Coinbase CEO眼中的2026年加密趋势Top 3：以及Davos的更多内容！

>> The besties are broadcasting from the USA House at the World Economic Forum. Our episode is sponsored by the New York Stock Exchange. Are you looking to change the world and raise capital? Do it at the NYSC. The NYSE is a modern marketplace and a massive platform built for scale and long-term impact. So, if you're building for the future, the NYSC is where it happens. >> I'm Jason Calakanis. This is the All-In interview show. Uh, last minute we got added to the roster here at the World Economic Forum, and we had time to do a halfozen interviews and, uh, Brian was here, and uh, this is your Brian Armstrong from Coinbase, of course, and friend of the pod. This is probably your fourth or fifth appearance on the pod.

You come to Davos because this actually isn't about networking for you. This is about serious regulations on a global basis.

Yeah.

Well, that's been the focus of this attendance at Davos. We are trying to get market-structure legislation done for crypto. There is a lot of networking that happens here, too. We've had a lot of commercial meetings. Five of the top 20 global banks are now using Coinbase to build crypto infrastructure into their products.

We meet with heads of state and leaders of different countries and talk to them about economic freedom and how crypto can update their financial systems. There are all kinds of good meetings.

You had embedded in there these partnerships with banks. Is that a white-label type of thing, so they can sell crypto to their customers? Is it disclosed which banks are doing that and how it works?

A couple of them are public. We've talked about integration with JPMorgan and PNC Bank. There are a couple of others that are not public yet, but 5 of the top 20 global banks are now using Coinbase for that.

We're also powering integrations with BlackRock. They've said they want to tokenize every single one of their funds. A lot of these financial institutions are coming on-chain, which is great.

I was thinking on the way over here how you've really struggled to work with regulators over the last decade. I remember under the Biden administration, the 46th administration, you went to D.C. and were like, “I'm here. I would love to talk to you.” They were like, “Yeah, we don't want to talk to you.”

Some people might have varying feelings about Donald J. Trump, our 47th president. But one thing he has nailed is interfacing with the business community and taking regulation—and creating a legal path for crypto specifically—very seriously. How have things changed for you in the last year?

I know you like to call balls and strikes, and I think, just looking at it objectively, the Biden administration really tried to unlawfully kill this industry in America, from my point of view. Donald J. Trump—you've got to give him credit. He campaigned on this idea of making the United States the crypto capital of the world. He's kept his promises. He's leaned in and tried to get clear rules and regulations passed so that American companies can thrive and American consumers can earn more money on their money.

He also understands that it's an important political issue. There's a huge base—there are like 52 million Americans who've used crypto now.

Right?

They want to see clear rules. They want to see better financial services in the United States. It's also, frankly, a global-competitiveness issue. China just announced that they're going to pay interest on their central-bank digital currency. Some of the largest stablecoin issuers are still offshore. He wants to repatriate that capital and bring it into the U.S.

This is the crazy thing we went through. I was never a fan, calling balls and strikes, of people doing things that weren't buttoned up.

Mm-hmm.

But I was even less of a fan of the prior administration not meeting and saying, “Hey, this is uniquely different. Let's figure out a way to give you a path to do it properly.”

In our industry, sometimes you have to reinterpret rules. Airbnb and Uber—the biggest successes of my investment career—bent rules, too. Crypto bent some rules. In some cases, people broke them, and they paid the price. But here we are now; the rule set is being refined.

The most important one, I think, for you is stablecoins and your competition with the banks. You have banks as partners, but you're also a competitor to them.

I'd say it's mostly collaborative. Of the bank CEOs that I've met with here, most of them are actually very into crypto. They're starting to integrate it.

I met with one of the top 10 global banks in the world yesterday, and the CEO told me, “Crypto is my number-one priority. We view this as existential. We're all in.”

Why is it existential for them? What do you think?

They're seeing it like when the internet came around, and you had Amazon competing with Barnes & Noble, or blogs competing with the New York Times in print.

Yes.

Anytime there's change happening in the world, you can think of it as an opportunity, or you can think of it as a threat and bury your head in the sand and pretend it's not happening. But the reality is that crypto is massive. Something like 500 million people have used it globally. Bitcoin was the best-performing asset class of the last decade.

The largest financial institutions in the world are now integrating this. At this point, I think it's foolish to pretend that this isn't happening. We also have the GENIUS Act. The stablecoin bill has now passed into law, so we're not going to undo that. That is the law of the land. Congress just put that into law.

It's very important because I think what David Sacks—my bestie—led there was that these have to be audited. These have to be above board. We can't have a run on stablecoins, which, let's face it, people anticipated Tether would have at some point. There were lots of fines they got. There were these attestations, and people didn't know if they even had the resources they said they had.

Now it's pretty clear: You have to keep your assets in Treasuries. Correct?

That's correct. Under the GENIUS Act, which passed into law last year, U.S.-regulated stablecoins have to have 100% of their assets stored in short-term U.S. Treasuries. Something like 30 days; 30-day Treasuries are the maximum, I believe.

That's pretty much the safest thing you can get. You're basically trusting that the United States government is not going to fail in 30 days, which I think is a pretty safe bet.

I'm going to go with safe bet.

I've been making this point as well: Banks do something called fractional-reserve lending. They actually don't store all your money there; they're lending it out. That's why they have such high regulatory overhead, because there can be a run on the bank, and it gives them a very unique business model.

They can basically lend it out. The old joke is, “You lend it out at 6%, you pay 3%, and you're on the tee by 3:00,” or whatever. But that business model is not available to you unless you have a bank license. In a stablecoin world with 100% reserves, you don't need a bank license for that, and you can give people—

Because it's safer, right? We saw this with Silicon Valley Bank, essentially—

It had mistimed its allocations with Treasuries, I guess.

And what happened? They had a run on the bank. Literally, I was in a board meeting—I think it was a Thursday—and the run happened Thursday afternoon.

I got a text: “Get your money out of Silicon Valley Bank.” I'm in the board meeting, and we were having it on the docket. The third thing was to talk about Silicon Valley Bank, and we had 100% of our money in there. Two of the board members were like, “We can't just take all the money out of Silicon Valley Bank. They've been incredible partners for 30 years.”

I said, “How about we take out half so we can make payroll?” I insisted.

Yeah.

Literally that night, boom. The key issue now is your business model. You need to have revenue, and the revenue from these stablecoins is paying some interest, with the people who are putting their money in there being able to make some interest on their hard-earned capital. That's the sticking point for you.

It's not interest; it's a rewards program. This was carefully negotiated in the GENIUS Act.

Yes, that's our view. In my opinion, it's actually—what's the difference there? The rewards program: We should think of it like American Express points.

There are lots of credit-card reward programs, but the legal difference is that rewards can't be based solely on the balance you're holding. The customer has to do some sort of other activity, like payments or trading, or have a subscription to Coinbase One.

When customers do that, we pass along about 100% of the economics to them for holding those stablecoins with us. That's a big driver of growth.

There's always been this balance between whether people want to put their money in money markets or whether they want to put it in bank deposits. I don't think crypto is really new in that dimension. It's just another flavor of this happening.

There’s been a lot of hand-wringing about this destroying the entire lending market, and I don’t think that’s true. Money markets are already trillions of dollars, and there are high-yield checking accounts. But these banks haven’t had to deal with a disruptive technology competitor that’s really good at what it does, so they’re a little bit nervous about their franchise.

Is that my interpretation? Am I correct?

Some of them are nervous, and some of them are leaning into it as an opportunity. I think the latter. We want everyone to win here. I don’t speak for the president, but my interpretation of his comments is that he wants all American businesses to win. There is a win-win outcome here. But if someone is going to try to undermine his legislation that just got passed in the GENIUS Act, he’d probably—

Is that what’s happening now? Are the banks trying to retrade the deal?

I want to be careful here. The bank trade groups, which I believe are trying to undo the GENIUS Act, are doing so even though it just got passed into law 4 months ago. For us, that’s a red line. I’ve talked to many others in the industry, and for them, that’s a red line as well. I think we have to accept that this is law and that it’s going to continue to exist. But that doesn’t mean banks and crypto companies can’t both win in this new world.

Yeah. So this is just a classic tale of incumbents.

Yeah, incumbents and new folks. You want to partner with them. You want to enable it. You have a good partnership with Jeremy Allaire, an old friend of mine at Circle. Is USDC the default stablecoin on Coinbase, or how do you think about the relationship with them? How should we think about the relationship with them?

Yes, we’ve got a strong relationship with Circle, and USDC is the largest regulated stablecoin because it is compliant under the GENIUS Act in the U.S. It’s compliant under MiCA in Europe, et cetera. There’s another one that you’re familiar with.

But I think it’s in the process of—

They’re trying to clean it up, is my understanding.

Yeah.

The likely scenario is that there’ll be 2 Tethers: a United States one that complies, and then the Wild West one outside the U.S. Is that what you’ve heard as well?

Yeah. And I should mention that we don’t have an exclusive with Circle or anything like that. We actually list other stablecoins on our platform.

Do you list Tether?

We support it in certain ways, especially for people who want to convert Tether. We support it. We also support PayPal’s stablecoin. We’re open to listing others, too, so we don’t have an exclusive on USDC.

Yeah, but you’re not endorsing it. Do you let people trade into Tether, or do you just let them trade out of Tether? How does it work mechanically?

