Anthropic's $2T IPO, Zuck's AI Manifesto, Nvidia's $500B AI Bet, Grok's Comeback

Welcome back to the All-In podcast. David Sacks and Gavin Baker are with us this week. We've got a short crew but a long docket, and we're going to rock it. How are you doing, Gavin?

Fantastic.

You're still in that SpaceX afterglow. This is the biggest win of your career, correct?

SpaceX was a magical moment, and there's only one SpaceX. But to quote Bill Belichick, it's on to Cincinnati. There's a lot ahead for SpaceX, so the SpaceX story isn't over.

In this business, you're only as good as your last investment, so everyone's asking, "What now?"

Sacks, how's the summer wrapping up for you?

Good. I'm looking forward to talking to Gavin.

You missed a really great discussion we had on the All-In interview show with Rahm Emanuel.

How did that go? I heard it got kind of spicy.

It got super spicy. Rahm came in full guns blazing, but it's a really great interview. We'll take anybody who wants to do the interview show who's running in 2028. Did you catch it, Gavin?

I haven't watched it yet.

Okay, you've got it in the queue.

The verdict on Rahm is that he's kind of a throwback to the Clinton-Obama DNC, with no chance whatsoever in the Mamdani-DSA Democrat party.

That's the great irony, and it's a very similar moment to what happened in your Republican party, where you had this internal civil war. I'm hoping the moderates have a chance here, but who knows what's going to happen with these lunatic democratic socialists.

He's got an incredible amount of experience, having worked for Obama and Clinton, been mayor of Chicago, and ambassador to Japan, so he was awesome on foreign relations. And to your point, he thinks hard no on democratic socialists—the socialist policies are a road to nowhere. He went totally hard at it, as hard as you would, Sacks. Nothing would be greater than to have more moderate candidates; I think we can all agree on that. It's worth checking out.

Breaking news from the docket: we record on Thursdays and publish on Fridays, and at the time of this recording, the Financial Times reported Wednesday night that Anthropic is targeting a $2 trillion IPO. That would break the record-setting $1.75 trillion IPO by SpaceX. Investors expect the company to go public in October — six to eight weeks away — which lines up with Polymarket, where sharps put an 80% chance on an Anthropic IPO this year, and Brad Gerstner has also signaled it would happen sooner rather than later.

On revenue, as we've discussed, the annualized run rate will end the year between $100 and $120 billion. That's an extraordinary ramp-up of revenue we've never seen in Silicon Valley. At a $2 trillion valuation, that's 16 to 20 times sales — a fraction of what SpaceX, Palantir, and other highly valued stocks went out at. Revenue growth over the last year was 10x. Anthropic still has AI's top model, but Grok is catching up quickly and OpenAI isn't too shabby either; Polymarket gives a 67% chance Anthropic still has the top model at year's end. Also breaking Thursday: Anthropic is in talks to buy an AI startup called DART for $6 billion, whose software lowers the cost of AI training and inference by making chips more efficient.

Gavin, what's your take on the revenue ramp and what to expect from the IPO — and are you an Anthropic shareholder?

A few things. I'm not currently an Anthropic shareholder. I probably came a little too late on Anthropic, SpaceX, and Databricks, and I generally don't invest in Elon's competitors — I don't think it's a wise course of action.

Right. And Databricks — Elon decided that when he had so much extra compute with Colossus, he would rent it to Dario and Anthropic.

Absolutely. And probably more importantly, he got comfortable with Dario's morals — that Dario is a good person, and like all humans, to err is human. Maybe he's made some mistakes in how he's handled government relations, but Elon got comfortable that his morals were in the right place. As Elon has said, they don't set off my evil detector — and he has a pretty good evil detector; he famously cut his conversation with SBF short.

On Anthropic itself, the rumored numbers are exceptional — we've really never seen anything like this. I still think those moments in April and May were some of the most extraordinary moments in the history of capitalism. And what's wild to me is that, probably on the margin, Anthropic is losing share to OpenAI, to open source, and to Grok — and they're still growing this fast.

Why is that? This is important for people to understand: the pie is getting ginormous. Even if Anthropic is losing share on a percentage basis, the real numbers keep growing — and remember, open source is dark tokens that aren't tracked anywhere. This is a major pie-growing moment.

That's the point. AI broadly accelerated in the month of July, which was crazy to contrast with the public stock market action — now we have a higher-resolution view of what was happening.

A few things on the IPO itself. It's an exceptional business; they've executed really well, and not just on the model — their products are amazing. I always think of Eric Vishria, a good friend who I don't think you've ever had on the pod but would be a great guest. Nearly two years ago he had a thought experiment that OpenAI and Anthropic would still hold a lot of value even if they lost at the model layer, because the product, the harness, and user familiarity are all so important. I think there's an element of truth to that.

As for the IPO, the first thing I'd say is that I'm a little skeptical of the $2 trillion figure. We saw this with SpaceX: if everyone is jockeying to be lead left on the deal and you lose that slot, the first thing you want to do is make whoever is lead left look bad. So you leak to the press and say, "Hey, we're in the room — the bar is a $2 trillion IPO."

You said on a previous episode three to four trillion.

Absolutely — though that's probably where it might trade. We haven't seen the S1 yet. But if you're bringing a company out responsibly, you want to price it in a way that lets you absorb the lockup. You don't want to price it really high and create a lot of volatility — that's very disconcerting for employees, and you don't want anyone distracted at this moment. So I think they'll price it smartly. Maybe that is two trillion; maybe it prices at two and trades to three.

Just remember, these leaks always come from the bankers who probably lost lead left and want to make the lead left look bad. So if it comes out at two trillion, it probably means the testing-the-waters and the roadshow just went incredibly well — people do want to own this. These numbers are unprecedented.

So are you saying the roadshow probably already occurred, or quietly occurred?

No, definitely not. I'm saying that if they price it at two, it means the testing-the-waters went well, the roadshow went well, and they've got a lot of confidence about where it's going to trade.

Can the revenue ramp continue?

I think it can continue, although it's a legitimate question. If you're ending the year at, say, a hundred billion run rate — up 10x-plus year over year, and by the way they've grown 10x year over year for the last three years, so it's growing exponentially — then if that rate continued, they'd hit a trillion dollars of ARR by the end of next year. The question you have to ask is: is the TAM big enough for that? But also, is there enough compute? Is there enough energy? You start to get into physical constraints.

I do think the TAM is big enough and the demand will be there. There are so many new applications of AI — agents are just taking off now, as you saw with the launch of Grokbot — and all these other companies will keep extending AI into more and more contexts. So demand for tokens will keep growing exponentially. The question is whether they can physically meet that demand, and then also Gavin's point about market share: other companies are seeing what they're doing and competing more effectively to take share.

Remember, Anthropic made this huge bet on coding, and it ended up being a very prescient bet. I don't know if it was ideology — that they thought coding was the way to get to recursive self-improvement first — or whether they saw Cursor's utilization and said, "We're going to take that business." Either way, they've started moving into verticals where they see their own users building successfully on top of their platform, and that was definitely a signal to get into coding. Nonetheless, they got into it first and very successfully, and it led to a boom.

We hear that OpenAI has now basically pivoted to coding and is doing a better job of it with GPT 5.6. Their growth rate is accelerating: they used to grow 3 to 4x a year, but I've heard that over the past two months their growth is now over 20% month over month — which, extrapolated over 12 months, would be a 10x growth rate. So OpenAI is now growing as fast as Anthropic after making that pivot. We can discuss whether xAI may get there as well.

In any event, regardless of whether Anthropic can grow 10x next year, I think they're going to grow very fast. The question is just what their market share looks like relative to their competitors.

