of people think that when a Chinese company publishes 1 of these models, it's still somehow theirs. Can you just explain this?

No. But the reality is, once it's published, it's no longer theirs. It belongs to anyone who wants to take that code, and you're not running that on a Chinese cloud or something like that. The data is not going back to China. You're taking that model and running it on your own infrastructure.

When I first started 8090 [?], my only solution was Amazon Bedrock, which is a service that Amazon provides that allows you to essentially get inference as a service. As we were building our product and needed inference and inference tokens, Bedrock basically handled everything. It's what AWS is, but for this vertical of AI, right?

They have the servers. These are in American data centers. They're managed by Americans, and what they do is take a handful of models and make sure that they can support usage of those models. That was how we started.

But as with everything, we have to manage our costs and our operating profile. We're always looking for other models and other places, other than Amazon, that can service our needs. In fairness, Amazon is very expensive.

A different company that I helped get off the ground, Groq with a Q, has a cloud. What they've been doing is working initially with Llama. Then they worked with OpenAI to bring their open-source model, but they also brought a couple of these Chinese models online.

And what they do, exactly as you said, Sacks, is they take the source code, basically implement it, and fork it. Now it's implemented domestically, on American soil, by Americans, inside an American data center. China gave us the road map, if you will—the architectural plans—but we, as the American company in this case, Groq, built the house and launched it. So we, as 8090 [?], basically made a cost decision to move to this open-source model because it was materially cheaper.

Right. And what Groq, with a Q, will give you, if you're the application company 8090 [?]—

They're like Amazon for us. They're—

They'll give you an API.

Exactly.

So, the same way that if you want to use a closed model like OpenAI or ChatGPT, they'll give you an API: you submit prompts, they give you answers—basically, tokens in, tokens out. What Groq does is take this open-source model, run it on its own infrastructure, and then give you the API so that you can get tokens in and tokens out through their API.

For me, as a consumer, it reduces us to a pure economic decision: where is it cheaper? It's not dissimilar to the last generation of the internet. You'd run on AWS, but then you'd bid it against GCP. You'd bring in Azure and say, "Who's cheaper?" Ultimately, you're running a database or some service—pick your service, Snowflake.

Right.

It didn't really matter where it was. You were just trying to find the cheapest vendor.

Right. Now, here's what's compelling about it. First of all, like you said, it's cheaper to run it on your own infrastructure if you know what you're doing. Also, enterprises like it because it's more customizable, and there's going to be a lot of fine-tuning of these open-source models for specific applications.

100%. Enterprises frequently want to run these models on-premises, in their own data centers, because they want to keep their own data on their own infrastructure. But now the challenge is that you've got these models that are no longer Chinese. They've been forked by an American company, but they originated in China.

That's right. And they could be running on some critical infrastructure, and that does raise issues. What is Groq doing to test whether these models are safe and whether they could be backdoored? How do they think about that?

They have an entire pipeline of things that they do. I don't know the details exactly, because I've never asked what they run through, but—

Yeah, that's the big rub in this.

They go through an incredibly rigorous—

They basically do safety testing to make sure.

Absolutely. A lot of people think that if you run a Chinese model, the data must be going back to China, but that's not true if it's being run on your own infrastructure. I think the issue is more theoretical: could a Chinese model somehow be backdoored with an exploit or vulnerability?

Well, if you take a compiled version, sure. But if you just take the open source and do it yourself, no.

Right. That's the thing. If someone did discover a vulnerability, it would get widely shared in the community very quickly. At this point, you can expect that every major company involved in security, every cloud vendor, and every major model maker is trying to prove that the other models are inferior or bad in some way. That's where the competitive cycle is really valuable, because you have the best and brightest computer scientists hammering everything to figure out whether there's a vulnerability.

Yesterday, I was talking to a certain person—he's Italian—who's a leading security guy at one of these model makers. He's in charge of the security work, and they're hammering everything to try to figure out whether there's a vulnerability, because it slows the other companies down. That made me feel quite positive that we haven't seen anything yet on any of these models, which is to say that, generally, everybody has actually been a pretty good actor so far.

Yeah. Well, brave new world.

The other piece to this puzzle, Sacks, is that there are a lot of crypto-distributed projects. The one I've been working on is Bittensor and TAO. I think you've also done a deep dive on this, Chamath, and I'm a partner in an emerging crypto fund called Stillmark Capital. We're buying TAO and looking at Bittensor and all of these subnets that are being created for distributed computing.

