DeepSeek、开源、关税、DOGE与市场影响｜BG2 对话 Bill Gurley 与 Brad Gerstner

I think you make a good point and a good defense for globalism, but the exact response would be: If you’re the president of the United States, you’re not looking out for the standard of living for all humans or for all people around the globe. You’re looking out for the standard of living of people in the United States, to whom you have a constitutional oath.

No doubt.

Hey, Bill. It’s great to see you.

Good to see you, sir. Are you going to apply to run the sovereign wealth fund? Actually, better yet, will you give me permission to nominate you to run the sovereign wealth fund?

I would think there are people I’ve met through my days who are LPs at rather large funds who have way better experience for something like that.

I think you would be a hell of a choice, certainly maybe on the board of the sovereign wealth fund. It’s going to be a fascinating experiment.

One observation I had as we were getting ready for the podcast today is that last night I was helping Lincoln study for his AP History exam. He’s studying the Gilded Age and, more importantly, the McKinley presidency. It’s all about the McKinley tariffs. I said to him, “It’s pretty amazing that you and I are working on the same thing. I’m reading on ChatGPT about the McKinley tariffs.” He said, “Yeah, but, Dad, don’t worry about it. You don’t have to take the test.” I said, “No, worse yet, I have to figure out how much money to be exposed to the market and what our risk is going to be in the hedge fund.” I said, “Don’t worry. I get a scorecard on that, too.”

A little weightier than the test.

Yeah, no doubt. You and I could go full Lex Fridman today. By the way, his 5-hour podcast recently with Dylan Patel was great, but we’re going to try to keep this pretty tight and jam quickly on DeepSeek, tariffs, DOGE, and maybe what the market reaction might be to all of these things.

As I was thinking about this setup, Bill, I thought it would be helpful. You and I got into this because we said we wanted the podcast to be about the intersection of tech and markets, investing and capitalism. We specifically wanted to stay away from Washington and politics, but it’s really impossible at the moment to talk about capitalism and markets and what’s happening without talking about the big things occurring in Washington.

We’re going to do that, but we’re really going to try to stick to the economic and market lens of the events that are happening rather than the political lens. There are incredible podcasts you can listen to that give you the political analysis of all of this. Maybe we just dive in.

That would be great. When you expand the lens to include that, and you look at the ridiculous pace of innovation in the AI space, I don’t ever recall a single time in my career where it feels like you could just have your ear to the ground 24/7 and you’re picking up something new constantly.

1. DeepSeek & Open Source

No doubt about it. Speaking of that, we covered DeepSeek last week, Bill. It seems like it was just yesterday, but a lot has happened since then. For one, these usage charts for DeepSeek are really quite amazing. Andrej Karpathy tweeted this, showing the percentage of daily active users relative to ChatGPT. They top the geographies—I think they have something like 15% in India, 10% in China, and 8% in Indonesia. It’s really a global phenomenon.

One of the things that struck me after all of this is that you tweeted, “It’s a better world if the most disruptive LLM model is foreign and open source versus domestic and proprietary,” right? As you said, it’s better for safety, security, free speech, and so on.

Talk us through where you think we stand today, with a little space now to reflect on R1. We know that you and Benchmark have been the staunchest proponents of open source over the course of the last few decades. Help us understand how you think open source versus closed source is going to evolve.

Let me give you some reflections now that this has been a week in the rearview mirror, and I’ve been reading and watching as much as I can of other people talking about it.

Here are some things I think we know about DeepSeek R1. It was quite innovative. If you have 5 hours to listen to Dylan Patel and Nathan Lambert on Lex Fridman, they get into some of this. DeepSeek had put more experts simultaneously against a problem. They were able to do that because they figured out a way to separate the parameters and work on things with smaller parameter counts faster, which no one else had done before. You end up with something that’s cheaper and faster. It’s important to recognize that they did innovate versus just copy, because I think there’s a lot of noise around this whole thing.

They also chose to be the most open model we know of today. There are a lot of people who like to get into nuanced conversations about whether it’s truly open source. I guess people are hoping for one day when someone shows all the data and all the training processes, so they may have fallen short of that. But with the MIT license, which has no restriction whatsoever on how you do this, and open weights, it gives a lot of freedom to a lot of people to take this thing and run in opposite directions.

The thing you mentioned is also worth noting: The success was a breakthrough. Maybe not compared to OpenAI, but certainly I think Mistral, Anthropic, or any of these players would have loved to have had the launch moment, if you will, that DeepSeek had, and weren’t able to achieve it for whatever reasons. I don’t know if we can fully explain why this app is still number 1 in the App Store today.

The thing you mentioned that I think is super interesting is the rest of the world. We’re going to go into how this might play into China, U.S. sanctions and restrictions, and the rest of the world could be up for grabs.

A couple of other points are worth mentioning. It’s now validated that they went around CUDA, and I just think that’s interesting. It’s interesting to think about why, and to think about performance optimization and how you might get down to the bare metal. When we went from on-premises software to SaaS, places like AWS started pulling out as many layers as they could to get to optimization. You ended up with very special versions of Linux where things were just ripped out and ripped out and ripped out. I think that’s worth noting. I know of at least another example where someone working on inference optimization went underneath as well.

The last point I would make, which I think is worth understanding, is that you and I stumbled upon, with the help of Sunny, an interview in Chinese with the founder. We translated and read it, and I think it’s impossible to read that and say this isn’t an exceptional founder. He’s intelligent and independently minded. I think it would be very hard to make an argument that he’s not a remarkable founder.

Of course, you’re referring to someone likely named Liang Wenfeng, the founder of DeepSeek. We’ll put the link to that interview in the podcast. You can copy it and put it into ChatGPT or whatever and get a quick translation.

As you know, our team knows him quite well and knows a lot of members of the DeepSeek team. He’s become really a national hero in China. There’s a little bit more of a microscope on him today, but we did learn some things when we talked to the DeepSeek team, and a few of those are pretty salient.