I think it’s different in different countries. I want to make sure I get it exactly right, but in countries where we’re allowed to do it, we support Tether.

Right?

It’s nuanced.

Are you concerned about, or have you historically been concerned about, Tether and its somewhat loosey-goosey approach to regulations and trading? I’m giving it that descriptor, not you. I would think that having it on your platform, with regulators pretty focused on it over the last 5 or 10 years and this belief that it could all come apart and create a run, might seem like a risk that’s just not worth taking—that you don’t want to be too close to it in case it does flip over.

Yeah. We’ve definitely gotten questions about it. I want to be careful here: I actually like the Tether guys. I think they’ve done a lot of good things in the world. There are people who are really struggling with local currencies that have 70% to 100% inflation year over year, so there was high demand for the dollar. Tether got great distribution in a lot of the emerging markets. I actually think they’ve done a lot of good for the world.

But it’s not currently compliant under the GENIUS Act in the U.S., and it doesn’t follow those same requirements for 100% reserves in short-term U.S. Treasuries, to my understanding. People have to make their own determination on that. I think other countries are following suit in terms of cleaning this up.

Is there a way in crypto now to give consumers an objective rating? This one has this grade and follows these regulations; this one follows a different level and is a lower grade; this one doesn’t follow anything and is a memecoin. Whatever—this is the Wild West, with no crying in the casino, coins. What is your responsibility as a platform, or your opportunity as a platform, to inform the people who are participating?

What we try to do is have minimum listing standards. If we believe there’s a cybersecurity risk—for example, the developer could rug everyone—or if it’s illegal from a compliance point of view, there are a few different areas we look at. If it meets the minimum bar, we’ll list it, and then we let customers decide.

I don’t feel like it’s our job to recommend investments. In the traditional financial world, there are AAA-rated bonds, and it always felt a little bit like the organizations that do the ratings are politicized.

Go see The Big Short.

Yeah, exactly. I think of it a little bit like the app stores, or let’s say Amazon. You want to have the everything exchange; you want to have the everything store. Everything that’s legal should be in the store, but maybe there are customer reviews we could add at some point.

We actually tried that for a little bit. If you see a 2-out-of-5-star rating on Amazon, you can still buy it, but at least you’re informed. We tried making user ratings at one point, but it didn’t go that well because people were basically voting with whatever they liked.

Sure. They’re talking their book.

Yeah, talking their book. Anyway, right now we’re in the regime of disclosures and minimum listing standards.

Which crypto projects do you find the most fascinating right now? Bittensor? Some of these projects that are popping up and actually providing technological solutions to problems, like distributed computing? I find those fascinating.

They are. People are trying to tokenize data centers and oil reserves. I think the biggest trends happening in crypto right now are, number 1, the everything exchange. It’s not just crypto that you can trade; equities are increasingly getting closer to being able to trade on-chain. Prediction markets are—

You have a partner for that, right?

Yeah, we’re working with Kalshi currently.

Is that exclusive, or are you willing to put anybody up on the platform?

It’s not exclusive, so we’re looking at others as well. I know you guys work with Polymarket on the show. I was one of the original angel investors in Robinhood, and I think they’re doing Kalshi, too.

It seems like people are plugging different ones in. Polymarket’s our favorite.

Yeah. We’re talking to Polymarket, but we can also list our own prediction markets.

Oh, yeah. So you could fire up your own.

Yeah. Anyway, we’re along for the ride. I think the biggest trends are that all assets are coming on-chain for trading, prediction markets are growing like crazy, and stablecoin payments are growing like crazy. Those are probably the 3 biggest trends in crypto right now.

When do you think stablecoins tip into the area of businesses using them for payments to reduce friction? Maybe a designer does some work for Coinbase and makes a new logo, and you want to send them $25,000. When does that start to happen? And for consumers, when do people at a poker game start settling up through their Coinbase accounts with a stablecoin?

The biggest growth area over the last year has been B2B cross-border payments.

Yeah, cross-border especially. There are a lot of companies that might be buying goods from Asia or Europe and trying to sell them in their shop in Brazil, or whatever it is. They have to wait 7 days, and there are high foreign-exchange fees and all this kind of friction to move the money.

Crazy, the fees.

Yeah. That’s been growing really nicely. We launched a product called Coinbase Business, which serves lots of small and medium-sized companies that want to do cross-border payments, invoicing, tax, accounting, and all that.

How do you find those customers? Is that a really unique group of people, or do they just find you?

Currently, they’re beating a path to our door. We actually have a huge backlog of people waiting to onboard, so we need to staff up that team.

We also launched something called Coinbase Developer Platform, which is kind of like AWS. You can white-label anything, like with the banks, but lots of other businesses are using it for wallets, trading, payments, staking, financing, and all kinds of things.

You mentioned tokenization. I was talking with Vlad, and he did a little experiment: “Why don’t I tokenize some OpenAI shares?” Sam Altman wasn’t too thrilled with that.

How do you think about that opportunity? I’m a private-market investor. I would love to be able to take my early position in Robinhood or my early position in Uber as it was going up, put it into a market, and let people trade it. That would be very interesting for VCs and angel investors, to be able to move that around. How do you think about it?

Well, I think it has to be done with the permission of the companies, because if you’re a private company, you don’t want your employees to be able to get liquid after 1 year. You’re trying to retain them; that’s why you have vesting. It’s a retention mechanism, right? Let’s all build this together, and maybe when we go public. There are stories of founders who took a little secondary too early, then the company didn’t work out, and the company was worse off.

So, I think what’s going to happen in crypto is that, first of all, we should make on-chain capital formation way easier for private companies. If you want to go—this is what we’re chatting with the SEC and others about—you know, can you go register a security? Right now, you’d only be able to raise money from accredited investors in the US. I know you and I agree on this: we’d like to expand how you can become an accredited investor.

There is a bill right now that’s working its way through, and it would basically mean the SEC—and they’ve already been charged with this, but I don’t know if you’ve studied the SEC at all. They tend to take their time and then not do what they’ve been told to do. One of the things they were supposed to do was create an accreditation test.

Well, I’d say the SEC is actually moving very quickly. But this one is—yes.

Yeah, I think that would be a fairer way, because otherwise it’s kind of like a regressive tax: only rich people can get richer on private investments. Anyway, I think on-chain capital formation is going to be massive for private companies. Eventually, I think you’ll actually just be able to go public totally on-chain, too.

That would lower the cost massively, reduce the friction, and increase the democratization of wealth creation.

Yeah.

If you think about when you were a private company—

Mhm.

You had all this pent-up demand, with people trying to buy the shares like crazy. They were all doing backdoor, kind of shady stuff, popping up SPVs. I don’t know if you’ve been following the SPV market now, but it’s turned into a boiler room. It’s no longer Chris Sacca representing Twitter and doing an orderly thing, or Elon doing an orderly thing every 6 months where SpaceX keeps control of it. Now people are going out, raising money from dentists and civilians—high-net-worth individuals—to buy SpaceX or Anduril or whatever it is. Then they go try to find the shares and charge a 10% load-in fee, with no carry.

Mhm.

I mean, think about how crazy that is.

Yeah. There’s such high demand for some of these large private companies.

You know, it’s kind of a good example of the unintended consequences of higher regulation sometimes.

Yeah.

Sarbanes-Oxley and all that kind of stuff really cut down the number of companies and how long they stayed private before they went public. Uber, Airbnb, and a lot of these companies—all the money was made by private or accredited investors, like yourself. Then, when they finally went public, it kind of went sideways.

Oh, yeah.

Airbnb and Uber all had a 5-year indigestion period, I would say. Some people, I think Instacart, wound up going from $30 billion down to $10 billion when they went public. It was like, “Okay, we’ve got to dig out of a hole for the last series of investors.” That is the unintended consequence of this, because you don’t have anybody setting a proper valuation for the company in some reasonable way. What about funds? I get approached by a lot of people offshore saying, “Hey, take your next seed fund. You’re going to do a $50 million fund. Put it on-chain.” If you were one of my LPs and you needed liquidity, I could just sell it to somebody else. If I were an LP in a Sequoia fund and I wanted to sell you the interest, you could buy it from me, and we could just take our wallets out and zip, zip.

Yeah, I think that’s absolutely going to happen. Coinbase launched a product actually called Coinbase Tokenize. We’re helping any fund, real estate project, or anybody who wants to tokenize their products. It democratizes access and increases demand. It gets rid of a lot of the back-office fees and the settlement risk, because it can be settled instantly on-chain. Some very innovative companies—the top funds in the world, like BlackRock and Apollo—have come out publicly and said they want to tokenize every single one of their products. It’s absolutely happening.

How do they keep control of it? You have these more liquid backdoor consequences, or downstream effects—second-order effects, third-order effects. What are the second- and third-order effects that would happen if a venture fund or a REIT were on-chain? Have you thought that through?

Have you thought it through? Yeah. There are different types of funds. Some are going to be available only to institutions and accredited investors. Some would be open to retail.

For the retail side, you could actually get tens of millions of people around the world, in 5 minutes, to all put in money. The average price might be $100 or $1,000, right? It starts to really democratize access. We recently published a report, and people have heard about the unbanked, but there are actually 4 billion adults who are unbrokered as well, which means they don’t have any ability to invest in these high-quality assets.