This is a big push for Apple as well. You've seen a lot of M4 Mac minis out there. The plan is to put all of these LLM stacks on people's personal computers, distribute them, and have something like SETI@home with an incentive layer. I think that's going to be a big part of this. People aren't necessarily going to want their AI workloads to go to the cloud. They might want to do them locally, and I think that's where the phones and all this silicon are going, with Apple's big focus on it.

Yeah. You bring up an interesting point. In the early years of this AI revolution—I'm talking about 2023 and 2024—there was an analogy that AI was like nuclear weapons. You'd hear the doomer crowd and the safety advocates saying that AI was this really threatening technology. They'd even say things like GPUs were like plutonium.

I think that model of the world is just wrong, because what we're seeing is—and Jensen actually had a pretty good line about this—"Nobody needs nuclear weapons. Everyone needs AI." It's true. Every consumer and every business is going to want to run AI. A lot of them are going to want to run it on their own infrastructure, and consumers are going to want to run it on their own phones. You're going to have an AI that's highly personalized to you.

Everyone is going to have AI. It's not like a nuclear weapon, where we want to stop all proliferation. AI is, first and foremost, a consumer product that is going to proliferate. Bearing that in mind, the question is: how do you create an appropriate response to the national security risk? But this idea that we're just going to stop AI and only have 2 or 3 companies that have it—which I think was the view a few years ago among policymakers—is ridiculous.

What we're seeing now is that, regardless of what certain policymakers might want, it's already highly decentralized. You've got 5 major American open-source companies, 8 major Chinese models, and everything that's happening with startups. This is going to be highly decentralized—

—and verticalized, right? All the Hugging Face models: there are specific ones for images and specific ones for video. It's going to be super fragmented.

The vast majority of this activity is benign. These are business solutions, consumer products, and viral videos. Most of this stuff does not rise to the level of a nuclear weapon or anything like that.

This is a good chance for us to talk about AI regulation. A lot of states are starting to look into regulating AI. California SB 53, the Transparency in Frontier Artificial Intelligence Act, is working its way through the system and could serve as a template for other states.

It was introduced in January as an alternative to the more sweeping bill, SB 1047. That bill would have required AI developers to conduct extensive safety tests before rolling out their models. It got a lot of pushback from tech, obviously, and Newsom ultimately vetoed it. This new law focuses only on the most advanced, large frontier models that we just talked about. It requires companies to release a framework explaining how they're approaching safety issues, including standards and best practices—whatever that means, and however safety is defined.

These are models, I guess, that have $500 million in annual revenue. I don't know how they picked that out, but it requires these companies to release transparency reports before deploying. They're going to be like the App Store, I guess, approving frontier models and their updates. That sounds great: you have to go to the government to release a new model. Your thoughts, David Sacks?

I think it's very concerning. There's a regulatory frenzy happening in the states right now. To be very clear about what happened in California, there was an original bill, SB 1047, that was incredibly obtrusive. Newsom vetoed that, but now they've passed a new one called SB 53. Like you said, it's not as burdensome and intrusive as the previous version. It focuses on making frontier AI models report safety risks. They're supposed to report if they have—

Can I stop you there for a second? What is the safety risk they're going to be required to report? It's such a nebulous term. What safety? What? Is it that the AI is going to jump out of the computer and murder me? Is it safety that it's going to give me the wrong answer?

They're supposed to report on potential catastrophic harms related to cyberattacks, biothreats, and model autonomy, which is the Terminator scenario. They're also supposed to—

Okay.

—let the government know if there's a safety incident. All these things are quite nebulous.

It’s almost like a nuclear power plant having to report if there was an incident. Are any of these, in your mind, thoughtful?

Let me just interrupt for a second. I think it’s the equivalent of saying, “I need any factory to report to me on the risk of a nuclear explosion,” even though the factory might not be working with nuclear material. You see what I’m trying to get at here?

I’m confused.

It effectively uses terminology that makes everyone nod their head and say, “Oh, yeah, that makes sense. That’s a good idea.” When the reality is that the legislators have no concept of what they’re talking about. They have no concept of how these models are built or deployed, and they’re using language that they think is inevitably going to result in giving them, ultimately, tools and control over a private market system. That’s fundamentally what I think a lot of this comes down to.

So—

Yeah, actually, I think that’s a really important point. Just let me give you some stats on this regulatory frenzy that’s happening. All 50 states have introduced AI bills in 2025. There have been over 1,000 bills in state legislatures, and 118 AI laws have already been passed across the 50 states. The red-state proposals for AI in general have a lighter touch than the blue states, but everyone just seems to be motivated by the imperative to do something on AI, even though no one’s really sure what that something should be exactly.