Number 1, he’s an incredible founder. There’s no doubt about that. He’s been working on this problem for upwards of a decade and has been thinking about it for a long time. He’s very successful and has made a lot of money in the hedge fund business. It reminds me a little bit of the Jim Simons story at Renaissance. These are brilliant people who happen to apply this early AI and deep-learning edge to the hedge fund business and quantitative trading.

A couple of the key things that were debated last week: One was the total compute capex. When you take power into account, it really does get you closer to the $1 billion of total cost of ownership that was discussed, which is similar, I think, to the TCO of the comparative models.

Of course, they talked about the $6 million final training run, and it’s important to understand that this is also correct. As I said on CNBC, this compares to about $10 million or $15 million for o1 from OpenAI. Apples to apples, they were 30% to 50% more efficient. To your earlier point, that was due to real algorithmic breakthroughs.

Of course, they were doing this a few months later—maybe 6 to 8 months later—than what was going on at OpenAI, so you would expect some of those savings. But take nothing away from them. A lot of them came from algorithmic improvements, many of which I think are going to be copied, but they were breakthroughs nonetheless.

I would add one thing that’s worth paying attention to. Dylan and Nathan went into this on the Lex podcast, but they believe—and I can’t really defend it—that if you look at models that are apples to apples on the API right now, DeepSeek’s pricing is about 1/12th of OpenAI’s. They argue about whether OpenAI might just have much higher margins or whether DeepSeek might be subsidizing, but that differential is bigger than the one you described for training.

You’re referencing what they’re charging the customer for inference. Remember, they could run it well below cost, or they could choose to be running at well below cost in order to have these outcomes. I don’t think that’s going to last. OpenAI might be charging well above cost for that.

Lyft subsidized rides well below cost for a decade.

You could do it for a very long time. But let’s get back to the compute stock, because this was something we learned from their team that I think is really important to understand.

Their compute stack was smaller than the compute stack OpenAI used to train o1, but it wasn’t that much smaller. The problem they now face is that this changes dramatically. Remember, these are log-linear scaling functions. In order to get to the o3 level, they now have to 10x the amount of compute, assuming they don’t come through with some massive architectural improvement that allows them to do log-linear scaling without more compute.

To get to that next step function, they acknowledge it’s going to be a lot harder. Said another way, this was the moment in time when their compute comparison to OpenAI was the closest it’s ever going to be, for a few reasons.

Number 1, with the export controls, they acknowledge they’re going to have a very hard time keeping up with o3, and Stargate is going to be even more challenging. They’re not going to have access to Blackwell, which is a 2x to 3x improvement on top of the Hopper series, which already exists. This differential in GPUs, as they begin to train o3, is far greater than it was for o1.

On top of that, we learned they are massively compute-constrained right now. You’ve seen some tweets about this—people getting server delays and all that stuff with these guys—because of their massive success.

They have a limited cluster to begin with, and they’re currently taking most of that compute and deploying it against inference just to support the demand coming in the door. That further constrains their ability to do training, and we know they have to 10x the training to get to that next function.

This is a really tough situation. On o1, they already had somewhere in the order of 30,000 to 50,000 GPUs that they had previously purchased. There’s a lot of debate about exactly how many they had, but the differential wasn’t that great. Now, when you have to step up to 10x that, it becomes very challenging.

Those are things that we think we’ve confirmed directly from the company. One of the things that would impress me even more, Bill—I asked the team at OpenAI what surprised them. Did DeepSeek surprise them? They said the only thing that surprised them was that it was a Chinese company that was able to get there before Meta and others.

I think it’s a really important question. The thing that would impress me even more is if somehow they figure out an architectural or algorithmic way to catch up with o3, Deep Research, and the stuff that’s now truly frontier without having access to GPUs. That would be the definitive statement that they have somehow broken the paradigm on cost and scaling.

We wouldn’t be the first people to say it. Even Fareed Zakaria said it on his show on national television. Everyone is talking about the fact that constraints can lead to innovation. The reason Llama probably didn’t do it is that they had access to scaling, and only if you’re limited on that function might you make an algorithmic change.

My guess is—and Nathan Lambert might disagree with me about the 10x requirement—that because it’s open and they publish the paper, I’m 100% certain that people at Anthropic and OpenAI are studying what DeepSeek did. If there was some innovative breakthrough, the borrowing is going to be bidirectional. It’s going to go right back.

That gets you into Jevons’s paradox, which everyone else is also talking about: If we make this stuff cheaper, won’t people just buy more and more?

But let me go back to your original question on open source, because I never really got around to answering it. There are some things worth mentioning. As you said, I tweeted that I think a lot of people were worried about these models and whether they control what people say, disinformation, what’s embedded in them, and what they do.

I’m a big believer—and many people in academia are as well—that more transparency leads to more understanding, more safety, more security, and more free speech. I’m not the only one who has said this. If you had a singular, proprietary model from a singular company, it would give you more risk on all those fronts and more ability for someone to control that kind of thing.

Let’s be very clear: The very reason OpenAI exists—I remember when Elon first talked about it on stage at launch—was to defend against singular control by Google of a closed, tyrannical AI. I put all of those in one group.

Then I said it’s also better for innovation, startups, cost performance, and global prosperity. I’ll give you a data point. I was talking with Clément Delangue over at Hugging Face. About 48 hours after R1 was posted, they had 500 variants on Hugging Face. Today I pinged him this morning before we started, and they’re up to 1,300.

These are just forks in different directions. It allows people to do massive optimization and solve any problem you may have with R1. Someone can go work on it. It’s Linus’s law from the original “The Cathedral and the Bazaar” paper: Given enough eyes, all bugs are shallow. The R&D force becomes the world, not just an individual player.