This is the engine of wealth creation for capitalists like you and me. A lot of people are just stuck. The only way they can earn is from their labor, right? They might want to put 10% of their $100 or $1,000 into the S&P 500, Coinbase stock, NVIDIA, or whatever, and they can’t do that.

But they can use PrizePicks. They could use some other thing. No, and I like PrizePicks. I use PrizePicks to bet on Knicks parlays. But they wind up putting it somewhere else. Why not be able to invest? If they want to bet on the Knicks, that’s fine. Or go to Vegas; that’s fine, too, and play in a poker tournament.

Maybe they hear about a company like LinkedIn because they work in the HR department, and everybody in the HR department is over the moon about it. That rank-and-file, $75K person working in HR understands what the next big product will be. They’ll know whether Indeed or LinkedIn is going to work, and they can make a life-changing bet with just $1,000. They could make a 1,000x return.

Yeah. There’s a financial-literacy component to this as well. I think the AI agents are now getting really good; we’ve integrated one into the Coinbase app. It can teach people about dollar-cost averaging and tax-loss harvesting. Financial education is a part of it, but then, yeah, let’s make high-quality investments available to them and democratize access. It’s just like lifting people out of poverty. It’s great.

I mean, you’re thinking about the whole globe, but just thinking in the United States, a lot of what people are upset about—and the topic we’ve been talking about a lot—is the rise of socialism in New York specifically, California. I’m not sure if you’re still a resident. I won’t put you on the spot here, but—

Considering options.

Considering—I mean—

As we all are. Three years ago, I moved to Austin. I was like, “I’m done.” The social issues—I saw the writing on the wall. I actually think there’s a chance that California goes bankrupt. I said that on the podcast: This feels like it’s trending toward insolvency. I never thought they would get to the wealth tax or just seizing people’s assets.

Yeah. What’s your take on all that? I was reading something this morning that said the people who have already left—which, by my estimate, is probably 20% of the billionaires in California—have already created a negative $10 billion tax hole, even with the amount they hope to raise from the people who stay.

Yeah. So, it’s one of the biggest self-owns I’ve ever seen. It’s a disaster, and I’m torn, actually, because there’s always this question of voice or exit: Do you try to fix it from within, or do you leave like you did?

I think the incentives are strange because, on the one hand, I love California. On the other hand, it’s been like an abusive relationship. It just keeps coming back with another thing and another thing. In some ways, if you do make the decision to leave, you don’t really have much incentive to try to fix it at that point. You actually want to get a lot of the builders and top talent out of California and resettle somewhere new that’s welcoming to us and to businesses.

Yeah. I don't think people understand how easy it is for somebody who's in a certain stratum and is already operating globally. I'm on planes and on 4 different continents every year. It doesn't matter where I am. What matters is that my wife and my kids are happy, that they love the place we live, and that we love Austin.

The thing I've seen—probably one of the hardest things you had to deal with with your employees and your team at Coinbase—is the price of their housing. How many times did you try to recruit somebody to come to California, and it's a family of 4? They need private school, they need a house, and you're like, "Oh my God, what is their nut going to be here?"

Their nut. I love that. [laughter] There's a great South Park episode on that. But, yeah, you're right. That's a major barrier whenever we make an offer to somebody in California or for NewLimit, the biotech. It's always like, "Well, my cost of living is going to double. You need to pay me more." So, it's getting expensed through to you, the business owner.

Yeah, it does. San Francisco had this revenue-based tax that was very punitive on financial services companies. Stripe moved out when that happened. I think Mark Benioff regrets supporting that one.

Yeah, I think so.

Yeah. But in fairness, he did want to finance the homeless industrial complex, which has been completely ineffective in reducing the number of homeless individuals because—

They're not homeless. They're addicted to drugs.

A home doesn't help that problem.

Exactly. I mean, I'm preaching to the choir here, but people would have a lot more tolerance for paying higher taxes if they felt like it was working. The history of the last 10 years in California is that the budget has gone up dramatically and the services have gotten worse. It's actually creating the wrong incentives: The more money we spend on homelessness, the more homeless people there are. And then there's the waste and fraud. Oh my gosh. You guys saw Nick Shirley and all that.

What could we even guess is the level of abuse in California? It's going to make Minnesota look—

Like peanuts. Such a big economy, with so many nonprofits and so many homeless organizations taking down hundreds of millions of dollars in San Francisco alone.

Yeah. So, tell everybody about the side hustle, your other company.

Oh, the biotech.

Yeah.

Well, when Coinbase went public in 2021, I got some liquidity from that, and I thought it through. I was like, "All right, I want to continue to be CEO of Coinbase. Being a public company CEO is a really cool thing. I just feel like we're at the beginning of our journey."

But I also felt like I wanted to start using some of that capital to go after these big bets. I was a little inspired by Elon, actually. He did the thing with PayPal and X, and then he went into the world of atoms, not bits. Software is more forgiving. Startups are all hard, but software is a little more forgiving because you have higher margins. The world of atoms is much less forgiving.

Anyway, I was lucky enough to meet some really amazing co-founders who came together with this idea in the longevity space. The fundamental science behind it is called epigenetic reprogramming. You can reprogram your cells to restore the function they had when they were younger.

There was some really cool research being done. I hosted a couple of dinners. Anyway, I decided to fund these guys, and a bunch of other people have invested now. I'm a board member, and I've been helping them.

And the name of it is—

New Limit.

New Limit.

Yes. They've made incredible progress.

When will they have a product? This feels like a 20-year investment, not a 2- or 5-year investment.

Yeah. Biotech does move more slowly, but it's moved faster than I would have thought. I thought this was going to be 5 years of just basic research, but it turned out that within the first 2 or 3 years, they were able to successfully demonstrate the reprogramming of human cells to restore the function they had when they were younger. The first drug candidate is probably going to go into clinical trials next year.

Amazing.

Yeah, that's super rewarding.

Five minutes. Yeah, okay, great. So, coming out of Davos, what's your take on the state of the world? Everybody, when they get here, seems to be saying that the ESG and DEI kumbaya stuff has switched in the last year or two to brass-tacks dealmaking, whether it's between countries and businesses.

This is turning into a business conference. It used to be—this is what everybody's telling me on the streets, in the houses—it's about business now, and on the margins there's a patina of globalization versus nationalism. What's your take on the state of the world in 2026, talking to regulators and people who work in government?

I think you're right. I've only been at Davos once before, but it did feel more like, "How do we make a global government? How do we do lots of ESG and DEI?" That's really not what anyone's talking about now.

I think partially it's because of Larry Fink coming in, the new leader of Davos, more or less. I also think it's because of Donald Trump.

Yeah, he shook it up.

Yeah. The numbers that the United States is putting up in terms of GDP growth and low inflation, and just that business environment—it's like, "Hey, how do we all win?" That's how you create prosperity for everyone in society.

I do think it actually benefits everyone. Even the poorest people in society do the best in high-economic-freedom countries, anyway.

Growth solves a lot of problems.

It does. It does. The growth is objectively spectacular. We have not seen this level of GDP growth since we pumped a bunch of money and printed a bunch of dollars during COVID. 5.6% GDP growth is pretty spectacular. Let's hope it keeps up.

Unemployment is very reasonable at 4.6%, the lowest of our lifetime. I think 4.3% was the lowest it hit. Inflation is closer to 3% than 2%, but the actual average has been around 2.8% or 2.9%. So, 2% is the target.

Yeah. So, we're right around the average, not hitting the target yet, but I think we'll get there. It seems like growth does not come from government spending, right? That's the key. This Keynesian economic argument, I think, is basically wrong.

Growth comes from deregulation, low-cost energy, allowing the private markets to build, letting them have clear rules about what's allowed and what's not, and then creating a level playing field. Everyone competes, the consumer benefits, the companies benefit, all the employees, the shareholders. Capitalism is the biggest win-win. Someone had a great rant about that recently.

Yeah.

We're seeing the private companies in the United States really cook. It's great, right? And if you're cooking, you create more jobs, hopefully pay more taxes, and all of that just starts the cycle in the right direction.

How do you—I’ll end on AI. You haven't been on with the 4 of us in a while, but when somebody gets sick, we'll definitely rotate you in. Everybody loves when you're on, like the quartet. You're a fan favorite, by the way.

I'm curious what you think about AI and job displacement. Sacks and I have been debating this. When is it going to be here? Is it here? Young people can't find jobs, but we're still at a pretty low unemployment rate overall. Then Elon's position and Bernie Sanders's position are in sync: "Hey, listen. It's going to be a lot of job displacement."

How do you think about it? Obviously, Amazon is also the one I'm watching, because the idea that somebody's going to drive packages or pack packages in the age of Optimus, robotaxis, and Waymo sounds crazy. Those jobs are going away. How do you think about job displacement, and what are you seeing with the most AI-first employees at Coinbase?

Just zooming out for a second, I think crypto and AI are the 2 most important technology trends happening in the world. What's cool—and most people don't realize this—is that they're actually going to come together, because AI agents need to get work done and they have to make payments.

The whole traditional financial system is built around knowing there's a human behind every product. You upload your—

Oh, know your customer.

Yeah. So, AI agents, I think, are going to use stablecoins and crypto wallets.

Know your agent.

Well, I don't even know if you need to know the agent, but, yeah. Anyway, that's one of the important trends that we're trying to help happen.