It's worth double-clicking on two points there. The first is whether there will be a market for these tokens and whether growth can keep up—in other words, is there demand? I think all three of us agree: yes, of course there's demand. But the next two pieces, which Sacks mentioned, are energy and data centers, and that's a question I want you to answer. Can they get to one trillion in revenue? Let's slow down and do some back-of-the-envelope math—this is your wheelhouse. What would it take in terms of infrastructure for them to go from a hundred billion dollars a year in revenue and 10x it, and is that humanly possible in the next year?

A few thoughts. I think those are the right questions. First, internally Anthropic is very confident. I've been told by multiple people I trust that Dario has said Anthropic might be the only private company in the world at some point. Think about that. In this Anthropic-maximalist vision, there's Anthropic, and then there are governments, and that's it. So there's a lot of confidence. I'm sure they have more advanced checkpoints than Fable up their sleeve—they've executed really, really well. But I would probably take the under on them being the only private company in the world. I think it's going to be a long time before they lay a rocket or—

—deliver your lunch. There might be other businesses in the world. Zipline might want to bring you a burrito, Uber might want to take you somewhere, Waymo might want to drive you home. But sure, Dario.

I might interpret that as a negative signal, because it's so hubristic. This is getting into SPF land a little bit. I would certainly discourage Dario from saying that ever again to anyone.

So AI Jesus should not claim that he is AI Jesus. Okay.

Pride cometh before the fall. And that will be a long fall if they—

But look, I have heard from other people at the company that when they were at about 60 billion of ARR, they thought they could get to 600 billion in one year—meaning they thought they were on that 10x ramp from the middle of this year. Now they're at over 80, and it doesn't seem to be slowing down. So I think the big question is how fast they can grow next year: does it go all the way from 100 or 120 to a trillion dollars? If they merely got to 400 to 500 billion, that would make them the biggest software company.

If they get to 200 billion, that's an amazing outcome.

Where does that put them versus Microsoft's Office franchise, or Google's search franchise? We're starting to become a peer to those two companies.

Absolutely. Just to break the question down into demand and supply: on demand, depending on how you count it, there's 25 to 65 trillion in knowledge work. 25 is kind of the number I've used, and I was told by the head of the AI institute at one of the three largest investment banks that it was way low.

So this is the TAM they're going after today, before we have robotics.

And then it's a question of whether it's labor substitution or accelerating growth. Thus far, it really does look like it's accelerating growth. There are more job openings for software coders today than there were a year ago—more demand for coders—and that's the epicenter of the blast zone. There is some evidence that for very young people coming right out of college, maybe there's been a negative impact. But broadly speaking, it has to come from one of those two, and it's better for everyone if it comes from accelerating growth. These companies are now at a scale where, if it's coming from accelerating growth, we will start to see that in the national accounts over the next 18 months.

Okay—you want to go to the energy part?

Supply, yeah. And listen, this is where the rubber meets the proverbial road. The people who build these data centers—the best way to understand it is go watch that series Landman. It's like Billy Bob Thornton's out there, you know, per basin time.

The Permian Basin is where the rubber meets the road—you're in the patch, actually worried about narcos. It's the tip of the spear: 110-degree heat, and you have to orchestrate thousands of people in these remote locations. It's atoms, not bits, and it's really, really hard. But America is actually very good at this once we gear up, and capitalism is very good at solving these problems.

Everybody talks about the ultimate constraint being turbine blades, which are only made in a couple of facilities in America and Europe—each one a giant building with a 40-ton press. Now they're running 24-hour shifts. Necessity is the mother of innovation, and nothing like capitalism will make that happen. Boom Supersonic is making a new jet, and I know Sacks has one on deposit. Their response to demand was, "We're going to build a turbine business first." And I understand there was a news report that Elon bought a turbine company personally—which tells me he just wanted to go fast. On top of that, residuals for private jets are extremely strong these days, largely because people are taking jet engines off old planes and repurposing them as turbines to power data centers. It's a big business, growing really fast, and Caterpillar, Cummins, GE Vernova, and Siemens Energy are all expanding capacity fast. So I think the supply question will get solved—the Billy Bob Thorntons of America will work with those companies and get these turbines on the ground.

The real constraint is politics and regulation. Texas has announced what I think is probably— an energy audit, yes. Governor Abbott announced this week he's going to do an energy audit. Not that he's going to slow things down, not a moratorium, but everybody has to go through an energy audit. If you bring your own power, you'll be fine. But if you don't have a solution for that and want to tap the Texas grid, you may need to pass an audit, because we don't want the whole grid going down—which it does every January when things freeze over.

Right now the narrative on AI is not good. It's that data centers take all the water—based on a mistake in a book three years ago that was off, I think, by a factor of 100,000 in the amount of water a data center uses. The claim that they raise your cost of electricity is just wrong; they lower it. The Wall Street Journal just wrote a story about a town called Ellendale where a data center revitalized a dying town: tax revenue went up 10x, and it's been the best thing that's ever happened there. But nobody's telling those stories, and people need to. As we've discussed on the show many times, the Chinese Communist Party runs a very active anti-AI, anti-data-center campaign because that's what's good for them and bad for America. Someone needs to tell the truth about AI. One of the best phrases in the Bible is "the truth shall set you free"—but it can't set you free if nobody tells the truth.

The data center panic is a testament to the power of the media to create hoaxes and hysteria over literally nothing. Every single aspect of the data center story is wrong. They don't use up a community's water—the water use is quite manageable, less than a golf course, and it recirculates. They bring down electricity costs, assuming the data center stands up its own power generation, because they sell the excess back to the grid and pay for grid upgrade costs. The real reason electricity prices are rising in blue states is decarbonization and other green regulations that make power generation more difficult. What about the other accusations? Well, noise pollution is legit—nobody in their right mind is putting one in anyone's backyard today. We have a lot of people in this country. People did do that for a bit, though.

If we just use common sense, there's no problem. The idea that you have to have a complete ban on data center construction is just silly. And the communities that are embracing them are seeing they have better schools and better services because they're making a lot of revenue.

And to your point, Sacks, when entrepreneurs get on something, we're suddenly seeing solar, batteries, nuclear, and more clean energy being tapped to power these data centers. That demand will lower prices over time—solar panels and batteries will get cheaper because they'll be scaled much faster, which is good for humanity. This is basic, folks.

The green stuff can't scale in the short term. It's all gas-powered for the new data centers—that's the only way to get the amount of scale you need in the next year or two.

I think solar will be viable, but it's obviously not as fast as popping up natural gas.

Yeah, I largely agree. I will say, Elon is building a $10 billion new solar manufacturing plant in Texas, my home state, for a reason. I think the world will eventually run on sunlight. But we're going to run on natural gas for a while. We have a lot of natural gas in America. It's a clean fuel—the greenest carbon fuel. And this media hysteria about data centers is doing such a disservice to low-income Americans, to small-town Americans, often to minority Americans. Data centers are so good in every way, and the media is doing such a disservice. No one's taking the other side, and someone needs to take the other side and tell the truth.

I don't know who's relying on the media for their facts now, but I think there's a reason we were number four in the world last week—we actually are part of this buildout, we know about it, and we can explain it to you a little more succinctly. Noise pollution is an issue. Air pollution from natural gas is an issue. Keep these things far away from people, as Sacks just said, and you don't have as much of an issue. There is a minor issue: if you put these at scale in a certain location, you can raise the temperature of that valley or wherever you put it a couple of degrees—that's something worth monitoring. These are all absolutely manageable issues. And if you really want to get obsessed about water, please, for the love of God, stop eating almonds—and maybe the golf course doesn't have to use a billion gallons of water every year.