It allows for so much optimization. It also makes enterprise companies happy. Aaron Levie and Marc Benioff were out there very boldly supporting this R1 breakthrough, and it just makes sense. If there’s a piece of technology that’s a commodity they need to be successful, they’re better off than if there’s some proprietary piece they have to license from someone.

It was quickly deployed in all the clouds—in AWS and in Azure—which was shocking to me, especially the Microsoft deployment. Obviously, that feels like a piece of the strategic back-and-forth between OpenAI and Microsoft around their contract, but everybody went up fast. It just shows you what’s possible.

I think the amount of innovation you can have isn’t just up the stack. People say, “That’s going to allow people to build special models,” and all this stuff, but it’s down the stack as well.

If you’re trying to compete with NVIDIA with a TPU or a non-GPU—we’ve talked about all these companies before—or if you’re an investor in a company like Fireworks AI, which is trying to be this high-performance inference optimization middle layer, knowing more about the model allows them to optimize even more. They have a lot of amazing customers doing runtime inference in production right now.

I think everybody else, other than the big proprietary models, is probably thrilled to have this type of product out there. Once again, the variants will just go, everyone will borrow from it, and—

Can I just pause for a second?

Sure.

A year ago, there was a lot of excitement about Llama, and Zuckerberg really took the leadership on open source in the U.S. Here we are evangelizing about open source, but it’s not about Llama. It’s about DeepSeek and R1. What do you think happened there? Do you think it’s just scarcity and the leapfrogging?

Obviously, there are lots of reports that there was a lot of trauma within the Meta complex last week—people very upset about the fact that they were leapfrogged here, the amount of money they’re spending, and that they didn’t get there first. But do you have any speculation as to why it wasn’t them?

I’ve watched different open-source battles in a whole bunch of different verticals since my firm was an original investor in Red Hat, I think in 1999—25 years ago.

There’s always a continuum of openness, and there are a whole bunch of licenses. I tweeted this list of a hierarchy, or continuum, of licenses. They run from most open to least open.

Almost every company that tries to play in the open-source area is playing this weird game where they want the proliferation of openness, but they want some kind of hook to be able to claw back and have a proprietary advantage.

We know that Meta had not gone fully open. They had never published the weights, and there was this clause that said if you use it too much, you have to come see us again. That clause got spread around last week.

Amazon actually has to pay Meta for the use of Llama.

They were playing the same game that all these companies have played. They’ve all had to play this game. It looks like these guys just decided to be more open, and the MIT license is pretty against the grain. As I said, there are people who say they could be even more open, but this is the most open for sure today.

China is not a newcomer to open source. If you look at all of the major projects, like Linux or MySQL, most of these open-source projects have a website where you can see who the leading donors are. Go to the Linux one—I’ll put a link in here from the Linux Foundation member group—and you’ll see a ton of Chinese companies.

Someone may ask, “Why is China pro-open source?” For the past 30 years, the West has done nothing but accuse them of being intellectual-property thieves. If you believe you have the fastest, cheapest, most capable entrepreneurs or engineers who can run faster and work harder than everyone else, you’d rather live in a world where there’s no IP protection than one where you’re just being held back.

I think they jumped into open source full throttle. It’s not just Linux and it’s not just this. If you look at RISC-V, they’re one of the biggest supporters of RISC-V. Every time we put more constraints on what they can get access to, they invest more in RISC-V.

This is particularly important relative to the rest of the world, as we brought up earlier. We’ll get into sanctions and whatnot, but if you pull the wall up and we don’t support open source while they do, and everybody else likes them leading the way, that could be a dangerous situation.

The last point I want to make is about Fareed Zakaria. I’m a big fan of his. He had 2 takeaways on DeepSeek, and I was impressed that he landed on both of them.

One was that there was a lot of discussion, especially in Washington, that the U.S. was 2 years ahead of China. He said, “It looks like, after the fact, that that’s hubris. If that’s now 6 months, 3 months, whatever, it’s closing, and we need to think with our eyes wide open as we make policy decisions.” I think that’s important.

The second one was that I was impressed with his understanding of open versus closed and how you can reach a tipping point where things just move in that direction because so many different entities get behind it. I like to see water run downhill.

Sam Altman did this AMA last week and was asked about DeepSeek and open source. I thought his response was really interesting. He said, “DeepSeek is an impressive model,” and, to your point about Jevons’s paradox, “We’re going to need a lot more compute,” because demand for this is exploding and they’re compute-constrained.

On the issue of open source, they asked, “Would you consider releasing model weights and publishing your open-source research?” Sam said, “Yes, we are discussing it. I personally think we have been on the wrong side of history here, and we need to figure out a different open-source strategy.”

I tweeted in response to something Mark Andreessen had said that I thought all current closed-source model companies—let’s just say OpenAI and Anthropic—would open source their models. In the case of OpenAI, I could see them open sourcing o1, which competes head-to-head with DeepSeek.

At the same time, I could see all companies that are currently open source, including DeepSeek, closing source in the future. Mark Zuckerberg has said he reserves the right not to release all the models in the future.

I think we may end up with a world where the true frontier—the actual underlying model—is not released at all, and the only thing that gets released is the agent. But then, 1 or 2 generations behind, you’re going to see them all open source these models.

Let’s start with Sam’s comments. Are you encouraged that Sam has said, “Yes, we need to come out and say we’re on the wrong side of history here”?

To a certain extent, it validates what R1 did—that he would feel the need to say that. I’ve found, and you talk to him more than I do, that whenever a threat or a challenge is made to OpenAI, Sam tends to go toward it. That’s his go-to move, and I think it works for him. I’m not surprised that he said that.

I was surprised, on a side note, that a couple of our friends who are co-investors in OpenAI, when the R1 thing hit, very quickly took to X to say that DeepSeek cheated or that the government needed to come in. To me, that was a validation point as well. You wouldn’t take the trouble if this thing wasn’t real.