In terms of job displacement, I don't know. Maybe this is a bit of a techno-optimist take, but I actually think, if you go back and look at the early 1900s, I think it was like 80% of the U.S. population was working in agriculture. That's hard manual labor out in the fields. When agriculture got automated, now it's 3% or something of the workforce working in agriculture.

That happened over 30 years.

Yeah.

Yeah. And so I think they would look at what you and I do for a living—we're just having a cool conversation in Davos, talking—and they'd be like, “That's not a real job.” A real job is manual labor in the fields, right?

You just—you’re on vacation all the time. But we think of it as a job, and people who are typing on a keyboard get to sit in an air-conditioned office or whatever.

It’s stressful. That’s for sure.

It can be stressful, but I’d rather be doing that than backbreaking labor in the sun, digging a ditch or something. So I think job displacement is not a bad thing, actually, if it means that people can do new kinds of work and new kinds of jobs. Is there going to be a transition period? Yes.

I basically think if AI plays out as we all think it will, with robots, there will be a lot of job displacement. But it means that we’ll be in a world of more abundance, and people are going to have jobs—streaming video games on YouTube or whatever. I don’t know what it’s going to be, but great philosophical works might be written because we don’t have to burden ourselves with the tedium of packing boxes. I’m basically an optimist on it. I think it’ll be good.

I’m pretty optimistic about it as well. Having watched the robotaxi self-driving thing—

And just watching the velocity that it’s getting better, and having watched the Uber story up close for 12 years, I’m like, yeah, that’s going to happen in 6. Mhm.

I think it’s just going to—

Yeah. It feels to me—and the thing I’m starting to see in the field is that, in Wuhan and Beijing, they’re having protests and saying, “Well, we’re just going to give out a certain number of self-driving licenses. We’re going to contain it.” Then you have Boston and a couple of places in California saying, “Hey, listen. We’re only going to allow a certain number of robotaxis.” Or Boston’s saying, “We’re not going to let you have them here. We’re going to protect these jobs.”

This is going to become one of these class debates over the coming years. But what about employees in the company? You have the same number of employees as you had a couple of years ago. You overhired for a bit, maybe, or were hiring for growth. How do you think about hiring and training young people versus just automating things or having AI? You must have some people on staff who are using Claude Cowork or something, and they’re just 10x knowledge workers. Developers are obvious, but talk to me about knowledge workers and what you’re seeing with your most AI-first employees.

Yeah. One of the big pushes we made in the last year was getting our own internally hosted AI model connected to all of our data sources. Every Slack message, every Google Doc, all of our Salesforce data, Confluence—you know, it’s all linked up in one place. The data is aggregated, and you can ask these agents questions. Every team is using it—legal, finance, everything.

It’s like the Oracle of Coinbase.

Yeah. I’ve started to ask it more than just, “Can you write this kind of memo for me?” I’m asking these AI agents, as CEO, “What should I be aware of in the company that I might not be aware of?” It’ll tell me, “Did you know that there’s actually disagreement on this team about the strategy?” And I’ll say, “Actually, I didn’t know that,” because it can read every Slack message and every Google Doc.

I’ve been prompting it in a different way. Tobi, on my board, said this—he’s calling it reverse prompting. Instead of telling the AI agent what you want to do, you ask it what you should be thinking more about.

Right.

It’s a mentor.

Yeah. It’s like a coach.

Yeah. Like, “What could make me a better CEO?” It’ll say, “I looked at how you spent your time in the last quarter. Here’s how you said you wanted to spend it, but you actually spent 32% of your time on this instead of 20%.”

I’ve asked it other questions, like, “What’s the thing that I changed my mind on the most over the last year?” Things like that. It’s now becoming something that prompts you with information you should be thinking about, instead of the other way around.

I recently did this, and I don’t know if you’ve played with Claude Cowork yet. Have you played with it? It came out this week. There’s Claude Opus 4.5, or something.

Yeah. But there’s Cowork, which is kind of like—

You describe what you want to do as a knowledge worker, and it starts to build it. Instead of doing vibe coding and saying, “Hey, I want to write code for this,” you describe the end application and the result you want, and then a wizard takes you through it.

It’s pretty scary because I connected my Notion to it and my Slack to it.

And my Google Docs. It did the same type of thing. I was like, “Tell me about myself.” It was like, “Whoa, you need to spend more time with your founders who are winning, as opposed to more time with your internal team.” It was really interesting to analyze your teams, and I think that’s going to be the future of this.

All right. Listen, Brian, you’ve got a lot more meetings to do. Thanks for coming on the program.

I’m thrilled because my guest started building AI chips 6 or 7 years before ChatGPT was launched. Andrew Feldman is, of course, the CEO of Cerebras Systems, and they are building the Wafer-Scale Engine, or WSE.

Yep.

That’s the category of chips you’re working on, and they’re for inference—

For inference or for training—both.

Or for training, or both. And you have one with you?

I do.

So here is a Wafer-Scale Engine. Usually, chips are the size of a postage stamp.

Yeah.

This is 56 times larger than a B200.

Wow.

And it’s a 4 trillion-transistor part. For AI work, big chips process more information and deliver results in less time, so you get faster results for your query.

What does that cost for the typical system today, and how does it compete with the H100s and B200s?

Well, the first one cost us half a billion dollars to make.

Yes. The first one I’ve heard is the most expensive.

Turns out the first one’s the kicker, right?

Yeah. These come in a system.

All right. We can deliver the system on-premises, or you can use it in our cloud.

On-premises, they’re about $1 million to $1.5 million, depending on how you have it configured. In the cloud, you can rent it by the token. By the million tokens, it’ll vary by different models, from $0.50 per million tokens to several dollars per million tokens. Or you can rent it by the month or the year.

How did you know 7 years before ChatGPT was launched—or did you know—that the AI revolution would be this fast, furious, and unstoppable? Has what’s happened in the last 2 years surprised you?

For sure. I think anybody except maybe Sam and Ilya really saw it.

Yeah.

We talked to them in 2015, and what they were saying then—it was unbelievable how right they’ve been. But I think what we saw was the rise of a new computational problem called AI. It would put new and different pressure on a processor, and we saw this on the horizon. We said, “What would happen if this got giant?”

We had no idea how big it would get or how fast it would come. But as a computer architect, you try to think about whether you could build a machine that’s way faster at this new thing, and whether there would be enough of it to build a business around.

We saw AI on the horizon and asked ourselves, “Could we build a processor that would be unique in its performance?” Could we build something not 1, 2, 3, or 5 times faster, but 20 or 50 times faster? We came to believe we could.

We chose an approach that solved a problem that had been open in the computer industry for 75 years. Nobody had ever built a chip this big. Many smart people had failed. We delivered it, and it’s blisteringly fast.

What were the first applications? Nvidia got to perfect their compute—and really their company—off the backs of video game players, then Bitcoin and crypto. It was almost like there were a number of waypoints before AI emerged. You didn’t have those.

We didn’t have that. If you look at Nvidia’s stock price from 2004 to 2010, it was flat.

Yeah. [laughter]

Right. They were trying to find a new market. They had a lot of the graphics market, and that market was sort of flat. They found love with gamers. They tried to go into the supercomputing world.

We were focused entirely on AI. At first, we found love with the national labs, the military, and some pharma.

What were the applications they were using?

They were training various types of models.

Got it. And this is before large language models.

This is before language models existed.

So they were doing models for sequencing proteins.

Sequencing models.

They were doing different forms of vision models. They were doing work at the edges of high-performance computing and AI.

What is now taking the most compute? We see a lot of applications now: images, video production, training of the models, and really deep learning—deep thinking, I guess—where it’s firing off many, many threaded jobs. Which one of those uses the most compute, the most limited?

Right now, deep research uses an enormous amount of compute.

Explain to the audience what happens when they do one of those deep-research queries. As an example, I’ve been playing with the latest Claude, and they have a Cowork, a copilot-type application, and I made a prompt—

Yep.

—every time we have a guest on the podcast. I had it do maybe 15 or 20 steps: every podcast you’ve been on, every news item, a timeline, right?

Right.

It was unbelievable when it made this document. It’s better than anything a human has ever made for me, and I’ve been doing interviews for 20 years with lots of pretty smart assistants whose job that exact thing was.

When I tell you they would ask for 48 hours to do a dossier that was 20% of what this does in under 10 minutes, I’m not even joking. Last year, I told them, “You can use it to get ideas and get some links, but keep doing it the old way.” So it cut their time from 16 hours to 8. Now it’s 16 hours to, “I don’t need them.”

Literally, I don’t need them to do this work. Right. Now imagine if you could get it in 10 seconds.

Yeah. Right. That’s what we do.

Yeah.

That’s it, exactly. So what happens—remember, we make AI with training, and we use AI with inference. That’s the simplest way. The reason inference is going through the roof is because everybody’s using AI.

Yes.

A task that you kicked off starts a bunch of little threads, and each of those asks queries. Each of those queries delivers results that are the input to other queries, so you’ve got a cascade of queries going on.

It’s wild.

Each of those requires more compute. You have 20 different queries being kicked off, each query asks 20 queries, and each one of those requires 10, 15, or 20 seconds to get done in traditional compute. So you have this giant waterfall of time and answers. We built this part so you can get all those answers back in 4 seconds or 10 seconds.