On the demand side and the pricing side, I think Anthropic's growth rate is going to slow considerably next year. It's been on a tear, but there's no way they get to a trillion in revenue next year—maybe they could triple and get to three or 400 million. The reason is watching startups, and also some large companies, now embrace GLM 5.2, and Zuckerberg—which we'll get to in a moment—releasing really great open-source models. Just look at the pricing of GLM versus Claude Opus: it is 90% cheaper to use these products. I'm using them, the founders I invest in are using them, and increasingly corporate America is going to embrace these solutions. It's just a little bit harder to set them up, and it's coming, folks. If you think corporate America is not going to embrace open source, you have not been paying attention to the last 20 years, when they did exactly that. Twenty years ago there was a big argument—oh my God, corporate America will never trust open source. And corporate America said, "We don't want to get rug pulled or have some large tech corporation hold us hostage for their technology, which they're experts at doing. So open source is the safer bet for us." That's what's happening in corporate America today and inside of startups. This is going to be very real.

Gavin, you're jumping out of your seat.

No, this is a larger point. For sure it's good for corporate America, for sure open source is good for the world, for sure it's good for America. But Dorcash, a podcaster—a thoughtful guy close to the AI scene—posted something that really resonated with me. He said:

At the end of the day, Anthropic's AI and their constitution is wired to do what Anthropic thinks is best for humanity.

Let's take Elon's judgment as gold-plated. His "don't be evil" detector did not go off, and they're genuinely good people—which I actually believe them to be—doing what they think is good for humanity. I actually believe that. The problem is that there are always unintended consequences, and the world is chaotic and unpredictable. What's great about open source is that it means we're going to have a rich variety of AIs, a diversity of AIs, and that is a good world for humans in America. We do not want one, two, or three dominant models.

That's why—not to always talk about Grok, but it's great—I think history will judge Elon kindly for having Grok dedicated toward the pursuit of objective truth.

"Objective truth."

Yeah. The whole point, Gavin, is that if you made the bet, that means you used your intelligence and your savvy to pick what you think the winner is. I think that's super accurate. And I think Elon explicitly believes in open source—he believes this is a technology that should be shared for the good of it. In fact, that was his first foray into this: backing OpenAI and OpenAI's manifesto.

That manifesto was rewritten this week, topic two, by Mark Zuckerberg—quite paradoxically, or ironically—in a 6,500-word essay titled "The Future Is for Everyone: The Path to a Positive AI Future." He laid out an extremely optimistic AI worldview: as "OpenAI 2.0," Facebook would release open source models, create an agent for everybody, make it free, and provide everyone in the world with their own tutor and unlimited abundance. He has picked up the abundance crown that Sam Altman dropped when he went for the private company. He says superintelligence is invention, not automation, and that AI safety should be based on the balance of power—no singular, centralized intelligence. Sacks, I'm assuming you read or scanned or got the coverage of Zuckerberg's manifesto—any thoughts on his positioning?

Yeah, I read it and found myself nodding in agreement with much of it. I really like the point he made early on, where he took a shot at the effective altruists and Anthropic. He said, look, Dario, why are you guys rushing to create a future that you don't believe in, that is so dystopian? If you think this is such a bad future—where everyone's out of work and AI goes rogue and creates all these safety issues—why are you so excited about creating it? I think he points out a fundamental contradiction.

There are two possible answers to that question. One is that they could be trying to create a future they don't believe in for mercenary reasons—it could be lucrative—although I don't think that's the real reason. I think most of those guys are missionaries, and what they believe is that the future is dangerous and only under their enlightened control can humanity be protected. It's right out of The Vision of the Anointed, a book Thomas Sowell wrote 30 years ago, where intellectuals throughout history have thought that if they're maximally empowered, they can engineer society in a more benevolent direction.

And this has always backfired. It's always led not just to empty or broken promises, but also becomes an excuse for totalitarian schemes and state power. Zuckerberg goes on to say something that gets at this. Right after making the point about why you'd rush to build a future you don't believe in or want, he says:

The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic. Historically, hoping that an absolute power will benevolently provide for humanity, if sufficiently enlightened, has not led to safe or positive outcomes.

The point he's getting at is that a major frame on the whole AI debate is not just open versus closed—it's also centralized versus decentralized: whether you think a centralized outcome or a decentralized outcome is the right one. I'm on the side of decentralized.

How could you not be on the side of decentralized? It's the most obvious argument ever.

The EA types are not—and by the way, they're delusional. I experienced this when I got to Washington. All the former Biden people had just moved over to Anthropic's government affairs, and I also dealt with think tank people who are part of this clique. Their frame on AI was that it was inherently dangerous and that we had to centralize control over it in the government—just two or three companies that should essentially work together, made into a cartel. Remember, Marc Andreessen had a meeting with those Biden officials where he came out—

Don't bother meeting.

They dispute that characterization. Maybe they didn't use those words exactly, but I believe Marc was correct in his interpretation of what he heard, because I heard similar things. They really believe that the way to control what they saw as malign, or at least very dangerous, was first to limit it all to the United States: don't export the chips, don't export the models.

The idea was essentially to retain total control over it in the US, which is frankly a pipe dream, because the rest of the world is going to get AI. That was part one. Part two was to form something like an atomic energy commission to cartelize the industry, deeply embedded with the government — effectively a merger of corporate and state power. It was an extremely centralized vision of the future, and I think they still believe in it to some degree.

Clearly Zuckerberg listens to the All-In podcast. We have a seat open next week — Mark, somebody reach out to him. We tape Thursdays around 10–11 a.m. your time; whether you're in Hawaii or Tahoe, we'll get a Starlink to you and you can do it from your e-foil. It would be great to have you on.

To wrap up, I think where Zuckerberg goes with this, and where I agree, is that we want maximum empowerment of the individual. That's the way to solve this problem: decentralized outcomes, open models as JCal was saying, data sovereignty. There can still be proprietary frontier models from Anthropic and OpenAI, but we want a decentralized, empowering outcome for individuals — and I think that's the side Zuck is on. I very much agree with that.

Gavin, I assume you read it — or did you not have time for 6,500 words?

If he writes nine more, he's got the average non-fiction book, which runs around 60,000–70,000 words. So Harper Business, Hollis — call Zuckerberg and get the other eight parts of this manifesto going.

I read it and agreed with it. I'd just super-agree with everything David said, and boil it down to this: Anthropic and people in the effective altruism movement believe this technology is too dangerous to distribute. What Mark Zuckerberg — and I think Elon and Jensen — believe is that it is too dangerous to centralize. And history has spoken: when given the choice, it is always better to distribute and decentralize. I can't think of a counterexample. Forget being an American — as a human, I want an AI that looks out for me, the individual, the sovereign individual.

My conception of what's good for America: I do not want Dario Amodei or Anthropic deciding what's in my best interests. I fundamentally don't want that.

You don't trust AI Jesus?

I mean, AI Jesus seems like a great guy. Although it is funny — somebody said he was cursed with resting smiling face. It's a little jarring to listen to him describe this terrible vision for humanity while smiling happily. He's literally got that grin on while explaining the Terminator scenario, that only he can save us, and that even he doesn't know — there's a 20% chance he might not be able to save us all.

Gavin, I love that framing about it being too dangerous to centralize. One closely analogous point is how you think about competition. The EAs think competition will create a race to the bottom and a lack of safety, whereas I believe competition brings out the best — they think it brings out the worst. There are probably examples of both, but fundamentally they want to centralize control and have less competition; we want to decentralize and have more competition.