And there’s no real agreement on what all these AI regulations are supposed to do. They’re just making things up.

Or what the risks are.

That’s what I’m trying to get at. So let me ask you a specific question.

Yeah. Well, I was going to finish the point about California. California has gotten to this point where now it’s about reporting on all these safety risks. If this is all it was, then it would just be basically a bunch of red tape, and it wouldn’t be so bad. The problem is that you’ve got to multiply this by 50 states.

You’ve got 50 different states, each with their own reporting regime, which is going to be a trap for startups. They’ve all got to figure this out: what they’re supposed to report on, what the deadlines are, and who to report to. I mean, this is very European-style regulation, actually maybe even worse than the EU, because the EU tried to basically harmonize to get to 1 authority. We’re going to have 50; they’re going to have 1.

The other problem is that this is just the camel’s nose under the tent. Even in California, Scott Wiener, the legislator who did SB 1047, has now done this. He’s got a bloc of legislators, and they have 17 more AI regulation bills that they want to pass. So this is just the beginning.

If you want to see where this is going, look at Colorado. We should talk about this Colorado bill because it has already been passed into law, and it’s called SB24-205, Consumer Protections for Artificial Intelligence. It was passed back in May 2024, so it was one of the first to pass. Even though they didn’t really know what they were trying to regulate, no one’s quite sure how to implement it.

What the law does is ban something they call algorithmic discrimination. Algorithmic discrimination is defined as unlawful differential treatment or disparate impact based on protected characteristics—things like age, race, sex, or disability. If any of those factors drive an AI decision and it results in a disparate impact, then both the developer of the AI model and the deployer, which means basically the business that’s using it, can be in violation of this law and can be prosecuted by the Colorado attorney general.

Let me give you a practical application here. Let’s say you’ve got someone like a mortgage loan officer who’s reviewing applications. Let’s say they don’t even discuss race; it’s not on the form. They’re just using race-neutral criteria, like a credit rating or financial holdings, something like that. If the result of their decision nevertheless had a disparate impact on a particular protected group, their decisions could be found to be discriminatory.

Moreover, the developer of that model could be liable even though their model just gave an answer that, under the circumstances, was truthful. The only way I see for model developers to comply with this law is to build in a new DEI layer into the models—to somehow prevent models from giving outputs that might have a disparate impact on protected groups. So we’re back to woke AI again. I think that’s the whole point.

Yeah, that’s the whole point of this Colorado law. But let’s get Chamath in on this discussion. Chamath—

I think that this is really, really dumb. If you have 50 sets of rules, what you will have are some conservative versions of AI. You’ll have some progressive-leaning versions of AI. These 50 sets of laws will essentially render this industry impotent and incapable of maximizing itself and actually doing what’s necessary to drive productivity and GDP on behalf of the country.

There is no conceivable way, as Friedberg said, that anybody in Sacramento or Little Rock—or, you know, name your state capital—will have the intellectual wherewithal to get to an answer as good as the federal government will and, as Sacks would tell you, just to be totally honest with everybody. So what should happen here is that there needs to be a complete moratorium, and the federal government should be given the time to figure out what the framework should be so that there is one size, one set of rules.

Now, if that doesn’t happen and this is allowed to stand, there is a perfect example of where this has happened before, and that is in the car market. What happened was that there was a complete set of rules in California for emissions that was entirely different from the rest of the country. You can look and see what it did. Now, that’s just 2 sets of rules—

And what—let me, let me, let me finish.

Okay.

So what did going from 1 set of rules to 2 do? It drove most of these companies toward barely breaking even or massively losing money. It has been something that the entire industry has been fighting back on for 10-plus years now. Can you imagine that instead of 2 sets of rules, you have 50? I think you know what the economic consequences will be. You’ll render this entire category incapable of generating any positive economic output.

I guess the steel man, if we were to make one, is transportation, education, abortion, taxes, alcohol, and cannabis. I think I mentioned those are all state—

Cannabis is a poison, and it is the worst thing in the world.

Right, but for you—

Okay, that’s your opinion. Great. But should states have some general rights?

Oh, okay. We know your position on that. I’m talking about the difference, which is: what should states do?

Perfect. Well, I don’t disagree with that statement.

The question I’m asking is: We let states, just to steel-man this for the audience, decide how they want to execute against things like taxes, alcohol, education, abortion, and transportation. Should, David Friedberg, states have some rights here? I’m just steel-manning this. I’m not saying this is my opinion. But if this is the most transformational technology of our lifetime, shouldn’t states have a say? Or what’s the argument for states having a say?

It’s the United States. It’s a federated republic. I am 100% in favor. I think what we’re pointing out is the idiocy of these decisions, for number 1.