It could tip us more toward open source. As someone who really enjoyed my business-school classes on finance and economics, one of the reasons I like open source so much is that it’s the closest thing to pure competition.

If you look up pure competition in an economics book, it’s like a commodity: It’s hard to have pricing power, and it leads to innovation and low price points. Jevons’s paradox blows up. I’m thrilled that it’s tilting in that direction.

I like to see water run downhill.

Sam Altman did say they’re going to need a lot more compute. On the issue of open source, I think it’s notable that OpenAI is now saying that it may have been on the wrong side of history. Let’s transition into the other thing that happened as a result of R1, which is the increased effort to raise the wall of regulation and sanctions.

It’s funny watching a number of people say, “Oh my God, look at R1. Washington must act quickly,” while the thing each person intended was radically different from one another. Some people think this means we need to embrace open source and encourage more open innovation in America. Other people think, “Oh my God,” and Dario Amodei put out a long piece, consistent with his entire tenure here, begging for more lockdown and regulation.

Let’s go into that, because I do think the industry response seems to be tipping more in the direction that we’re going to go open, too. My suspicion is that you will see that this year out of folks like OpenAI.

The discussion about DeepSeek clearly touched a national nerve. Jensen Huang got called to the White House last Friday. We all saw it. They said it was prescheduled. I’m not sure whether or not that was the case, but let’s assume it was.

It’s opened up this broader conversation about whether the U.S. can—or even whether it’s wise for the U.S. to try to—stop China from advancing along the AI frontier.

Some of the arguments are that human talent in China will always find a way to innovate; keeping China 6 months behind is not worth the cost; scarcity fuels innovation and turns AI into a global arms race; or, as I’ve been arguing, we’ve focused so much on slowing China down that we haven’t focused enough on speeding America up.

We need to remove the regulations around power generation and do all the things necessary to get America running at full speed.

This is probably the biggest divide in Silicon Valley among technologists regarding this president’s policies. There’s a camp of China hawks led by people like Alex Karp, who was on the Palantir call last night, who view this as an existential holy war and believe we must battle on every front in order to slow China down.

Then there are people who I would consider China constructivists. I would put you in that camp. I’d put myself in that camp. Frankly, I’d put Elon in that camp, along with others who seem to think it’s a losing battle just to focus on slowing China down. What we really need to focus on is more engagement and speeding the U.S. up.

Can you lay out your views on those 2 competing sides and where we may end up?

Brad, I think you framed it perfectly. By the way, on the anti-China side in Silicon Valley, you have 3 groups.

You have people like Dario, who may be worried about competition, maybe worried about more, but who certainly question the threat. You have the new venture-capital-backed defense companies. I might put Palantir in that group, but they all have an incentive to have tension with China, because it actually increases revenue. I call them the new neocons.

Then I think you just have a large group of people who were raised to be anti-China. It’s what they were taught growing up. It’s the anti-communist thing. It’s what got us into the Vietnam War. It’s been around forever. Your parents might have taught you that; it’s just in the ethos.

One thing I would add on the risk side is that protection of U.S. companies causes harm. Detroit is not globally competitive anymore, and putting tariffs on BYD’s cars is not going to make Detroit more competitive. It’s going to make them less competitive, and they’re going to fall further behind.

I think this idea of raising the wall and increasing decoupling is a super-dangerous idea. We may find that the rest of the world is perfectly fine buying $10,000 BYD cars and using DeepSeek, while we may simply be shutting ourselves off.

To highlight my fascination with China, I’ve studied it over a very long time frame. It turns out that most people have no reason to know this, but in 1820 China’s economy was wide open, and it actually had 33% of the global economy. Thirty-three percent of global GDP was China. Most people probably wouldn’t know that.

The reason they might not know it is that 150 years later, at the end of Mao’s reign, Mao had raised the wall and turned China inward. China had fallen to 5% of global GDP. They’ve been working their way back from that, so we know the emerging China from that place.

That’s the real risk. If you’re not a globalist, if you don’t believe in all the great economic work that shows how specialization can work to the benefit of everyone, and you close that wall, you may be surprised at what happens.

To wrap this section, that’s the perfect segue to talk about tariffs.

I would say there’s a middle ground. When I first read your tweet that it’s better for an open-source Chinese model to win versus a—

Foreign model.

Foreign model. I knew what the hell you meant. I want Team America to win on this. I would love to see the U.S. frontier labs open source more stuff. I agree with you fundamentally on the principles of open source. I believe they will, and unquestionably I want to see the U.S. win when it comes to the race in AI. I know you do, too.

I would also say that a lot of the things we’ve done in the name of being tough on China are actually counterproductive. They take the eye off the prize, slow us down, and don’t focus on speeding us up. Frankly, they’re not very effective, or they backfire entirely in terms of slowing China down.

One of the places where I think this takes us, Bill, is to Trump’s tariffs, which brought us to the fore this week.

I want to say 2 things about what you said, and then we’ll go to the tariffs.

I was being provocative when I said “foreign,” and you could read that as China. But think about it: Linux doesn’t have a geography. One possible reality that may not meet your goal of America winning is that you get to a place that’s Linux-like, where the model doesn’t have a sovereignty.

Because of the 1,300 variants of R1 already on Hugging Face and because of that MIT license, it could be a model that wins that doesn’t have a geography. It doesn’t have to be the one DeepSeek is developing; it could be a fork of it.

On the risk side, there are proposals in Congress right now that would kill open source and say we couldn’t use variants of R1. There’s a huge breadth of perspectives, as you already said.

I can imagine that if you just poke China enough—if you keep poking, keep raising the constraints—you increase the odds that they make a run at Taiwan. I think it’s important to always think from their perspective. We need to be careful about how hard we push. We may end up with the exact worst outcome through unintended consequences.

This week we woke up—or really ended Friday—with a late press conference the president held, and then it went into effect over the weekend: 25% tariffs on Mexico and Canada and a 10% to 15% tariff on China.