I want it in the Constitution — I have the right to bear arms, and I think that's important. I grew up in Texas; it's kind of a side of defense against tyranny. And I want the right to my own AI, where its values are aligned with me — not with someone else, however good a person they might be.

The right to bear arms is a good analogy, because we do allow you to have a gun even though you might misuse it — guns are used to kill things. AI has much more creative and salutary uses; it's fundamentally beneficial, a consumer technology everyone will want to use. So if we allow you to have guns, I agree, you should be allowed to have AI.

Let me correct you there, Sax, now that you're in the great state of Texas with me — guns also help you feed your family and defend you against evil forces in the world. Those are virtuous uses of a gun.

Yeah, I get it — self-defense. But you'd still use the gun to kill or hurt the person in self-defense. I'm not against guns. I think this is the delusion of the effective altruists — you can look them up, folks. They think they're going to make all the money in the world, create the one company, the money-printing machine, and then save all of us. It's massive delusions of grandeur. And the technology is going to diffuse anyway; the idea that you keep the genie in the bottle is absurd.

I think the way they view it, Sachs, is that they see this literally as superintelligence and the nuclear bomb. If you convince yourself that you're building the nuclear bomb and the Terminator — something you yourself cannot control, a superintelligence that will make a billion Terminators and a billion bombs — they have literally deluded themselves, while taking peyote or LSD at Burning Man, into believing that's the reality. It's not the reality. You could unplug it at any time.

They're suffering from delusions of grandeur, period, full stop. This is not the nuclear bomb.

You're right about that. When I first came into Washington, that was the dominant analogy promoted by them and their related think tanks — the idea that AI was like a nuclear weapon and that we needed a global atomic energy commission to regulate it. A total pipe dream, never going to happen. But they were using that analogy, and I think the reason is that you don't have a right to build your own nuclear weapon. You have the right to a gun, but not a nuclear weapon, so they wanted to put AI in that bucket. But as Jensen has said, no individual, no consumer needs a nuclear weapon, while every consumer and every business needs AI. It's fundamentally different — a consumer technology before it's a military technology. The idea that it's primarily like a nuclear bomb is just absurd.

In fairness, Gavin, Cyberdyne Industries also thought they were building good technology. Science fiction has literally polluted people's brains.

If we're going to have a Terminator thing, I want my own Terminators who like me.

Open source the Terminator.

But simply boiling everything else down: I'm an American, I'm a patriot. If we ban open source, if we let control of this be centralized, we will lose AI, we will lose geopolitics, we will lose everything to China. It's just that simple. So if you're an American patriot, you should be for the distribution, democratization, decentralization, and individualization of this technology.

Hey, Jal, what are your thoughts on the essay?

Well, like you, Sax, I found myself nodding the whole way through, because he ran in front of the parade, grabbed the baton, and said, "I agree with the best takes here" — and the best take is open source empowering the individual, and that this isn't nuclear technology. I want to point out two other points that got buried because he tried to cover a lot of ground. The first is that I want to give him credit for his view of job displacement. He says it is possible, but that people fear automation will outpace individuals' capability growth, leading to job displacement followed by a difficult period as people learn new jobs. I think we can all agree that is a concern.

But there is no rule that AI must increase automation faster than it increases individuals' capabilities or demand for new skills. Recent statistics suggest it may be more likely that individuals' capability growth could match or outpace automation. I thought this was a really good framing, and it's just another way of saying what Jensen said: you're not going to lose your job to AI, you're going to lose it to somebody with AI. And that's the game on the field right now. Every time I use this technology, three new business opportunities open up and I create two more job reqs, and people are doing five or ten times as much work if they embrace the technology. That's a big if, though—whether people will do it.

The second thing I found pretty interesting was his play at having the right to index anybody's content and learn from it. The quote is: "All AI models are derived from human knowledge." Okay, fair enough—pretty plain vanilla. Some have tried to frame distillation as harmful, but he argues it's important to protect the principle that you can learn from anything you can observe. In other words, he's preemptively saying, I want to scan the world's data. Well, I have a message for him: Zuckerberg, please explain how this concept is compatible with the fact that you have sued every startup—many of which I've invested in—that has ever tried to index or use the social graph at Instagram or Facebook. There are hundreds of lawsuits where they sued people for trying to distill things like who the top influencers on Instagram are or who's in my social graph, and they locked that whole thing down. So there's a bit of hypocrisy here that reminded me of Sacks and our discussion on previous episodes about Claude ripping apart old books.

We had the distillation debate many times.

Let me just go back to this point: there's a huge irony in the fact that Dario did not get his way in Washington, for the most part, during this administration. If he had realized his dream of a highly regulated AI apparatus, I don't think the company would be set up for a hugely successful IPO. Why? Because their whole right to charge this massive premium for tokens—which is the point you were making before, wondering how sustainable it is given all these open models offering tokens for much less—depends on them being six months ahead of those open models. And that six months would be eaten up in the blink of an eye if they were subject to regulatory approval. Dario himself keeps writing these blog posts saying we need an FAA for AI. You know how long it takes the FAA to certify and approve a new airframe? A minimum of five years.

If you want to change a table in your plane, it takes eight years.

You know how long it takes the FDA to approve a new medicine? Years. You're going to apply that regulatory framework to AI model releases? Game over. Forget about it. Your six months will be gone like that, and your model will be commoditized. To Gavin's point, the Chinese competitors will catch up, and that's bad for the United States—bad for our competitiveness and bad for the AI race. Obviously, we don't want China beating us, but aside from the geopolitical dimension, it'd be horrible for Anthropic as a company, because all their pricing power depends on them remaining in the lead by six months. If they ever slip up and lose that six months—

What does that do to revenue?

Exactly. Now, I actually think their pricing power may be quite sustainable, because a decent subset of the market—not the whole market—is willing to pay a premium for true frontier intelligence. If the top 20% is willing to pay 10x, they're golden. It's like I said before: Apple versus Android. Most of the market goes with Android because it's cheaper, but a significant subset is willing to pay a huge premium for Apple because it's a better experience. Same thing: if you're a company in a competitive market and you want the absolute best, you're going to pay a premium for true frontier intelligence. But Anthropic and OpenAI are on that hamster wheel—they have to stay in the lead to justify that premium. And by the way, I think they can, and they think they can; in fact, they think their lead is growing because they're on the way to recursive self-improvement.

The "Going Direct" Movement and CEO Authenticity

Leaders are increasingly taking control of their own messaging. Mark Zuckerberg has reactivated his X account, Palo Alto Networks' Nikesh Arora has started writing long-form posts, and Nvidia's Jensen Huang recently opened an account as well.

Zuckerberg previously let PR teams craft and distribute his messages, but he has now transitioned to publishing his own long-form essays directly. While he likely received assistance, AI detectors confirm the text is bespoke human writing. Getting ahead of narrative issues by speaking directly is a major shift—one that would have been valuable had Meta adopted it earlier regarding social media addiction and teen mental health. Overall, X has become a primary platform for direct dialogue and long-form missives from tech leaders.

Having an authentic voice on X has become essential for public figures and corporate leaders. Corporate PR departments cannot fully protect a CEO or speak for them with true authenticity. While a founder with super-voting shares like Zuckerberg cannot be easily ousted, most CEOs risk finding themselves defenseless—especially after retirement, under new management, or when called before Congress. Cultivating a direct, authentic public presence serves as both personal protection and brand power.

How Open Source Increases the Value of Frontier Models

The availability of open-source AI actually enhances the value of frontier models rather than diminishing it.

If closed frontier models from Anthropic, OpenAI, or xAI operate at a metaphorical 200 to 250 IQ level, their value increases when they can orchestrate cheaper, open-source 150 IQ models. A superior intelligence becomes more valuable when it acts as an executive system, delegating and farming out subtasks to lower-cost computational units—much like lead physicists directing broader teams on complex scientific endeavors.