When does that happen? Right now, it seems like when I do these kinds of deep-research queries, it’s grab-a-cup-of-coffee time, right? Five minutes, right? Not 15, but it seems like about 5 minutes is what it averages. When you’re doing an image or a 5-second video, it seems like it’s 90 seconds or so. When does that come down to the experience we had with dial-up going to fiber?

That’s the perfect analogy, right? When the internet was slow, Netflix delivered DVDs in envelopes. I know you remember this, right?

Oh, I do.

When Netflix got fast—when the internet got fast—Netflix didn’t get better at delivering DVDs. Netflix became a movie studio. It enabled them to be something different. It wasn’t a change in degree; it was a fundamental change in kind.

What speed does for AI is the same. We have customers like Cognition who use us to power their coding engine. If you read the tweets and people’s comments, they’re saying there is zero latency between their requests and their answers, so they can stay in the flow as they write code. This is the idea: you shouldn’t have to wait at all.

Anthropic is not a customer, but we recently announced OpenAI.

Right? They were original investors.

And now they’ve just put in a major purchase order. They have, and this is really exciting. Part of it, I think, was because we could deliver extraordinary speed so that the user experience changed.

As we know, having watched Google—Larry and Sergey, Marissa, and the team over there came to a conclusion: when we shave off milliseconds, it’s the number 1 way we get usage to go up.

That’s exactly right. They published that paper years ago that said even milliseconds—even amounts of time that the individual user doesn’t recognize as noticeable—

That’s exactly what it is. What is the psychological just-noticeable difference? I believe your mom would know. She’s a behavioral—

She would know.

Just noticeable. It’s 15% of whatever the number is. So if you could cut 15% off the time, people use it more and leave less. Yes. Paul Graham had a great tweet. He said, “I’d use Google half as much if ChatGPT weren’t so slow.”

If you think about that, that’s what happens, right? While you’re waiting for Claude or ChatGPT, you get a coffee or poke around somewhere else, and you’ve lost the customer. The cost of being slow is that the customer has gone somewhere else.

Or you do what I do. I have a nice, wide Dell monitor. I have 3 browser windows open. I pay for all 3 services—I have them all: Gemini, Claude, and ChatGPT. I pay for all of them. I’m paying probably close to $600 or $700 personally a month, so I’m spending $10,000 a year just for me, right?

I just take the same query, go bing, bing, bing, bing, and start them all. I’m probably burning like 10 trees. I mean, that’s probably being a little greedy.

It’s not 10 trees.

I think that’s a really interesting way to manage how slow it is, right?

That’s why we exist: to fix that problem. What we partnered with OpenAI to do is deliver blisteringly fast speed across the world’s most popular models.

What’s the scope of the deal?

What we announced was 750 megawatts.

When did we switch from talking about the number of chips, the number of units being sold, to the amount of power being sold? It’s a little bit confusing for folks, and it started probably about last summer.

Actually, the change has been coming for a lot longer. We used to talk about data centers in terms of square footage: “I’ve got a 100,000-square-foot data center,” right? Now nobody cares how many square feet you have; they care about how much power you have.

The limiting constraint on data centers is always their power footprint. Right now, for large deployments, the limiting constraint is how much power can be delivered. By talking about how much power is delivered, you’re talking in the unit of the limiting constraint.

Mm-hmm.

The limiting constraint is power. We’re trying to find this huge amount of power for OpenAI. It’ll be delivered over several years.

And you’re responsible for the power as well, or is that a joint effort?

It’s a cloud deal.

Oh, so they’re utilizing your cloud. So you’ve got to do all the work.

We are building the cloud infrastructure for it.

Got it. Where are you building your data centers? What’s the best location here in 2026 to be placing these things? Is it natural gas? Is it near hydro? What’s the state of the art now?

The cheapest power in the world is hydro.

Yeah.

Without question. After that is natural gas. Where places have natural gas, you have an abundance of relatively low-cost power.

Which is Texas.

West Texas, Wyoming, outside the US, in the Caribbean, and in Ghana, you have a huge amount of natural gas. You also have geothermal, which is its own thing in the Nordics.

Natural gas is a very inexpensive way to get power, particularly if it’s coming as a byproduct from petroleum mining. What you have is what used to be called flare-off gas. They used to just throw it away. They used to just burn it at the top. Bitcoin miners found that, right? And so we’ll take that.

That’s right.

Yeah.

So basically, you look for the existing flare-off.

Tell me about hydro, because it does seem to me that people knew for a long time that they were moving data centers there. Is heat still an issue with your chips and others?

We’re water-cooled, and water is an extremely efficient coolant. We knew early on that we’d be going to water. We were some of the first production AI systems to use water. The TPU moved to water early on as well. Before that, there had been some water cooling, mostly in the Department of Energy supercomputing labs, where they used some water.

Does it matter? I remember early on, when people were talking about water to cool things 10 years ago, the source of the water and how cold that water is coming in. Or is water just cool enough?

No.

And you’re fine?

It depends on your particular design, but you’d like cooler water.

Sure.

Cooler water is better. Alaska and Canada feel pretty good about that. Or you bring chillers, or you cool the water, right? You can often take general groundwater or other forms of water from other locations.

There’s a huge misperception today that this water is not recycled and that AI is using all this water, when it’s not even comparable to golf courses, let’s say.

First, golf courses are extremely water-inefficient.

Most of our data centers use a closed loop, right? We're passing the water by the back of our chips. They pull the heat off, warm the water, and the warm water goes down through a closed-loop system, is chilled, and pumped back. So you're not using new water, and the water is not damaged.

The water's not damaged. It's just not like some chemicals or anything that gets put into them.

No, not at all.

There's a lot of misperceptions about AI right now. It seems like there are almost some dark PR forces at work trying to make the data center build-out seem worse than it is. Then there's also, I think, maybe some valid concerns around jobs. When you look at each one of those issues, what do you think are the ones that are most frustrating as an AI executive building data centers?

It's a really good point. I think some of the hyperscalers made a bad call in the way they went into some of these rural communities. You're looking for a place where land is cheap and there's an abundance of power, and they went to these communities without doing a good job talking to people.

Right. Tech people didn't do a good job talking to humans. What a surprise.

They went into these communities and cut deals with the power company. The power company was looking to build new infrastructure to support them.

Yeah.

Traditionally, the regulated power industry would then amortize that cost over 20 or 30 years. So they ended up increasing the local people's power rates, right? The people got upset, and that's very reasonable. If instead you'd gone in and said, "Look, great, we're going to be good citizens. We're going to be big taxpayers here. Let's build more schools. We can build a school for you. It's a rounding error in the cost of this facility. We're going to make a bunch of construction jobs, and we're going to be good citizens."

Yes.

They would have had a very different approach.

And not only that, they were a little heavy-handed early on, saying, "We're going to play 3 communities off each other. Who's going to give us the biggest tax discount?"

Right. That was another Silicon Valley mistake. Yeah.

Now, what I just saw is that Microsoft put out—we talked about it this week on the show—I thought it was a very thoughtful and reasonable approach for a company that's sort of a national champion. They're going to be good citizens and want to make sure that your rates don't increase. They basically said, just to catch the audience up, "We guarantee you our usage of energy will not increase the cost of your utilities."

That's fair. I mean, very reasonable.

I think the next step—we were brainstorming on the pod—is that there are people putting solar on roofs, and there's Base Power. Michael Dell's son is doing a really interesting project.

Yeah. They just put batteries on the side of your house. They don't have to be super intricate; they load those batteries up when there's extra power and it's cheap, and they deploy it when the duck curve, or whatever, demand hits. If you think Microsoft—and you talk about rounding errors—if you give everybody a battery at home to store some energy when it's cheap, that can be used to flow back into the data centers. I think we could live in a world where you say, "Hey, we're going to put a data center here, and everybody's energy is free."

I think—well, there are a couple of things. First is, we chose as a nation not to invest in our grid for 40 or 50 years. Our grid is behind and vastly in need of improvement. Our grid is decrepit compared to other advanced nations, and in particular compared to what China has done.

I think the ability to store power at your home and use it when power is the most expensive is an obviously reasonable thing to do, right? Obviously, it's a reasonable thing to do, and it takes load off the grid as well. People like the idea of being a little resilient, right?

They do. If you do lose your power—which in California, I think they turn it off on the Peninsula about half a dozen times a year for you?

Only when it's really hot or really cold.

Either one.

Yeah.

And they'll leave it off for 2 days because of wind, and it's just a complete disaster. Oh, and by the way, I don't know if you knew this: There were subsidies given for nuclear where people living around nuclear power plants in France were told, in exchange for living near a nuclear power plant—which some people might have concerns about, maybe they're reasonable, maybe they're unreasonable; put that aside—"We're going to just give you free energy for life." Very interesting. What is your thought on small modular nuclear? Is it just too far out for you to be concerned with right now?

I think it's both obviously the right thing to do and probably not the source of power for data centers for the next 3 or 4 years.

And if you had the ability to do one, you would do it.

Oh yeah, for sure. You're seeing some of that in the more aggressive nations. The UAE is building modular nuclear power plants to support data centers, putting huge amounts of power on the grid with nuclear. What a great idea.