2. Trump Tariffs

Before we dive into the economic debates for and against tariffs, let’s lay this out. On Monday morning, with the markets falling overnight on Sunday, Kevin Hassett, chairman of the National Economic Council, came out on the White House lawn and said these were all being misinterpreted. This wasn’t about a trade war; it was a drug war. It was about fentanyl.

He did happen to say that they may revisit tariffs in the future as part of a tax-reform strategy, so he left the opening for tariffs for other reasons. Then the president spoke to both sides. They each committed 10,000 troops to the border to fight fentanyl, and the tariffs were delayed for 30 days.

The market’s reaction now is that they’re not going to hit at all. Let’s start with what we think happens here. Help me predict what you think happens, and then I’d love to get into the merits and demerits from an economic perspective and maybe from the tech industry perspective on this tariff strategy.

I’m going to be brief because this is more your world than mine. You’re looking at a lot of large public companies and all the things that impact them, even in the medium-sized public companies.

Especially if you have physical goods, you have supply chains all around the globe. I imagine one thing you’ve had to do in your shop when a new tariff pops up is immediately ask, “Who’s impacted? Who sources there? Who has what?” It probably just creates a lot of chaos in the short term as we try to figure those things out.

There are different Foxconn plants all around the globe, and different people source different products from Vietnam, Indonesia, China, and elsewhere. During COVID, I think one of the things we realized is that there is some flexibility. They can move a lot faster than people thought, but it’s still chaotic.

As I said, my default is globalist. I don’t know enough to know whether we have unfair deals that need to be honed and whether this is just a means to an end. If so, maybe it’s not that big a deal.

I don’t believe that creating a lot of barriers and bringing the wall up around AI will be in the U.S.’s long-term best interest.

You make a good point about the fact that Trump clearly extracted a concession from Canada and Mexico when it comes to defending the border. As a tactical negotiating tool, his batting average is exceptionally high, whether it’s getting Colombia to take detained deportees when he threatened a tariff or other examples.

I think that’s the market’s reflexive belief. Scott Bessent, the Treasury secretary, used this concept in a letter he wrote to his investors about a year ago. He said you shouldn’t be afraid of Trump’s tariffs because his strategy was to have a fully loaded gun but rarely discharge it.

The interpretation is that he’s just using this as a big stick to achieve very tactical goals. I call this the Bessent consensus. I believe this is the market consensus view.

But I want to throw out an alternative view. Trump may in fact have a much deeper and more principled belief in tariffs. If you listen to his speeches and read what he’s said, this goes back over a decade. He believes that McKinley was one of the best presidents. He believes that the country was at its peak, or at its best, during peak tariffs in the 1880s.

You could, in fact, replace income taxes with tariffs. Prior to 1910, which is when we got the income tax, the vast majority of revenue to the U.S. government came from tariffs. Then, starting in 1910, revenue from tariffs plummeted, while the amount of revenue that came from income tax and Social Security tax skyrocketed.

I think there’s a belief that replacing tariffs with high income taxes and corporate taxes has gutted the middle class. Not only did it destroy jobs in America, but it also burdened people with taxes to pay for social services that could otherwise have been paid for by tariffs.

That’s a variant view. If you believe that, there’s a much more principled architecture that he wants to move toward. This is not the Bessent consensus—a fully loaded gun that’s rarely discharged—but a fully loaded gun that you fully intend to discharge.

Let’s assume for the moment that he does have that view, perhaps more of a Fortress America view. I think you’ve outlined where you stand on this. You believe that hurts us technologically and hurts us in terms of our global economic standing. Is that right?

I’m not a tariff expert, but when you told me you wanted to talk about this, I did some research. Maybe we could have your son or your son’s professor on to talk more about it.

All of the success stories around tariffs, if you ask your favorite AI to tell you about them, seem very short-windowed. Even the McKinley tariffs were only in place for about 4 years. I don’t know of a long-term successful high-tariff program.

It gets back to what I was saying about China. Pulling up the wall is not something I think works to help a country in the long run.

I do want to take a brief second to talk about what you said about the American middle class. There’s a time and place when a country is in a great position to be competitive globally in scaling out production. It relates to having an educated workforce with a very low standard of living that’s willing to work for a wage that’s highly competitive globally and may be willing to work 996—way more hours than 9 in the morning to 9 at night, 6 days a week.

If you look at when America was mostly successful at scaling out, it was after World War II, when Europe and Japan had been decimated. We had a lot of people moving up the social and prosperity ladder as a result of being willing to do that.

If you fast-forward to where we are today, I don’t think there’s any way to say this other than bluntly: There are people in China, Vietnam, Indonesia, and Mexico who are willing to work harder and longer for a wage that is radically lower than what people in the U.S. are willing to work for. They’re going to move from a place on the prosperity and social ladder that’s low to a place that’s still beneath the average American.

I don’t know how, as a humanist, you can say they don’t deserve that. I think we misinterpret this as somehow being a result of tariffs or trade. It’s just global fairness. That’s my point of view.

No, totally. I think you make a good point and a good defense for globalism, but the exact response would be: If you’re the president of the United States, you’re not looking out for the standard of living for all humans or for all people around the globe. You’re looking out for the standard of living of people in the United States, to whom you have a constitutional oath.

No doubt.

But even with that lens, you have to think about the dynamics around the globe. Pulling up the wall and trying to get people to make a $40 microwave in America is going to fail. You’re just going to end up with more expensive products. We watched this happen in Europe. We watched it play out there.

I do think there are some great economic arguments on this. Kevin Hassett, chairman of the National Economic Council, is at the Hoover Institution, and he’s going to be on the front lines of carrying and defending the tariff policy.

If you look at the McKinley tariffs, they certainly caused a lot of strife, but there are very good arguments that they helped us industrialize in a way we never would have. More importantly, they helped us build critical strength heading into World War I. Had we not industrialized in the 1880s and 1890s, would we have been prepared?