Who ran the Manhattan Project? Oppenheimer. You had Oppenheimer and effectively ten Nobel laureates—maybe thirty of them—but they needed the roughly 2,000, or top 10,000 in the world, physicists. So having more affordable intelligence actually makes the frontier even more valuable. A very plausible outcome is one in which frontier tokens are 65 to 85% of the economic value, open source tokens are like 80% of the volume, and that's good for everyone—everyone wins.

Tell me the revenue: Gavin Baker, Anthropic by the end of 2027. If 2026 ends at exactly 100 billion, what is their 2027 revenue?

I'll take the over.

Over what, a trillion?

A lot of estimates. I'm not going to take the over on a trillion.

Okay, I'll put 500 billion—over or under, Gavin Baker?

How about this: I'll take the over on whatever the Wall Street consensus is for '27, which I don't know—a 25 to 30% over on that, assuming this regulation campaign doesn't happen.

Basically going out on a limb saying the underwriters are going to sandbag revenue. Okay, congratulations, Gavin. Give me my answer: over 500 or under?

I think four to 500 billion is within the zone of realistic and achievable.

Hold on—under 500 billion, Sachs?

I was actually going to say 4 to 500 billion as exit ARR for next year. Seems very comfortable.

So right under. Okay, two unders. The only reason it's not higher is that I do think you somehow get into physical limitations at that point—but maybe not, maybe there'll be enough compute.

There's also competition. We all believe they're going to hit 400 billion next year and 4x—4x 100 billion?

Keep in mind they'll be at 120, so 4x would be more like 500.

I think the over/under really is 500 billion exit ARR for next year, and I think they'll do the over.

We were talking about them at 10 billion ten months ago. 10 billion to 400 billion is 40x, gentlemen.

Are they going to thank me for preventing Daario from shooting himself in the foot? He put two guns on two ankles and emptied both clips—and you had taken the bullets out. They should give you 1%.

I don't think the bullets were taken out. I think he had them pointed at his temples—he did the Deer Hunter, and Sachs pulled him down, and he just shot himself in the foot.

I think this is where their hostility to open models comes from: if they're going to slow down, they have to slow down the open models as well, otherwise they do get commoditized and their business goes away. However, the United States does not have jurisdiction over what happens in China, and China was not going to slow down. That was always a huge flaw in this whole conception. These EA types thought they were going to institute global compute governance—somehow get some global agreement with China. Total pipe dream, never going to happen. If we slow down, China is just going to race ahead. They were going to create Accords 2.0, and that didn't work.

Slow your roll, kid.

Can I say one thing that I think is also important about open source, and why I think AI is going to be good for humans and good for Americans? It's that if we are short compute—we're talking about compute slowing down Anthropic's growth because we're literally running out of energy, we can't bring on—

Yeah, we're running into physics, planetary scale limitations. And I think that's a reality.

Just think about how energy-efficient we are—all of our brains. To approximate our brains, you'd need a data center consuming the power of a million American homes running for probably six to n months, and then you'd get our brains. And then every time you want to answer a question, you need the power of about a thousand American homes to operate at the level of our minds. We're smart, and all those figures are going to get better, but we're so energy-efficient. There's going to be a use for the most computationally efficient intelligences on planet Earth, and those are human intelligences.

It's so well said, because if you look at what Cerebras and Grok and Nvidia are up to, they're making inference five to ten times more efficient every 12 to 18 months. And you're the chip guy here, Gavin. Then you look at the models—and by the way, congratulations, you got Zuckerberg on your ass. Go ask MySpace, go ask Snapchat what it's like to have Zuckerberg in the rearview mirror. That's literally the T-Rex coming at the frontier models. He is a beast at copying, mimicking, and competing. He's going to be the open-source hero from America, and he's going to have the best open-source model in the next year—I guarantee it, here on the audience.

I might take the under on that.

There's another American hero: Jensen Huang. You might see Nvidia at the open.

He doesn't like us talking about that—I got the back channel. He's like, "We're not in the open-source business. I mean, we do have NeMo and we can't give you a full stack." But one of the hyperscalers, by the way—we know Gavin—one of the hyperscalers told me they just did a nine-figure double Lindy reverse, where they went from doing a frontier model on 100 billion dollars in infrastructure, and they're like, it's going to be open source, not frontier. So people are looking at this saying it's time to maybe move over to some open-source models.

The All-In Summit is coming September 13th–15th in Los Angeles. Last week we announced the giants — Jensen, Satya, Jared, Gwen — all members of the "first name club." Now for the new names: Steve Hilton, running for governor of California, will be there, and Spencer Pratt will talk about his failed LA run, which was successful in terms of capturing people's imagination. Abigail Shrier, author of the fantastic book Bad Therapy, will join us with a very special surprise guest. Bestie Gavin might even make an appearance, as might Blake Scholl, CEO of Boom Supersonic, whom I mentioned earlier in the program.

This year we're turning the outside of the Shrine into a festival-like experience: a Frontier Tech Hub featuring AI and robotics demos, plus a gifting suite — like you're a celebrity. Merch.com has made an incredible All-In swag suite, and there will be a wellness suite with red light therapy, facials, and IVs — useful if you stayed out a little too late, Saxy Poo; that sometimes happens.

Don't miss these important conversations. We'll also be announcing all the DJs and parties — the reason Freeberg does this — and it's going to blow your mind. Apply today at allinsummit.com.

Speaking of our guy Jensen, sixth bestie here on the pod: Nvidia is partnering with Goldman Sachs, BlackRock, and others to raise $500 billion for AI compute. Jensen laid out his vision in an X article titled "Nvidia AI factory: compute is becoming an investable asset class," in which the firms will establish independent financing platforms raising third-party capital. The idea is to treat GPUs like a financeable, income-producing asset — maybe like mortgage-backed securities, hopefully not with the same outcome.

Here's how it works at a basic level: instead of a company paying billions to buy chips up front, they borrow money, buy Nvidia-based systems, and use them to start generating revenue — i.e., renting out compute — then use the cash flows to pay back the loan. Nvidia's role is matchmaker, connecting customers who need compute with lenders. GM and Ford, obviously, have their own finance divisions.

Gavin, what do you think of this announcement vis-à-vis round-tripping and the hand-wringing and concerns we talked about on this very podcast a year or two ago?

What's essentially happened is that Blackstone, KKR, Goldman Sachs — these firms would not have done this, and would not have done that CNBC episode with Jensen, if they did not believe Nvidia GPU compute was a financeable asset. What's really smart here is that you're getting some of the smartest asset managers in the world, who charge 2 and 20 for many products, to essentially validate this market. Nvidia is being a matchmaker, and what they're basically saying is: our compute, because it's so flexible, is going to have a long enough life that you can finance it at lower rates than other kinds of compute. I think that's smart and good for everyone.

It's also interesting that if you look at the underlying architectures of the big Chinese open-source models — Qwen, Kimi, DeepSeek, GLM — they're actually evolving in very different ways. That architectural variation works in Nvidia's and the GPU's favor, because it means you do need this more flexible compute to finance it and believe it can have a long life.

What Nvidia is doing that I think is really smart is saying: we're going to help create and grow this market, and we'll provide two very important functions. To Blackstone, Goldman Sachs, KKR, Apollo, BlackRock — there are six of them, I'm leaving someone out, my apologies — come to us if you have a deal, and we will give a residual value guarantee, which will further lower the cost of financing these assets. That guarantee can then be incorporated into those companies' underwriting. What that basically means is that after three or four years, Nvidia guarantees these GPUs can be rented at a certain rate, and the risk they're bearing is the gap between that value and wherever the market rate lands. Right now everything's going straight up, and Nvidia has better telemetry than almost anyone on the supply and demand of compute, so they can set that guarantee at a very smart place — they're only bearing maybe 25% of the risk, with some of the world's smartest underwriters in terms of reputation.