Number 2, the internet created a virtual network system for media, communications, content, productivity. We’re talking about something that stretches across the federal landscape. What needs to happen is federal preemption. The federal government—Congress—needs to pass a law that says, “Here are the standards that we are going to set,” or, “Here are the rules that we think are relevant for AI. Here are the things that states can and can’t do if we want this country to succeed on the opportunities and advantages that will arise from AI.”

The second thing I’ll say is that much of the law being drafted by these state legislators consists of regulatory-oversight laws, not laws that define a new civil or criminal penalty because of something you did that caused harm. They are specifically written in such a way that they say, “We need to have oversight. We need to have review. We need to have control over your systems because we get to review them.”

They don’t say, for example, “If your AI kills someone, you are going to jail.” That is what they should say. In fact, one could argue that many of the civil and criminal statutes that already exist in the states cover much of the harm that is already being discussed as the potential safety risk associated with AI.

You don't actually need more because, at the end of the day, if the AI system, the producer of the tool, or the user of the tool causes harm to someone, something, or some business, there is already a statute to protect against that harm. The statute that's being drafted is all about oversight. It is about giving the government regulatory control—the ability to go in, interrogate and investigate, and create approval systems to determine whether what you're creating as a private-market business or citizen is appropriate to be used.

It is one of many points of overreach that this federal republic has historically been able to withstand. After 250 years, the day may be up.

So, Sacks, in the case of a large language model being constructed in a nonthoughtful way so that it could be used to conduct cyberattacks, dox people, or be used for impersonation, I'm trying to think of a scenario involving these security capabilities that would be concerning. I don't know—if OpenAI allowed its AI to hack credit cards, that's already illegal, right?

It's already illegal to conduct a cyberattack. If you manage to take an AI model and use it as a tool to perform a cyberattack, that's still going to be illegal.

The same thing applies in Colorado. They've got this bill that wants to outlaw algorithmic discrimination, but discrimination is already a violation of the law. What they're doing there is not just going after the business that's performing the discrimination. That's already illegal. What they want to do is get into the tool itself, and they want to make the developer liable if their model creates an output that supposedly ends up creating a disparate impact in a decision.

Imagine if we did this with the internet. Imagine if we went back to the start of the internet and said, “Hey, if someone uses the internet to do something bad, therefore the government needs to approve everything that's done on the internet.”

You can talk about mobile communications. You can say, “Okay, Verizon's responsible if people use it in a terrorist attack. Verizon's not responsible if people use it to coordinate a bank robbery.” That's so obvious. So, yeah, this does seem like overreach.

David Sacks, what is the situation on Capitol Hill regarding a conversation about creating federal preemption—passing a bill that says the federal government is going to set standards around AI utilization that states cannot intervene in—and creating a mechanism that allows this market to develop and allows things to prosper?

Here's the situation: in the One Big Beautiful Bill, there was a federal moratorium on state AI regulation. I think it was well-intentioned and motivated by the fact that we do see this huge knee-jerk reaction from state legislators who want to do something without knowing what it is they want to do. However, there wasn't enough Republican or Democratic support for it.

I think part of the reason Republicans, in particular, have been opposed is that there's so much anger at the big tech companies right now for all the censorship that happened during, especially, COVID, but even before—and you still see it. You saw it with this Wikipedia news, where they're banning all conservative publications from being sources. There's just a lot of anger toward the big tech companies and tech bros, and basically, there are a lot of Republicans who don't want to get on board with anything perceived as helping tech.

The reality is, who does that ultimately benefit? Ultimately, it benefits the blue states that are leading on this type of regulation. It's Gavin Newsom, who just signed this new bill. It's, again, Jared Polis in Colorado, who ultimately signed this Colorado law. If there is no federal standard, what you're going to see is that the blue states will drive this ban on so-called algorithmic discrimination, which will lead to DEI being promoted in models, which is what the Biden administration wanted. You will see the return of woke AI at the state level. It's not something any Republican should want.

I understand the justifiable anger at these tech companies because their behavior in the past has been really bad toward conservatives. They did engage in a lot of censorship, shadow banning, demonetization, debanking, and all that kind of stuff. I get it. But we have to look at what the results are going to be. The single federal standard is the best way to make sure that we do not have woke AI, that we do not have insanely burdensome regulations that allow China to get ahead of us in this AI race, and that we actually have truthful, unbiased AI instead of highly ideological AI.

Do you think you can get it done?

Let me go to Polymarket: “The U.S. enacts an AI safety bill in 2025.” Not getting done this year.