The world does look very different today. Global supply chains and the cost of shipping are radically different, but I thought what was interesting is that the Fed actually did a study on the tariffs in 2018. It looked at washing machines, and they basically said the tariffs led to much higher prices for washing machines. Prices also rose for dryers that had no tariff because they’re usually sold together.

Interestingly enough, even domestic producers raised prices because the competitive market had raised prices. They now had a price umbrella that they could raise prices into. I think the researchers concluded that very few jobs were actually created. They were looking at it only 2 years out, so they weren’t looking at the long-run effects of whether it played out over time.

One area I’m really focused on for Silicon Valley is a potential tariff on GPUs or chips, which has been threatened this week. They’re talking about a tariff on Taiwan, which would effectively be a tariff on chips.

Today, we can’t make those chips in America. We don’t have 2-nanometer or 3-nanometer fabs where we could build them even if we wanted to. When I look at that, it’s really just a tax on chip manufacturers and on the end buyers.

That would be a tax on NVIDIA, AMD, and others, and it would be a tax on Meta, Amazon, and everyone who would have to pay it. That likely means fewer chips purchased and less AI research.

If our number 1 goal is to win the race in AI, this is a classic case where, in the short run, I think it’s self-defeating. But there are ways to do this where you can bring more fabs to the U.S.

If you wanted to have a permanent 50% tariff on all chips in the U.S. starting now, I think that’s just going to have negative repercussions. But if you said, “There’s going to be a 50% tax on chips, but I’m going to delay it for 2½ years, and you have to meet these hurdles for building fabs in the U.S.,” that’s more of a negotiating tactic than a permanent, higher tax.

I think there are some really good outcomes from that. We would be less dependent on Taiwan, which is always threatened by what most people believe is a foreign adversary. Less dependency would help in the case of some situation evolving there.

The main point I wanted to make today is: Don’t be lazy in believing the Bessent consensus. Don’t think this is just about a negotiating tactic. Go back and read the speeches. I think Trump and his administration have a much more principled view here—that maybe a value-added-tax-equivalent tariff is a more efficient and better mechanism for helping the middle class in the United States than an income tax.

We saw that the door was left open when they said tariffs may come back as part of tax reform. Keep your eyes out for that.

Any analysis of this situation is made more difficult by the blatant reality that even if Trump is just using tariffs as a negotiating tool, he can never say that out loud, or it would take away their ability to be used in that way. He has to be obscure about it either way, and that makes it harder to know exactly which way is up.

3. DOGE

Let’s move on to DOGE. You were in Washington. What took you there, and what did you see? What’s your perspective on DOGE? I’d love the update on your specific visit, but then could you reflect on why or why not DOGE matters to a tech investor?

I think this is so important because, again, we’re in this fog-of-war situation, where change brings a lot of contentiousness and sometimes we lose the basic facts of what we’re trying to achieve here.

It’s super important to understand that there’s nobody like Elon. He’s truly an N of 1 when it comes to working on issues like this. He’s working 20 hours a day, working through the weekend, and his team is sleeping across from the White House in the executive office building.

Most importantly, Elon is a systems thinker. He showed up in Washington and didn’t do what normal people do when they show up there, which is feed all the politicians and understand what he needs to do to play by the rules everybody else sets.

He just starts asking questions. Not surprisingly, the first question he asks is, “Who sends out the wires? Who controls the wires? Can I get a list of all the wires that are scheduled to go out? What are we spending the money on over the next month? Has it been audited?”

As he started doing that, the Leviathan of Washington convulsed. They said, “Whoa, whoa, whoa. Nobody questions this. Nobody looks at the wires.” He said, “That’s what the president asked me to do, so I need to do that.”

The antibodies really started attacking, even though the only thing he had to do in the first instance was follow his first principles and figure out where to look.

I want to bring it back to the idea that change is hard, but change is necessary. The whole purpose here is to balance the budget, something both parties have proven unable to do through the normal process. Most people agree that the current path is bankrupting the country.

It’s not that hard. We showed this on the podcast when we went through what would happen if you just returned to the 2019 baseline. If we went back to that baseline and grew it by 2½% from 2019, we would balance the budget during this president’s term.

That requires cutting $1 trillion off the COVID high. We lost our minds. Remember the letter to Meta: “Time to get fit.” Silicon Valley has gotten fit. We’ve made some reductions, but government hasn’t gotten fit at all. It hasn’t done anything except stay at that COVID high.

All Elon is saying is, “Let’s just go back. Let’s start by getting $1 trillion of this out.” That’s the excess we put in. If we tell the American people we’re going to do this and put together a believable plan to cut $1 trillion and balance the budget in the next few years, here’s what’s going to happen: Interest rates are going to come down.

The reason bond vigilantes moved into the bond market and started shorting it is that they thought, “Here we go. Trump is going to stimulate the hell out of the economy with a continuation of tax cuts, and nothing is going to really change on costs.”

The big thing I came away thinking is that people are wrong. There’s a fundamental difference in how these people are attacking inefficient spending and getting us back to a very sensible 2019 baseline.

Nobody thought in 2019 that we were starving babies in the streets because our spending was so low. Everybody thought we were spending plenty of money in 2019. That’s all they’re talking about, yet if you watch the convulsion coming out of Washington, you would think something very draconian was going on.

You would expect that. We don’t have term limits. We have lifelong politicians in Washington. Because of Citizens United, you can basically raise unlimited amounts of money from corporate interests.

I gave a speech a year and a half ago on regulatory capture. I’m not surprised that the entity that is Washington is pushing back on someone who wants to take away the tools that give them power.

What people expected, frankly, was that immediately after people saw the relationship with Trump, they would speculate about how long it would be until the relationship blew up. Trump always fires everybody, and the opposite is happening.