What could go wrong, Gavin? I mentioned mortgage-backed securities, and obviously the lifespan is the key issue here. The different hyperscalers were saying four, five, six years — Amazon and the big debate happened. But what we've seen, to your point with the open-source models, is they can be tweaked to run on the oldest hardware, so maybe we get years six through eight out of these.

More than that. CoreWeave said they are renting Ampere GPUs at economically profitable rates today for 2029. So an Ampere card that came out in 2020 is going to have a nine-year life. I think that's what Nvidia sees — it's like old American cars that get sent overseas.

Where could it go wrong? I'll tell you where it could go wrong — in theory, just so we're steel-manning it.

The biggest risk, to me, is not on the demand side. The biggest risk is that you get a glut of compute and an overbuild. In the same way we had dark fiber after the dotcom crash, dark GPUs would be a disaster for everyone — especially if you built out your compute infrastructure expecting a spot price of $30 to $50 a watt, as Elon said they were expecting. If too many people race to provide this compute, there's an oversupply and the market crashes — that's the risk factor.

In a weird way, though, all the political headwinds insure against that outcome, because it is so hard to build data centers, for all the reasons we said. There's a whole moral panic, slash hysteria, slash hoax going on, and it's those political headwinds that will almost guarantee there's no oversupply relative to the exponentially growing demand. So in a weird way, you're protected against that.

But Jake, could I shift gears for a second? Tell you what's so brilliant about what Jensen did here. In return for the guarantee, they get a revenue share above that floor. The way I think the deals will work is: we're willing to guarantee offtake — nowhere near $30 or $50 a watt, but at a very low-risk level that, with our really high-resolution view of productive capacity and future model trends, we're comfortable with. And in return, we get some revenue share above probably a higher level. Morgan Stanley wrote a great note two days ago: these revenue shares are effectively royalties, and Nvidia could very quickly become a very large cloud with a capital-light business.

Let me tell you why what Jensen did was so brilliant. The numbers are getting so big that the TAM is getting constrained by the ability to finance the buildout, and what he's doing is alleviating that financing constraint so he can grow as big as the TAM actually is. To take one example: Elon wants to add somewhere around 6 to 8 gigawatts next year. We know that would cost $300 to 400 billion of capex. The company just raised $100 billion in equity and debt offerings, so obviously they'd have to go out and finance that somehow. As we talked about in our previous episode, the simplest way would be seller financing from Nvidia, especially given that the payback period could be as quick as one year. So now Jensen is creating the line of credit — using these big banks, these big private equity shops — and making it available to all these downstream purchasers.

I don't think it's circular. This is not the circular situation Gurley is worried about. I think Wall Street banks and private equity firms are providing financing based on the expected cash flows that will be delivered from these GPUs. In some ways, Nvidia is becoming the central bank of AI, the Federal Reserve of AI — that's a way to conceptualize what they're doing. And just as the Fed's monetary policy works through private banks, here you'll have private markets charging fees and underwriting these deals, with Nvidia helping to facilitate that. So I don't think it's really circular financing.

Yeah, I watched that CNBC special as well. One point that stuck with me was when the private equity guys said this is a little like what they do for plane financing: when an airline buys airplanes, they can finance it not just on the airline's creditworthiness, but because the airplanes themselves have value. Even if the airline gets in trouble, you're protected — it's asset-backed financing, and that's what they're doing here with GPUs. For that to happen, Nvidia has to standardize — create reference designs and so on — so it can all be standardized enough for the Wall Street guys to package it up and securitize it, turn it into securities, basically like they did with mortgages.

There's a reason I'm not worried about this buildout. There's a governor, a throttle, on it — the physics of the real world that we talked about in building data centers.

There's also a dead man's switch on this — the switch you have to hold down to run the train, so if you have a heart attack your hand comes off. That switch is Anthropic and OpenAI, the two biggest customers. One of them, OpenAI, has already scaled back its buildout ambitions from around $1.4 trillion to roughly $600 billion. Then there's Anthropic, which we just said is going to hit $400 billion in revenue next year, or end the year at something like that run rate. If they don't need the compute, or compute gets so efficient, or open-source models create a headwind — whatever it is — they're going to take their hand off the buy button, and that dead man's switch will slow everything down. So we don't get a mortgage-backed-security-style collapse.

That could still lead to a car crash, but this is why I think the Anthropic IPO is really important for the market: it gets those quarterly earnings out so we can see their numbers every quarter. I think it will probably become the most important signal the entire industry has. Why? Because you have all these people saying AI is a bubble, that this capex isn't justified, that there won't be an ROI. Everyone is going to be looking to Anthropic's quarterly earnings as the pace car — not just them, it'll be them, OpenAI, and SpaceX, but you'd have to say Anthropic is the pace car right now. People will look to them to see whether the demand signal is there, and if there's a hiccup or a wobble, it'll show up there first.

If Anthropic slams on the brakes, there'll be a pileup of companies behind it, because Anthropic is making $100 billion per gigawatt of compute. That's why they're able to pay SpaceX, say, $50 billion a gigawatt — a spot price for compute. That's why SpaceX is then able to pay $30 billion to Nvidia for chips for the buildout, and why Nvidia is able to pay TSMC, Micron, SK Hynix, and all the way down. It's the entire food chain. So if there's a wobble — if somebody slams on the brakes, if the pace car slows down — the whole thing is going to feel it. But it's still better for those numbers to come out on a quarterly basis, in a vetted, GAAP-compliant way.

So Gavin, as we're saying here, there's a kill switch, a brake, whatever you want to call it. Is there enough space — enough distance between the cars — to avoid a pileup? In other words, if Dario does slam on the brakes, if there are headwinds, does everybody go slamming into them, or is there enough room for the entire industry to say, "Okay, we're going to go 20% slower, 30% slower"?

It's very simple: if he's slamming on the brakes because there's no demand, there will be a pileup. There's not enough room between the cars. But if it's because Dario is getting passed — by SpaceX, by OpenAI, or by an American open-source champion — then there won't be a pileup.

I do think the Anthropic S-1 is going to be really important, and I think it's going to break a lot of people's brains. The reason is that there are a lot of macro and value investors very confidently making prognostications about AI on the assumption that tokens are subsidized. Anthropic is generating cash. This fact is, I'd say, kind of known in the world — I've never actually spoken to Anthropic, but reading the tea leaves of what has been said in public, there are enough signals: they're generating cash, they're profitable; open-source tokens are profitable, Anthropic is profitable; OpenAI, if they're not generating cash yet, will be imminently; and the same goes for SpaceX.

When I hear these macro and value investors who are ignorant — very smart, but ignorant of the facts — they're assuming tokens are subsidized and that this is all going to collapse in some circular-financing bonfire.

They're just wrong. The overwhelming majority of tokens are profitable for everyone in the chain, and I think the Anthropic S1 is going to make that clear. When I hear these macro and value investors talk about AI, it's a bit like me saying I'm very bearish on the world economy because oil is at $500 a barrel. If oil actually were at $500 a barrel, that would be a good reason to be bearish — it's just not. The simple fact is wrong. In some ways, they're paid to be doom scrollers and merchants of doom.