I went to see Elon a couple of weeks ago on a Sunday afternoon, and we were talking. He really thinks that putting data centers and chips in space—cooling is pretty easy in space, and solar is much more effective. What do you think? He's been talking about this publicly, so I'm not speaking out of school. What do you think about data centers in space? Have you started researching it?

We have. I think, first, betting against Elon's ideas is probably not a good long-term betting strategy.

He's never wrong. He's frequently late.

That's right. I always look at that. I think that's both the blessing and the curse of being a visionary: You see things other people can't see, and in your mind they're just some technical hurdles to overcome.

You took a couple of years to build that.

It took a couple of years.

Were you on time?

We were plus or minus a year.

Okay. In the delivery of something that nobody had ever done.

It's hard to predict. It's really hard. I think the idea of using space to grab solar power is obviously a smart idea.

Yeah.

You're miles closer to the sun. You have much thinner atmosphere blocking the rays, so you can gather up the power. I think there's a lot of technical work to be done. Yes, it's cold there, but you're also in a vacuum, so the actual cooling isn't an easy problem. It's a solvable problem. I think communication among satellites is a real, real issue, and figuring out which technology you want to use to get the data back to Earth is another issue.

Remember when you tried to get internet from those satellites? It was really glitchy.

Yeah, there were these big delays. Those were higher-orbit satellites. The ones he's thinking about are much lower orbit. They'd have lower latency, but there's some real work to be done. I think it's in the 8-to-10-year category, not in the 3-to-5-year category.

Yeah, I think it may split the difference. All of these are worth pursuing if you believe that we're not going to overbuild. So, knowing what you know and watching this build-out, is it possible that we're overbuilding right now and we'll need a digestion period, or do you think, based on what we're seeing, there's just going to be the next workload, next workload, next workload?

I think we're still really early in the demand for AI compute. If you think about what portion of enterprises have really adopted AI in a meaningful way that has changed their workflow, it's tiny. I think even the most frequent users at the consumer level are going 6 to 8 times a day. What happens when they go to 100 times a day? What happens when all their devices are working for them? What happens when everybody in G&A, when every engineer, is using it as a coding copilot?

We're going to see enormous amounts of demand for inference. The models are getting better, more people are using them, they're using them more often, and the amount of compute taken with each usage is increasing. So I think we're just at the beginning.

How do you portion out the effort when you're making systems right now, in terms of energy efficiency, raw horsepower, and then the transport layer? These seem to be the 3 most important parts of what you're doing. Correct me if I'm wrong. How do you allocate your engineers and your overall team to tackle those 3 major issues?

One way to think about it that I don't hear often enough is this: The way you make a computer is, you think about 3 things, right? How fast you can do a calculation.

Where you can store the result.

Mm-hmm.

Memory, and how fast you can get the result to somebody who wants to use it.

Transport—

Not transport. These are the 3 things that make a computer. If you do really fast calculations but your storage is slow—

All right, bottleneck.

You're bottlenecked. If you can do fast calculations and store it, but your I/O is slow, then you can't get it to the user.

As a computer architect, you're constantly thinking about the balance—

—of these 3 dimensions, right? You make a jump in the performance of calculation, and you've got to think about storage. Then you've got to think about the other dimensions. I mean, it is a constant.

Are you thinking about those 3 simultaneously, or are there teams grinding out each one of those individual verticals? How do you architecturally build a group of engineers to do that?

Usually, your most senior architects—your CTO and your technical leads—are thinking about that as the basis of a design, right? It doesn't matter how fast the car can go if it can't turn, right? It's not a good car, except maybe for drag racing.

The designers are constantly thinking about where we should use power in the design, what we can make faster, whether we can add memory, and what the cost of adding memory is versus making it faster. The GPU, for example, has a lot of memory capacity, but it's really slow. That's a huge bottleneck in inference. It's why they can't be fast. It's why they just spent $20 billion buying Groq: because they didn't have an answer for fast inference. Fast inference needs fast access to memory, and the GPU doesn't have it.

These are things we're constantly thinking about.

We're having a massive memory shortage right now because of this. How does that get resolved? Is that just a short-term bottleneck, or is that going to be a long-term problem?

I think it's a crazy problem. Everybody knew that demand would increase, and this is true among the major memory makers. People get a little scared, and what happens is they place a full year's worth of demand and get the wrong answer back, which is, "We don't exactly know when you can have it."

So then their response is, "All right, we'll give you 18 months of demand." Suddenly, everybody went from giving 6 months of demand to 18 months of demand.

Okay.

Everybody—the entire supply chain—is confused. We're making the exact same amount of memory now as we were 4 months ago. What's happened is that the signal to the makers has exploded, and it will take us about 18 months to digest. The prices will stay high.

This is a known phenomenon in the memory market. It happens every 6 or 8 years. What's different right now is that the GPUs are using a huge amount of HBM, which is a flavor of DRAM. They're chewing through that, and that's maybe leaving a little less for other devices and consumers.

How far along are the Chinese in catching up to your company, NVIDIA, and Groq, and how do you think about the geopolitics of the AI race? Is there a scenario where they win and we lose, we win and they lose, or is that overblown in your mind?

I think the geopolitics are a real issue.

Okay.

We are well ahead in chipmaking.

Okay.

Within a few square miles of Santa Clara, you had Intel, AMD, NVIDIA, our team, and ARM. You have one of ARM's great teams. You have amazing talent. You have 6 of the world's great 10 chip teams.

I think the way you get good at building high-speed chips is to build high-speed chips. That's really how you do it. You play the game, and you get better at the game.

That's right. It turns out you get better at the game.

That's been a weakness in the Chinese chipmaking ecosystem. Now they're running hard, and they know they're behind on that. On the other side, I think they have pushed ahead in the open-model category.

Yes, the open-source model.

The open-source model is an area where they've pushed ahead. I think that, because they're a top-down economy, they were able to make decisions like, "We're going to bring a huge amount of power onto our grid. We're going to modernize our grid." They were able to bring on huge amounts of power, and that's something that we're behind on.

I think it's unpleasant to think of them as adversaries, and we've got to figure that out together. The world is a better place when we're not adversaries, but right now we are.

I think certainly in an industrial context, we're adversaries. There's the industrial context, and then there's, as we discussed, what impact does this actually have? What's downstream of us winning? It's every developer being a 100x developer, every knowledge worker being a 100x knowledge worker, and every biotech innovation.

Systems that are recursive—that build on themselves at rapid rates—have a huge winner-take-all feel, right? By getting ahead, you get further ahead. Your iteration speed accelerates, and even small differences at the beginning are magnified very quickly. That's why this race is so important.

Right.

So here we are. We're at Davos. It's a lot of politicians. My friend David Sacks, co-host here on the pod, is our AI czar. Trump, whether you voted for him or not, is very focused on this issue. Biden and his team weren't courting Silicon Valley. In fact, they looked at us as the problem and demonized us to a certain extent.

How do you think, objectively, independent of how you might feel about ICE agents in our cities or Greenland, how do you think the Trump administration is doing on its AI policy and the support it's giving the AI industry?

I think on a lot of fronts they're doing really well. I think—

Unpack it.

I think we made a mistake in the previous administration by keeping our chips from our allies. Let's keep China separate for a second, but the UAE is clearly an ally.

Absolutely.

As an ally, right? It's a modern Arab nation, has been a source of peace, made peace with Israel early on, and has huge Western influence. We kept chips from them, right? We'd like KSA, the Kingdom of Saudi Arabia, to move in the same direction, and we kept chips from them, right?

We then made a hierarchy that made the Danes feel second-rate, right? We said, "You are a number-two friend." Bad idea. We should be empowering our allies.

That's the first thing, and I don't think the previous administration did a good job. They didn't understand that at all, and Trump did a good job of that. Not only do we want those nations and their institutions using our technology, we want them investing in the U.S.

Under the previous administration, we had a CFIUS organization in the Treasury that was difficult to work with, and all of those are much improved.

They were unclear. They were not communicative.

They were impossible to deal with.

Impossible.

Impossible to deal with.

This is super important because, as David has said many times on this program, we want to be the standard, right? All of that energy goes back into our standard, into our ecosystem, into development on top of us, into the recursive system we just described.

Exactly. If you look at Huawei and what they did with 5G—their networking up against Cisco and our national champions—they ran the table in a lot of countries. They clobbered us in Africa. They clobbered us in the developing world. They absolutely ran the table.

Right. And now those places have spyware.

That's exactly right.

It's a real issue.

It's a real issue.

So I think those were all areas where this administration did absolutely the right thing. Energy.

Energy. Another area: I think one of the things that kills a company like us, when we're trying to grow extremely quickly, is having to deal with different regulations in each of the 14 localities where we're trying to put data centers.

Right.

That is brutal. What you don't want when you're trying to grow really quickly is to have 17 lawyers, each of whom is trying to figure out the local regulations. Trump's effort to say, "Look, let's get some reasonable laws across the board," is obviously smart. If we could get some money to improve the grid across the nation, that would also be really helpful. All of those are extremely positive.

The work he's done with the Department of Energy—

Chris Wright, right?

Right. Under—I think it's called the Genesis Mission? I think it's sort of the equivalent of a Manhattan Project for AI. Of course, we need this. Of course, we need to be thinking among our researchers not about how we can get a little bit faster—10 or 20 percent—but what we can use AI to do to increase the rate of research by 5x or 10x, and how we can get the things that impede government out of the way.