People expected Elon to come to Washington and not do anything—to make recommendations to Congress on things that could be cut. But you and I know Elon. There’s no chance he’s going to Washington just to run some research and make recommendations. That expectation was misplaced.

Let me tell you how I think DOGE fits in with a normal budget process, because I also think this is very misunderstood.

We have 2 tracks going on. Track 1 is the normal budget process. In this case, they’re using a parliamentary tool called reconciliation. I’ll spare you the details, but this is coming out of the White House, led by Kevin Hassett; in the House, it’s led by the speaker and the House Budget Committee.

Reconciliation is a special budget process that allows you to get an omnibus budget bill through Congress without having to get the 60 votes in the Senate required to overcome a filibuster.

They’re working hard on this. I expect some meaningful improvements in spending to come out of it. I suspect DOGE will offer its ideas about how to save some money, but this is the normal process that occurs in Washington.

The president has said he wants something to sign out of the reconciliation process in April or May. It has to go through the normal process. All the committees will have their hearings, they’ll put together the budget they think complies with reconciliation, and there will be a grand negotiation with the horse-trading that usually occurs in Washington.

Track 2 is DOGE and cuts in spending by executive authority. This is the part that has set Washington on fire. That’s what you’re seeing. Elon is advising the president, and the president is deciding in real time whether certain people need to be cut and whether certain spending should be stopped.

When the answer is that certain amounts of money and certain people are not required to faithfully execute the laws, they say they’re going to downsize the executive agency tasked with executing the law. They’re going to stop spending money they believe is wasteful and not needed to fulfill the law.

They’re saying that, especially in the face of a national fiscal crisis where we’re falling further and further into a debt spiral, we need to do this.

At the town hall on Sunday night, Elon called out USAID, an organization that’s quite controversial. You can research it. It spends $50 billion a year on foreign aid.

There are thousands of people associated with the agency—among employees, it’s probably closer to 1,000 or 2,000, plus a lot of contractors. Basically, Elon said, “I called the president and told him, unfortunately, there’s no apple to be saved. It’s a total ball of worms. If there were just 1 worm in the apple, we’d pull the worm out, but the whole thing is a ball of worms. The whole thing needs to be shut down.”

They were going to let thousands of people go and save $50 billion on the budget. It subsequently looked like, on Monday, they made a deal in which Marco Rubio, the secretary of state, would become the acting director of the agency. Now it looks like they’re going to eliminate whatever they think is wasteful and perhaps consolidate other parts of that spending into the State Department.

This is what caused Chuck Schumer and others to come out on Monday morning and declare all of this activity unconstitutional. They said nobody elected Elon, he can’t do this, and it’s unconstitutional.

That’s where the whole challenge is going to move. Remember, this has nothing to do with Track 1, except that it’s angering a lot of people on the Democratic side. This is really about Track 2: Does the president have executive authority not to spend money that he deems wasteful?

You asked me earlier whether it’s constitutional. That’s a fascinating constitutional question. I’ve consulted with many people who are experts in the area. I expect Schumer or a group of members, as early as this week, to file a claim in federal court arguing that there’s a violation of the Constitution under Article 1, Section 99, Clause 7, where Congress has the power of the purse.

The Supreme Court has long upheld this. The argument will be that separation of powers generally supports the idea that Congress appropriates funds, and anybody else who doesn’t spend those monies would be acting unconstitutionally. They’re likely to say they have to immediately cease and desist from shutting off wires, not spending money, or shutting down USAID.

I happen to think that argument is on pretty weak footing.

Why?

Think about this for a second. The president has the authority to execute the laws, and there’s a doctrine known as “empowerment” that the courts largely recognize.

It’s basically the president saying, “I see the laws we’re supposed to uphold, and I don’t need all this money. In fact, I have a further and maybe supreme duty—an overriding duty—to the Constitution that supersedes Congress’s constitutional control of the purse. I have to faithfully execute the laws and protect the general welfare of the American people, which might include protecting the country from bankruptcy.”

He’s saying, “I’m doing my duty. I’m executing all the laws I’ve been told to execute. However, I’m doing it for less money, and given that we’re in a national debt crisis, I need to do that to protect the American people.”

I think this is eventually going to come to a head. Imagine it goes to the Supreme Court to decide. There’s a decent chance that, along the way, at a minimum, Chuck Schumer is going to have to defend some of this really crazy spending.

There’s nobody on either side of this argument who doesn’t think there’s a lot of inefficient and silly spending by the government. If you want to defend this lawsuit, effectively that’s what you’re going to have to defend.

The political pressure is going to be massive, particularly because DOGE is being so transparent. You do not want to be defending every single line item to the American people, which is exactly what DOGE is going to put you on the spot to do.

I see 2 potential outcomes. Number 1, the political pressure causes them to cut a lot more as part of Track 1, the reconciliation process. Number 2, the Supreme Court recognizes some more expansive executive power around “empowerment.”

Either way, before this is all said and done, I imagine we’re going to see headlines saying, “Elon causes a constitutional crisis,” with the courts involved and the solicitor general representing the executive branch and the White House on this matter.

4. Tech and Political Uncertainty

Now, bring it home. Why does this matter for tech investors? Let’s presume it goes either way. What does it mean for how a tech investor should be thinking about the markets and tech stocks?

Think about the 3 topics we talked about today.

The first was massive technological uncertainty. It’s a pace of change you’ve never seen in your career. Highly disruptive companies valued at $150 billion are being challenged by a Chinese startup on a shoestring.

We would both say that our ability to forecast the future and where this is going is challenged because it’s moving so fast.

Then we talked about tariffs. That’s massive economic uncertainty. Free trade has generally been established as a principle in the economy for the better part of your entire investment career and mine. The markets could count on that.

Now I’m suggesting there’s at least some probability that this president is going to move in a very different direction. Maybe it’s not the Bessent consensus. Maybe it’s something else. It might be the Trump new normal, where tariffs become standard practice and perhaps replace income taxes.