Here's one framing I've been working on: the United States employs roughly 150 to 160 million people, with collective salaries of $10 to $12 trillion. There is no world in which I don't see corporations and people spending five or 10% of employee salaries on the equivalent in tokens. In other words, someone making $80,000 a year will spend $8,000 — 10% of their salary — or a minimum of $4,000 to make them more efficient. We have an analogy for this: SaaS software, computers on their desktops, mobile phones — general use of technology. Put those numbers together and you're looking at a trillion dollars in AI spend just in the United States. It's obviously going to happen. And if I'm wrong by half and it's $500 billion, that's roughly where these companies are already trending — not even counting the rest of the planet. I think we're going to be fine, folks.

Let's wrap up here. I have two more stories I want to get to, so we'll go lightning round. A very interesting story coming out...

A notable story out of New Jersey and New York involves Amazon. New Jersey's attorney general sued Amazon last week, arguing the company is essentially cheating by hiring drivers through subcontractors. This practice is well known: the Delivery Service Partner (DSP) model is a clever—perhaps too clever—way for Amazon to shield itself, according to the lawsuits now piling up. If an Amazon driver crashes, liability falls on the subcontractor, sparing Amazon costs it would face if the drivers were employees.

On Monday, New York City mayor Zohran Mamdani jumped into the fray, throwing his support behind the effort in one of his signature Democratic socialist videos. He is demanding municipal laws that would require Amazon to hire the drivers directly. As he put it:

"Corporations like Amazon build billion dollar business models by insulating themselves from accountability through a system of exploitative subcontracting."

This has been an issue for a long time, and there have tragically been a number of deaths because subcontractors are put under massive pressure, with drivers killing people while operating in the real world. Because the drivers are technically employed by the DSP, Amazon claims it is not the employer, so it is not on the hook and doesn't have to engage in union bargaining or labor obligations.

The host asked Sax for his take: given the surge in people believing in socialism and that corporations are cheating on the margins, is Amazon cheating here? Is the DSP model too clever, and should it be rolled back? The issue is gaining momentum and will be big in the midterms and in 2028.

Sax replied that the practice is widespread, not focused on New York, and that Amazon has used these networks of independent contractor drivers for a while. Amazon's argument is that denying it the ability to use independent contractors would raise delivery costs—a study showed the average household could pay hundreds of dollars extra—slow down service, put thousands of jobs at risk at small contractor businesses, and could lead Amazon to shift operations outside the city.

The DSP model also enables rapid scaling for peaks like holiday surges, dense urban routing, and lower costs—a much more flexible arrangement than if everyone were an Amazon employee. This is how the market evolved: Amazon probably did everything itself at first, then decentralized to networks of small businesses and contractors. That actually creates more jobs and more entrepreneurship; the DSPs employ thousands of people in New York City, and those jobs would be threatened if the arrangement were made illegal.

Asked where he stood, the other speaker said he would let freedom of contract and voluntary exchange prevail. The market structure evolved for a reason: the workers liked it, the DSP owners liked it, Amazon liked it—these are mutually beneficial arrangements, and he sees no reason for Mamdani to insert the city into the middle of it.

He added that if he were Amazon—though he's sure they won't do this—he would add a surcharge labeled as the "Zohran Mamdani fee" imposed by this ill-advised policy, as a line item on every bill, since that's what the policy will result in. Lots of people use Amazon, not just wealthy New Yorkers.

He continued: compared to anything government provides, Amazon's efficiency is remarkable—orders arrive the same day or, at worst, the next day, optimized for low consumer prices. Getting the government into the middle of Amazon's driver network raises the question: what city service does Mamdani provide that is remotely as efficient as what Amazon does, that qualifies him to tell Amazon how to run its business?

It's insane. Look at what Starbucks did when Schultz was running it: he gave people a job worth having and took pride in giving the lowest-paid people in the organization a living wage — somewhere around 20 to 22 dollars an hour in these cities. These DSP drivers are actually paid quite well, 18, 19, 20 dollars. But this system was not created in the free market, as you're suggesting, Sachs — it wasn't just what everybody wanted. It was created explicitly to benefit Amazon and shift the burden of these employees onto the American taxpayer, and New Jersey and New York are rightfully calling them out on it.

I'm a long-term Amazon customer and a Prime subscriber since the beginning, and it would be a de minimis cost for them to make these employees full-time. You'd still keep some DSPs to scale up and down, but this is an opportunity to make those frontline workers part of the Amazon family. It would cost about 25 cents per delivery in New York to move them from $18.90 an hour to a slightly higher wage — maybe 21 or 22 dollars — and give them benefits. This is an easy thing to do.

What our industry has a crazy blind spot for is efficiency. Amazon is the most efficient company in the world, to the point of being abusive to the most vulnerable members of our society and against the spirit of capitalism. This is a level of capitalism that needs to be self-corrected, and I think the government will have very valid arguments here. It would be amazing for Amazon to get ahead of this, because otherwise this becomes a very valid argument for the democratic socialists — socialist communists, in my mind — that will sell with Americans who believe this group of people is being taken advantage of and that we can easily afford 25 cents more per package delivery. Yeah, Sachs, you're right, it might add $100.

Well, what if we have Uber do that too, then?

Absolutely. Uber was basically castigated for the exact same thing — using independent contractors instead of making them full-fledged employees.

That's actually a great point about Uber: this would be the equivalent of Uber having three or four thousand people representing their drivers with no exposure. Uber actually has massive insurance exposure and massive lawsuits to contend with for what those drivers do. So it's distinctly different, and they've gone through this.

I'm not a proponent of socialism, but I do think there's an opportunity for Amazon — which is going to be automating along the way — to take a little haircut on their amazing business and be on the right side of history, instead of leaving their employees' benefits, unemployment, food stamps, and health insurance to the public to provide. These socialists are going to point this out in the next two election cycles to the point of absurdity, and it's going to play. Socialism is playing; it's making inroads and infecting people's brains. It's up to capitalists to say, you know what, we don't need to be as cutthroat and as clever.

That's my message as a long-term shareholder — I have a very big position in Amazon and I keep increasing it every year. They should do the right thing here and get ahead of this, so they can say, "We are an incredible employer," instead of, "We really are trying to hide and move these expenses around." That's my only point. Not a socialist, not a cuck, not interested in messing with capitalism — I just think they got a little too clever here, and they should treat the most vulnerable employees in the lowest docket right.

Now I know why you put this on the docket — so you can get on your soapbox.

No, I put it on the docket because it's one of the top news stories, and because we talk about socialism every week and all the virtues of capitalism — and there are downsides to capitalism, where we can get too clever.

I think we spent too much time on this, but let me ask you a question. You said you want Amazon to employ all these drivers —

I think they should employ —

I know you said you wanted a bunch to still remain with the DSPs.

I think the DSPs can happen.

So how do you decide which is which?

I think they could proactively do that. It's a great question. Andy Jassy could proactively decide, just like Schultz did, to get rid of the DSPs, and I think there's a compromise here: keep DSPs to scale up when there's demand during the holidays. But he should get ahead of it.

As soon as you acknowledge that the DSPs are better because there are more flexible working arrangements that people want, why are you meddling in this arrangement to begin with?

That's actually not true. The people who work in the DSPs—the DSPs themselves, the owners of those fragmented, very small companies that go out of business the second they get in an accident, to shield Amazon—those employees would very much like to have Amazon equity. They would love to get RSUs and they would love to get healthcare. Some of those drivers also work for Uber a lot of the time and toggle back and forth between being a driver for Amazon and a driver for Uber. They like the flexible working arrangement.

You're actually misinformed here. These are people working full shifts—8 to 12 hour shifts; I think it's actually 10 to 12 hour shifts.