Yeah.

Right. Those are good. China's a really sticky problem. I'm not sure I agree with the current push to allow the selling of H100s there, but it's reasonable to disagree with me. I don't think it's clear-cut like some of the other issues at all. It's a hard problem, and there are going to be lots of different views there.

Yeah. I don't know if you've been watching the news, but Canada just made a strong alliance with China, announced, I think, yesterday or today, when we're taping this.

And this is where maybe the Trump administration can improve: We do need to maintain this alliance with our neighbors so they feel like they can trust us. This is what I've heard spending time in Japan, where the Japanese feel like maybe we are not the most reliable partner. Canada feels we're not the most reliable partner because of the tariff issues, military issues, and maybe just the constant changing of policy. Consistency is really important.

And I think for a country like Canada that has a huge amount of raw-material exports—they have wheat, they have lumber, they have a huge amount of stuff that either we import or they have to take elsewhere—we have to be aware of the realpolitik of the situation. They have to sell their raw material, which is a huge part of their exports. They have to sell it somewhere, and China is a big buyer. We have to go in understanding that there are nations that are proud, and if you're constantly attacking them and saying things to the populace there, it gives the leaders the ability to say, “Well, hey, China is courting us, and they're going to invest, so why don't we build some ports with them?”

Yeah, ports. You should do a whole show, if you haven't already, on the rise of Chinese ownership of major ports and shipping. It is crazy when you look—I mean, basically, they own the world's large shipping routes.

The Belt and Road strategy.

And then if you get out of—I don't know, I'm not here in Switzerland very often—but if you go into many parts of the Third World, you begin to see BYD cars.

Yeah. They're going to be shipping them to Canada now.

All over the rest of the world. We don't see it, but it's unbelievable.

I was just in Mexico City with the wife for a couple of days.

Isn't Mexico City fun?

It's my first time there. I had a delightful time. The food is spectacular. I love it. Really fun, trendy, great place.

Yeah. Good vibes. And every car is a BYD.

Yeah. And we've got to think about that. They talk about national champions. There's no way that the government isn't subsidizing those by 30%, 40%, 50%. I think their goal is to put the Germans out of business. The English car manufacturers have been at it for a while, but the Germans are still making pretty great cars. These BYDs are starting to get footholds in Europe, and the same thing will happen. Who's going to buy a $40,000, $50,000, or $60,000 BMW, Volvo, or Audi when you can buy a $20,000, $30,000, or $40,000 BYD? They're nice cars.

They're price-dumping, though, and that's what tariffs are meant to protect against.

They're subsidizing at the top of the finished product.

Right?

And that benefits the whole supply chain, right? That's what they're trying to do. They think about it as the battery maker and the transmission maker—all are benefiting while they subsidize at the very top.

All right, let's end on employment. Let's put the crystal ball out there. Microsoft, Uber, Coinbase, Meta, and Google: 4 or 5 years ago, did they have more employees than they have now, or are they flat?

Yep.

Youth unemployment is starting to hit 10% to 20% among some college-age demographics.

Yep.

David Sax and I have this debate all the time. Is it AI? Is it entitled kids who don't have a work ethic? Is it the overfunding and the digestion—or indigestion—of tech companies that hired 2 years out? It's pretty clear to me, watching startups that are the most resourceful, they're doing so much with AI. They are AI-first. They're building agents. They're doing everything with AI.

There's no world in which we're not going to have AI displacement.

Job displacement.

That's not why it's displaced now, but 100% it's coming.

Okay. So when you look at it, you're in the camp of it's coming, but it's not an issue today. Define when it's coming. Why it's not an issue today is when I look at the people who have been let go, in middle management in particular—

Okay, be candid. Where are you on AI?

No, I mean, this is middle management. What I think has happened is this is the delayed impact of good SaaS tools.

Ah, that's what's happened: Your ability to extend your reach as a leader and as a manager, to stay abreast of what's happening—your scope is much, much bigger. And so the role of middle management, which was frequently to move information—

Yeah.

—to manage small teams and move information—

Keep people on track.

That's right. That job has shrunk in value. I don't think it's yet AI.

I think we're halfway there.

That's right. I think AI is coming, but I don't think that's what this is. What happened was there was this ballooning of these jobs, and Mark Zuckerberg and Satya Nadella looked one day and said, “Competition is coming. It is much more intense. What are these waves of people doing?”

They're slowing us down, let's be honest.

That's right. They're flattening their organizations as well. So it's not just that they're moving people out, but they're changing the shape of the organization, which is why I don't think it's AI yet.

Yeah. I think what we're going to see down the road is whole categories that are vastly more efficient and therefore need fewer people.

It's pretty clear, and it's almost—rest in peace, Scott Adams, creator of Dilbert, but—

Huge fan.

Yeah, he just passed away this week.

I saw that. Huge fan.

What a giant.

What a giant of ridiculing corporate America. That exact layer is gone.

Corporate America. It's actually great that Scott got to see it.

Yes.

Happen. Yeah. Towards the tail end. And we didn't get to mention it on a previous episode, but rest in peace, Scott Adams. I think it's a good place for us to end there, Andrew. I know you've got a lot to do here. Enjoy your time at Davos.

Thank you.

Yeah. If you see any of the Germans, ask them why they turned off their nukes.

All right.

Yeah, that's right there to joke to them. How's Greta Thunberg? How is your secretary of energy doing, Greta Thunberg? What are you doing? They turned off 3 of their 6 nuclear reactors.

I know. So they decided instead to import natural gas from Russia.

Where did they get it from?

Right. From Russia.

Oh, from Russia.

Yeah. Bad. Dependent on Russia. Really not smart.

Yeah. Not smart.

And you know what? All because of Davos.

I blame the WEF and Davos. They literally got so caught up in virtue-signaling about the environment that they never just looked from first principles at how safe nuclear is compared to burning fossil fuels.

Nuclear is safe, and we can make it safer. We've got to put the time and effort in. I was just in Japan last week.

They're putting in new nuclear reactors, and they just got over Fukushima and realized, “Oh, we made some mistakes putting it below sea level. We're not going to make those mistakes again. Nuclear is obviously the way to go.”

Let's get better at it.

Hey, let's get better at it. That was awesome, dude.

Thanks for all the time and for a great discussion. You rocked it.

A friend of the pod, Jake Loosararian, is here. You've been on the pod before, both This Week in Startups and All-In. You're, of course, the CEO and co-founder of Gecko Robotics. You've been at it for almost a decade now. You build robots, as people who have seen the pod before know, that inspect ships, bridges, whatever it happens to be. And you started this long before ChatGPT and this recent AI revolution.

I'm curious: These robots, which are very purpose-built and, I think, very straightforward—have you started to put AI into them yet? I was just curious, thinking about your previous presentations. It was pretty straightforward, right? We know the bridge, inspect the bridge. But now can it do things and start thinking on its own, and maybe be more adaptable because of AI?

Yeah, good question. Well, I changed my title. It's now chief grifting officer.

Oh, chief drifting officer.

Thank you. Especially when I'm in Davos, this is my title.

You've been here a couple of times.

Yeah, exactly.

Did you catch the tail end of the DEI?

I came right at the heart of it. You had to learn a different language, actually.

Really? Did they check your fluency in ESG and DEI buzzwords?

They did. They did.

Well—

No, but everything was super precious, and now I guess since Trump is here, it's kind of brass tacks: doing business, negotiating, and less of this performative stuff. It's a lot, actually, performance tonight. But there's a lot more focus on, okay, let's get down to the brass tacks.

We hear a lot of CEOs talking about AI, but actually a lot of the conversations I'm having in Congress already are just about, okay, where's the ROI from all the AI business?

It's actually business. It's actually trying to get to the first principles, to the roots of, okay, how do you actually get artificial intelligence to deliver on the promise? Funny enough, a lot of it comes down to this really interesting gap that exists in AI, which is all the information and data sets that you need to actually turn all this into actual return on investment and productivity gains, especially for these large infrastructure and large asset owners, like the energy, mining, or manufacturing companies of the world.

So that's a big focus, and that's what—

It really seems to be turning into a business conference. I was astounded by the amount of inbound I had that was just pure business capitalism, building products and services to make life better. Also, the world has changed a lot since you started the firm. We've got a new sort of Military 2.0 thing happening, and I think a lot of your customer base moved from just maintenance of bridges, tunnels, and infrastructure to the military. Tell us about that.

Yeah, that's exactly right. We do about 30% of our business in defense. So we work with the Department of War, I guess I should say. A lot of it is focused on how you actually use technology to fight against the speed of development of countries like China, for example, in terms of manufacturing speed. A big part of that is understanding the quality of the welds—the putting together of the actual welds that put pieces together.

We have these manufacturing facilities and forges that are 100 years old, doing things in a 100-year-old way today like they did back then. The technology that we're deploying helps manufacture certain components of a submarine and expedite how fast a destroyer can turn around to get out and patrol borders and deter conflict. These are the things that our robots are being used to do: speed up the decision-making process and make sure you're accurate.

In some cases, Admiral Houston has talked about 90% improvements in manufacturing speed using the technology that Gecko builds. You're seeing companies like Anduril now working with us. Palmer Luckey.

Palmer Luckey. Yep, of course. That's his helicopter up there. He might drop a bomb any minute.