Here's the good news: it doesn't really matter what I think. The important thing is what President Trump thinks. In his July 23 speech on AI, he was really clear that there needs to be a single national standard for AI. He said it was impractical, that it doesn't make sense to have 50 different regulatory regimes, and that this could cost us the AI race.

He would like there there to be a single federal standard, just as he promoted for vehicle emissions. Again, we didn't have a federal standard there, and then it was California taking the lead, followed by the blue states setting the standards. President Trump didn't think it made sense for California to set the rules for the whole country, so the feds preempted that. I think we should do the same thing on AI. That's what the president basically said in his speech.

I think the administration ultimately will support this, and I think more Republicans will come on board as they realize that what the blue states are doing here is not helpful for conservatives and is not helpful for having an unbiased information environment.

I'm torn on this one. I moved to the great state of Texas to have certain freedoms that we have here that we don't have in other states, and I kind of like the idea of states having certain rights. But I don't like the way these laws are being written, so I remain torn. The devil's going to be in the details on this one. Chamath had to bounce.

Do you like the Colorado law? Would you like to have that?

No, of course not. It's how these laws are executed that's my concern. I had this concern with gun rights in California. You should have the right to own a gun, and then they're just like, “Well, you can't have a gun.” Okay, well, you know. Then the states have to go back and forth in these lawsuits to determine whether New York City or San Francisco can ban guns. One of the reasons crime is out of control in some of these places is because homeowners can't have guns, along with the stand-your-ground laws and so on.

One of the nice things about this country is that you can pick a state where you want to live. I want to live in a state where abortion is legal. I don't want to live in a state where abortion isn't legal. I want to live in a state without state taxes, or one with taxes. You get to choose. It's one of the powerful things about this country, and we get to debate these things in real time.

I do have a concern about centralized government and overreaching federal governments, especially with the way executive power is being deployed these days, from Obama to Biden to Trump. There's too much executive power, in my mind. So, I have concerns on both sides of it, but the devil's in the details of the execution. I trust you to come up with something good as our civil servant, so come up with something good, Sacks.

We will. But just to go back to one of your points on states' rights, look, there's a Commerce Clause in the Constitution. The reason that exists is to create a seamless national market economy.

One of the reasons why the U.S. has such a strong economy—why it's the number-one economy in the world—is because we have a single national economy, which is the largest market for products. Imagine if we had 50 separate markets, each with its own rules and regulations. Doing business in the U.S. would be like doing business in Europe.

Remember, one of the reasons why the U.S. dominated the internet in the 1990s is that if you launched a startup in America and won the American market, you were basically right there in terms of winning the global market. Whereas, if you were in a European country and won your local market, whether it was the U.K., the Netherlands, France, or somewhere else, you would have won only a small part of Europe. Then you would have to figure out all the rules and regulations to get into the other 30 European countries, never mind the rest of the world.

It's that seamless national market that's given our companies the scale they need to then dominate across the world. If you restrict that by making every state have different laws for every product, we're going to lose that massive advantage that we have.

Here's the thing. I look at the car standards that Chamath brought up, Friedberg, and, you know, Trump, I guess, doesn't want California to have its own car standards. That got rid of 70% of the pollution in California. I was in favor of that. I wanted to see higher standards, not lower standards, because I don't want to pollute.

The smog over California, especially Los Angeles, was insufferable at times. Those standards—which led the nation, which led the world—did they add extra cost? Of course. But they made California a great place to live, because it's a car culture there, and people were dying and losing years off their lives from the smog.

So that's an example of it working really well, I think. And I am for cannabis regulation and for it being legal. California led the country in that, whereas other states want to ban cannabis and don't want to have higher standards for pollution. I like the fact that California led in those two ways.

Now, it's all in the execution, of course.

The problem is that because California is such a big market, those vehicle emission standards that may or may not have been right for California apply to every other state because the car companies can't manufacture different models for different states. Nor should they have to.

They did, though.

Practically, they did produce different models for different states. But you want to have different AI models for every state. You want to have a DEI model for Colorado.

In the case of cars, I do like the fact that California pushed the car companies to make cleaner cars. In the case of AI, that's why I was asking you which safety concerns you have, because I'm trying to find a safety concern that we can all say is a legitimate concern for AI, and we can't come up with one.

That's the interesting part about this: There are obviously overreaching laws right now because we can't come up with something where AI is going to jump out of the computer and do something in the real world that regular laws don't account for. We can't come up with an example here, and we're deep in this industry. Can you come up with a single example of AI doing something bad in the world that we should be concerned about that isn't covered by existing laws? I can't.