I don't think you can speak for all these workers. They can speak for themselves, and they are organizing, and Amazon has been trying to break the organizing efforts when they try to unionize. Forcing them to be part of the DSP is a tool for breaking that—that's the game on the field, Sacks.

Well, you already acknowledged that you think the DSPs should stick around.

I think a portion of them should, and I think Andy Jassy and the team at Amazon should get ahead of this, because it's a bad look. It's not in the best interest of capitalism to be this cutthroat. He should do it proactively, not be forced to—that's my personal belief. The same way Schultz said there's an opportunity to make baristas have a living wage and benefits, and it was one of the best things Starbucks ever did. It's really what capitalism should start thinking about: what is our role in driving the move to socialism amongst young people? This is an example of one of the reasons people are starting to believe in socialism—they believe companies like Amazon are cheating a little bit on the margins, maybe trying to take expenses they should incur and put them on the American taxpayer. And you know what? Mamdani and New Jersey are going to win these lawsuits. I predict it now.

Well, look, JCal, that all sounds really nice, and I want to save capitalism as much as you do. If Amazon doing that would really save capitalism, then I guess I'm game for it. But I'm not sure it will. What we do know it will do is raise the cost of packages by about $5.20 per package, cost families $664 more per year, and put something like 5,000 jobs at risk. There are real costs to doing things the way you're saying. When people wear their hat as voters, yeah, they fall for the socialism stuff, but when they're wearing their consumer hat, they just want to pay less for packages. They will hate the fact that packages cost $5 more. And like Gavin said, they should put the Mamdani surcharge on their orders—and it's not 25, it's $5. $5.20 per package, $664 per household per year in New York. That's crazy. That's a real cost. Talk about a cost of living adjustment—that's a huge one. JCal has had internet issues, so we're just moving on. Gavin?

Gavin was asked what he thought of the xAI releases this week — Grokbot and Grok 4.6 — with people saying the model is close to or at the frontier but much cheaper. He had maintained that the industry was down to a frontier lab duopoly of Anthropic and OpenAI, and was asked whether that was wrong.

He thinks it was. A graph of the price-adjusted frontier shows quality or intelligence on the y-axis and cost on the x-axis, with the axes inverted so the upper right is best. Grok 4.6's pricing is disruptive on that chart. The data comes from Cursor Bench — and since xAI (SpaceX) has a right to acquire Cursor and works with them, that could be grading your own homework. But there's also a chart from Databricks, which just raised money at a $190 billion valuation and is a very sophisticated company, showing Grok 4.6 ahead of Fable 5 — the gold standard for most people — at higher quality and a slightly lower price.

The numbers are the numbers, though the vibes still need to play out. DHH (David Heinemeier Hansson of Ruby on Rails, Basecamp, 37signals) posted something really positive, and the Databricks and Mercor evaluations look good. The early vibes in these benchmark scores are bearing out in the real world.

In other words, Elon has caught up, for two reasons: he bought Cursor and got that dialed-in team, and he reshuffled xAI's management, bringing in his "killers" from SpaceX along with the Cursor people and some new hires — all in six months. Never count Elon out: he's frequently late, never wrong, and tends to over-deliver when it comes, by his own admission.

This is still a relatively small model — a 1.5 trillion parameter model — and a larger one, Grok 4.7, should be significantly more capable and is coming in a few weeks. Grokbot seems just as important: it feels like another OpenClaw moment for AI — personalizing, democratizing AI — the same thing Zuckerberg described in his manifesto, and what Claude Co-work, Claude Tag, and Hermes' Perplexia Computer do well.

One speaker admitted that some of those tools feel like boxes to him despite trying to be technical, while Grokbot feels much easier to use — OpenClaw and Hermes require real expertise. (A running joke: "Hermes" got confused with the luxury brand — "buying a scarf" — like the channel where your wives occasionally order Chanel.)

Catching up in six months is remarkable, but it creates a tension: Elon is hosting Anthropic on his cloud — "Elon Web Services" — at the same time as competing with them increasingly directly.

What I love about it is that he's creating these huge compute clusters, which gives him the ability to train a big, potentially frontier model. You can think of that as the call option: catching up and becoming a true frontier lab. But if he fails at that, he has the put option of just selling the compute at pretty attractive prices to the frontier labs.

And he has the right to take that compute back from Anthropic. According to reports, he put that clause into the Anthropic deal so he could reclaim compute if he wants to.

Right. Both sides have a 90-day cancellation, which means if the spot price of compute goes up, he benefits. If it goes down, it could hurt him, because either side could renegotiate within 90 days. Basically, Elon is able to charge spot minus 90.

That's the risk, but it all goes back to what we were discussing before: do you think compute will be supply-constrained going into next year and the year after? If so, that's a very advantageous dynamic for Elon, because he's very close to charging true spot for his compute, but he can also essentially reroute it to his own service if he's able to take the lead.

Can I say what's interesting? Sorry to interrupt. I talk to a lot of investors about SpaceX, and what amazes me is that everyone will have a nuanced view on Starlink—maybe they're a skeptic on Starlink direct-to-cell, though I think the plan to use these distributed Starlink satellites effectively as small cells is interesting. Everybody will have a nuanced view on orbital compute, whether bullish or bearish, and they'll have a view on terrestrial compute.

But very few of these investors talk about Grok at all, and it is dominant on a lot of measures. We have an existence proof that Anthropic went from a billion to 50 billion really fast, and if Anthropic is worth two trillion per the FT—maybe three or four trillion—

I said I thought that was a market-clearing price. I didn't say that's where I thought it was worth. Maybe people should start to consider Grok when they think about SpaceX.

SpaceX has so many ways to win, and I think that's the higher bit here. And we're obviously all talking our book—the folks on this podcast are amongst the earliest investors in SpaceX.

I was part of leading a small part of leading the second round of 10 billion.

One thing broke during the show that we should talk about: Silver Lake may be buying Workday, and the stock is up 17%. If true, this is a significant moment for the stock market. Software has been in a death spiral—it probably troughed three or four months ago, but it was in a tough spot for 12, 15, 18 months—and the return of a private equity bid for software really changes the investing landscape.

I don't know what the potential buyers of Workday are thinking, but the rise of open source and its increasing competitiveness is a godsend for the American software industry—an absolute godsend. If you pull up a five-year chart of Workday, you can just see it: in a world with one, two, or three dominant frontier models, the stock went from around 300 straight down to 100 in a really short period of time. Even with this bid, it's only recovered a little bit. A world where Anthropic, OpenAI, and maybe even SpaceX are the only model providers dominating the model layer is a hard world for software. If I had to guess why the private equity bid for software is coming back, it's because they see what's happening with open source.

That's the backstop, isn't it, Sacks? If nobody in the public markets wants to own these companies and private equity sees a way in—we talked about Bending Spoons a whole bunch last week—there are now, I don't want to call them bottom feeders, but sharp acquirers of money-printing businesses who want to extract even more revenue from them. It's a good turn of events that there's a buyer for these companies.

Absolutely. What Silver Lake is saying is that these companies have been oversold. There must be some value there. And if you use AI to run them—I think that's what Bending Spoons is doing really well—they get rid of 80% of the employees, make the remaining 20% AI-first, and figure out how to get all of this incredible revenue. There's your breaking news.

Just so you know, I looked at the metrics before we got on today: you're now tied with Brad Gerstner in the ratings, neck and neck for fifth bestie—which means one of you will drop to sixth if this continues. Brad may be on next week, so we'll see who takes the fifth slot. David Sacks, great to see you. Maybe we can get a steak later—I've got Thursday night off.

This has been an amazing episode of the All-In podcast for August 14th. We'll see you next time.