Yeah. I think he's doing a speech pretty soon. It's just amazing to see the adoption.

But on the energy side, that's been the biggest growth area for our company. It's been these large energy and power companies trying to figure out how to get infrastructure. All these hyperscalers are trying to figure out how to get infrastructure. They're focusing on capex a lot, right? What if we started to play a game where we have access to these problems, to these really GDP-driver companies? What if we actually took an AI-native—or, in our case, what we've seen a lot of companies be is robot-native first—approach to support the AI initiatives by aggressively implementing and putting robotics to use to help build up the data infrastructure, then layer on AI models?

That's at the heart of what Gecko does. That's why I started this company 13 years ago, with this premise: data matters as it relates to being able to have all the gains.

For people who don't know, the robots have sensors in them, different arrays that can inspect metal—whatever the fabrication is—and go right to the seams of a submarine and make sure it's all been done perfectly and measure it perfectly.

That's right. We build the robots and the sensors that go around and diagnose the health of the built world. That means understanding and getting the largest inventory and database of information about the health of built structures—bridges, dams, submarines, whatever it is.

Along that journey, you're able to figure out that if you centralize all that information and data, and then layer on top of it operational data—which exists, for the most part, as a decent infrastructure of sensor data at these companies—well, wow, you get to make some pretty interesting decisions. You can figure out how to extend the useful life of an asset. If I push an asset harder, can I produce more? My focus is: how do I help create cleaner, as well as more, barrels per day and at lower costs? I use the word “cleaner.”

So if you have a refinery or a nuclear power plant, you inspect it.

Robots should be dedicated to figuring out the business problem. What is the fundamental business problem that the customer is trying to solve? If it's making a barrel, making a kilowatt, or getting a ship out of dry dock faster, that is our initiative and our goal as a company: to build robotic solutions toward that.

We haven't gotten into building humanoids or playing the humanoid game. When I was on your podcast, actually at the summit, I talked about how we're going to be the biggest purchasers of the Optimus robot. The real question is, how do you actually employ robots? How do you get robots to return ROI? Folding laundry and cleaning dishes is not a high-ROI use case. It's going to be the $20-an-hour—

You're not going to pay $40,000, or $20,000, whatever it is. But the U.S. has to be the best in the world at figuring out how to use robots to create unfair advantages with these companies, whether it's oil and gas or power. You need somebody between Tesla, Figure, or Boston Dynamics. Those robots are going to be sold. There's going to need to be an application layer and operational excellence in the field.

That's exactly what we are: a nervous system to pull all this information from robots together. Then you can build and use AI models on top of that to use the information and data to begin taking actions back into the real world.

So you actually see it—not just finding problems or monitoring situations and confirming that things are being built properly, that there are no potential problems with this nuclear power plant or this ship—but down the road, do you see yourself actually taking actions to build and repair?

Yeah, that's exactly the road map for us. But first, you have to figure out what the state of the built world is and what the state of its health is. What sorts of actions should I take when it comes to repair? What sorts of automated welding solutions, for example, are the right ones? Which ones could use information about how well that weld was done as feedback to create a foundation model for welding—to be the best in the world at welding?

We're going to be the company that builds robots to both identify and solve the most important and highest-ROI problems for customers, whether they're manufacturing new assets or trying to operate and maintain existing ones.

Funny enough, I've been talking a lot about how to reduce hazardous work hours for humans, how to extend the useful life of assets for infrastructure, and how to increase capacity and production and prevent and catch failures in assets. These are all very easy-to-underwrite problems, and it's something that you don't hear roboticists or AI founders talk a lot about. But that's my bread and butter. That's the world I live in, and I wear the steel-toe boots to understand the problems.

You're going to need humans in the loop for some time to come, and there are going to be plenty of jobs for welders, but there might also be incremental jobs created because—

Yeah.

—and you'll have one welder maybe supervising 10 of these robots. Is that what you think is going to happen?

That's what's going to happen. You want to be able to get the experience and subject-matter expertise to make sure the robot is actually understanding the ramifications if I do this action versus that action. You also want to understand that there's going to be a lot of teleoperations, particularly with mobile robots. You're going to have humans in the loop; they just might not be in the field as much. They might be more in an air-conditioned building, being able to operate and build information and data to train the foundation model.

If you think about how risky some of these jobs are, it might be nice not to have a human risking their life to maintain this part of the bridge.

You know, as brave and amazing as it is, they're doing that work, and we obviously appreciate that over the centuries.

Yeah.

It might be nice to actually take the human out of the deep-sea welding and out of the bridge-climbing business.

Well, I mean, the story of Gecko has been a story of building robots to help reduce the barrier to entry for these jobs that sometimes take 10,000 hours to be great at, and actually make them something you can attain within a few months of being able to use the technology in these fields.

Whether it's manufacturing different parts and inspecting the quality of those parts, or actually gathering information and data and understanding what kinds of decisions to make, my goodness, you have a shortage of welders, a shortage of inspectors, and a shortage of all these trades. You have to be able to augment—to take a Home Depot employee and, in a couple of months, make them able to make $100,000 or $150,000 running your robot and doing it safely.

That's an exciting, bright future. I think the key unlock for us in the robotics community is that you have to get your robotics into the field. You have to fail fast and also rapidly prototype really quickly.

And then manufacturing them is the big issue. As we focus on these sorts of problems, over the next 5 years, our goal is to be the company that's the best in the world at taking robots and making ROI from them. That's what we're focused on in terms of setting ourselves up to be the world-dominant company there.

And you wrote an editorial on the way in here and dropped it. What was your take in the editorial?

It was basically on the concept that we're talking about here. I live in Pittsburgh. At some point, Pittsburgh had more millionaires in 1930 than New York. It made 70% of the world's steel.

The industries and the companies there are still the backbone of our economy. I was inspired by that time frame, the Industrial Revolution, and how steel was invented, then manufactured and distributed to help create all the infrastructure that we rely on. That was the infrastructure you needed to be able to have all these big gains that came from the Industrial Revolution.

The same thing is what I was talking about in the editorial: what we're doing with robotics—collecting information and data sets—to help support and create the infrastructure for AI models to actually be able to return the kinds of returns that we're all betting on. It's important for people to understand that robotics is almost the foundation for being able to get the massive returns in the sectors that we're all mostly here for.

Great. Yeah. Being able to eliminate some Dilbert-level middle managers who aren't adding value—

—with some automation.

Okay, fine. But we really need to get out there in the real world to get that serious ROI, whether it's a robotaxi or a self-driving car. I think the risk you have is that this forum is changing, right? There's an ecosystem in a bubble when you live in a certain place, talking a certain way. In Silicon Valley, we only exist in the world of the internet. We don't build the kinds of technologies that started in Silicon Valley.

This world of energy, metal manufacturing, mining, and defense—those sectors just aren't part of the conversation. When I was starting the company, they were taboo to talk about. So you just don't think about the kinds of applications and things you can build.

In some ways, we ran out of things to solve for. I mean, what's next? When I would be pitched 10 years ago on SaaS software, it was like, “Okay, great.” Then it was, “This is the 50th SaaS software company in this vertical.” Then it was, “This is the 15th in this vertical.” We're kind of running out of—

What's crazy to me—

—those verticals to go after.

You're exactly right. This is why you think of an incredible invention like a humanoid robot, and the first demo that you and I saw was folding laundry. Oh my goodness. That was the thing that I do when I go home, so that's what a robot should do.

If you think about it, we had Boston Dynamics doing backflips with these robots a decade ago, but they didn't have an LLM behind them or a vision model or a world model yet.

Yeah. And now when they have it, you'll be able to, I think, tell it, “Hey, I want to lay some bricks,” and it'll just go out to the web and find all the bricklaying YouTube videos, the history of bricklaying, every manual on bricklaying, and every SKU of every device ever used for bricklaying. It's going to know—

How to do it.

Without ever having to be trained—or is that the—

Yeah, that's right. I think that the—

How soon?

How soon? I don't think that's going to be as far out. I take more of a 3-year time frame for those kinds of things. You can see these big bets. SoftBank and Nvidia just put a billion dollars into Skilled AI, which is creating the brain for robots. It's actually at a $14 billion valuation. The founder, Deepak, is in Pittsburgh, by the way.

The big problem with that extrapolation is that, in the world that I live in every day—energy, defense, and so on—we don't have those videos. There isn't a corpus of information and data sets. I'm focused on that. I'm focused on—

How do you get that data? Do you put GoPro cameras and sensors on people's arms, like I saw?

Is that how the training will be done, or are you modeling and actually watching a human do it, then having the robot analyze it?

Yeah, we think about the fact that not many customers are going to pay for that because the ROI just isn't clear; it's not there. No big energy or manufacturing company is going to say, “Yeah, let's do that, and I'll pay you $10 million a year to do that.”

So we're collecting it by solving important problems on critical infrastructure and assets. We're walking around these Manhattan-size refineries all the time, and there are information and data sets that we're building.

And the refinery inspection is done by a human today. Yeah.

They take a bunch of pictures. They use a bunch of sensors, and now the robot is 100 feet up in the air on a rope, collecting data by hand. If you use a robot that has a bunch more sensors to fuse together, it begins to create a world that doesn't exist on the internet, which gives Gecko a very big advantage.

That's a lot of world-building you're doing.

That's exactly right.