There’s a lot of uncertainty around that. The third issue is political. It’s been a while since we had a looming political constitutional crisis where an issue between the congressional branch and the executive branch went to the Supreme Court.

That also yields a lot of uncertainty. When you add these uncertainties up, what does it do for the value of assets that you and I look at? We’re valuing future cash flows, and we have to apply a discount rate. The discount rate measures the risk associated with those future cash flows.

We have to take the discount rate up because we’re a lot less certain about technology, politics, and economics. To me, that means multiples come down and asset prices have to come down while the world sifts through all of this.

The surprising thing to me is how well the public markets have held up in the face of all of it. I think part of that has to do with the belief that Trump is going to be a super-pro-growth president, with lower taxes and so forth.

As a risk manager, what I have to say to my team is that we have downside risk. You don’t do the 10th-best idea, or the 11th-best idea. You have to make sure you understand this stuff.

For the long-term investor, perhaps they can ignore the noise and say, “I’m fully invested. I believe in this supercycle. AI is going to be great for everything.” But what I say to our friends in Silicon Valley is to expect much more volatility.

I think the next 6 months are going to have a lot of volatility. It’s exactly what we felt all weekend long. It’s exactly why the markets were gapping down overnight on Sunday, and then they did a U-turn because we got a change in policy—or what appeared to be a change in policy—from the White House.

Welcome back to 2017, Bill. All of this change may be absolutely necessary and totally good for Team America, but it’s going to mean more sleepless nights.

Because of so much chaos, massive uncertainty in regulatory action, and the fact that it can often backfire, I think of the example of DeepSeek. First of all, I think the economic issue got blown out of proportion. The paper originally said $6 million was just the post-training cost, and somebody interpreted that as the whole thing.

That led to a lot of people—particularly a lot of NVIDIA bulls—running out and saying, “No, no, no. They had way more NVIDIA. They had way more GPUs than people thought.”

But pounding the table on that may cause people in Washington to say, “We should have had higher restrictions.” If you’re an NVIDIA bull and think you’re protecting NVIDIA by exposing DeepSeek, you may end up with sanctions that hurt NVIDIA’s revenues.

It’s a dangerous place to play.

You just showed that there are all these forces at play. What did I do as an investor? What did I do in the fall of 2022 that led me into NVIDIA in the first place? I studied what was happening in technology and studied the company.

I didn’t have to think about free trade or tariffs. I didn’t have to think about export restrictions. I didn’t have to think about a constitutional crisis. All I had to figure out was whether the forecast for NVIDIA was too low because of the explosion we were about to have in AI. That’s the bet we made, and we won big on it.

But now, as I sit here today, NVIDIA’s valuation is much higher, and I have to take all these other risks into consideration.

All I’m saying is that, all else being equal, I think the supercycle is on fire. We’re going to need way more compute than we have. DeepSeek unleashes the amount of inference we’re going to need, and Deep Research from OpenAI unleashes more demand as well.

The fundamentals are bigger than ever. At the same time, I’m humble in the face of what’s known and knowable about the next 12 months around tariffs, export controls, and all these other risks in the economy.

You have to look in the mirror and acknowledge that a lot of this is unpredictable. That impacts what people are willing to pay and what multiple they’re willing to pay.

For our friends in the venture-capital markets, particularly some of these highly valued companies in the middle- and late-stage venture market, it’s going to have an impact. There’s always a lag effect, but the public markets, risk appetite, and the multiples they pay roll downhill.

We may get to the back half of this year or into next year, and it may in fact be the golden age and off to the races. But at the moment, it’s the golden age of uncertainty.

Hey, Brad, let’s close with where we started, with this sovereign wealth fund. Pick either the pro or the con, make the argument, and I’ll take the other side.

I love the fact that we have people with business sensibilities and incentives looking out for America, who want to negotiate on behalf of America. We sell wireless spectrum and licenses, and I would love to see that go to the benefit of all the citizens in the country. We have drilling licenses on national lands, and that money belongs to the citizens.

I love that idea. Here’s my challenge with it, Bill: We have $40 trillion in debt and probably another $50 trillion of unfunded liabilities. We’re a debtor nation, and we’re paying 5% on all that debt.

The hurdle rate for the return needed on the sovereign wealth fund is therefore 5%. Otherwise, you would just take all those monies and pay down the debt, if this were our personal balance sheet.

What you and I would describe this as is levering up the balance sheet of the United States to earn the spread between the sovereign wealth fund’s returns and the 5% we’re paying to borrow all the money.

I think I’d probably have it in place because I think it’s a good tactical lever. I do worry about what happens from administration to administration. It could lead to crony capitalism and dealmaking that benefits certain people.

I don’t know. It’s a close call for me, but I think it’s going to happen either way.

You took both sides, so I’ll do a quick both sides.

In addition to the scenarios you talked about—and you and I have debated this in the past—if I look at Goldman Sachs, GM, or United, if the government is going to be the lender of last resort, I would argue that it should take all of the equity. This could be a vehicle for that, although nothing has kept the government from doing that. In the GM case, it did take equity.

I’m much more skeptical than you are about crony capitalism. It would be a 99% certainty that this asset would be raided from administration to administration in the government.

I would highlight that the most successful sovereign wealth funds in the world are all in autocracies. They’re not in democracies.

I don’t agree with that. From Norway to Korea to Canada, we have a lot of great sovereign wealth funds.

I think one of the things you do have in all those countries, Bill, is consistency and independence in the management of the sovereign wealth fund—independence from unilateral control by the executive branch.

If you want to see how it can go really wrong, go read “Billion Dollar Whale.” It’s one of the most exciting books you could possibly read about what happened to the Malaysian sovereign wealth fund.

As always, it’s fun to get together. Thanks for making the time. We’ll talk soon.