走进科技投资者 Gavin Baker 的投资思维

Jas

Our guest today named his hedge fund of traders, which, for the non-nerds in the room, means the spice must flow. For those of you who remember spice from your freshman year in high school, it's not that kind of spice. For Gavin, spice is high-bandwidth memory. His Arrakis is Taiwan, and his sandworm is Jensen Huang.

His X bio reads, “No investment advice; views my own,” which may be the most expensive disclaimer in modern finance because he's got 288,000 followers who follow his advice anyway. He ran 17 billion at Fidelity. He beat 99% of his peers, although he just told me it's actually 100% of his peers. He now runs his hedge fund from Boston, which may be the most contrarian position of all. Ladies and gentlemen, the most knowledgeable man alive on semiconductors, please welcome Gavin Baker. Wow.

Thank you, and let me introduce you, Jas.

Jas

You don't have to do that.

Jas is one of the most senior partners at Blackstone, and he just took over a new job heading all of Blackstone's AI strategy. He was formerly the head of their tactical opportunities strategy.

Jas

We've been friends for a while, so it's fun that he gets to interview me and I get to interview him. But Gavin, I do get to interview you.

1. Baker Explains His Investor Mental Model

Since this is an investing conference, I wanted to start with your mental model as an investor before we get into the meat and potatoes of semiconductors, memory, the stack, and the bottlenecks, which I'm sure all of you, just like me, want to engage with you on. You started at Fidelity in 2000 and managed it through 2 bubbles. What is 1 piece of Peter Lynch-era Fidelity DNA that you've kept, and the 1 that you had to break in order to start, run, and be successful with the traders?

Good question. I'd say that Peter Lynch DNA that is forever woven into me, as I suspect it's woven into anyone who's had the great fortune of starting their career at Fidelity, which was an amazing place for me, is this concept: if you like the store or you like the product, you're going to love the stock. The importance of engaging with new products and with companies as deeply as you can as a consumer and user—that for sure is embedded into me.

What I have been working on my entire professional career as an investor is this: there's a stock market axiom that you want to—Peter Lynch said—you want to pull up your weeds and water your flowers. You want to sell your losers and ride your winners. For whatever reason, that is very hard for me. I'm extremely valuation-sensitive and very contrarian. I'm most comfortable on the 52-week-low list. I've been hanging on to the memory stocks for dear life. That's been a lifelong journey that I try to get a little better at every year.

Jas

Most great investors point to a single trade—a single bad trade—as the formative one in their own journey. What was yours, and what specifically does it stop you from doing today that you would otherwise do?

I would say there are 2. I had a really hard year in 2011 or 2012, and it was 2 stocks.

One was a company called Accredo Health. They were helping small hospitals do a better job of negotiating with big insurance companies. You feel like it's a good cause. Then you wake up and there's a front-page, 10,000-word article in The New York Times on Sunday about how they're denying care to people in need, and this is simply not true. Nonetheless, the stock went down 90–95% and never recovered. There are always unknown unknowns. There are always risks that, no matter how much you try to dimensionalize an opportunity, are present.

The second one is embarrassing: I was in Nextel International. I was a telecom analyst in the past. For a long time, you could make a lot of money in telecom. Whenever a new network was built in an emerging market, it was definitionally the lowest-cost, best, and highest-quality network because it wasn't loaded.

Nextel International built a new network in South America. It was the latest and greatest 3G network, and they were migrating off something called iDEN. I'd made money in this kind of pattern many times, so I made it a large position. I'll never forget: I wrote a letter to the board insisting that they do a buyback. The company was bankrupt 18 months later.

The lesson is to be very, very careful of high leverage because the company was just too levered. Sometimes not everything goes right. What happened was that a price war broke out between 2 unrelated competitors who were much larger. They got caught up in that. I'll forever have the black mark of having written my only letter to a board asking them to buy back stock 15 months before the company went bankrupt.

Jas

I appreciate that. Thanks for the candor and the vulnerability. I appreciate that.

We'll mix it up a little bit. Usually lightning rounds come at the end.

Sure.

Jas

Let's put it at the beginning. Overrated or underrated? Pattern recognition. Scuttlebutt. Position sizing. Sleep. You just get 1 word.

I think pattern recognition and sleep are highly underrated. I think scuttlebutt is overrated.

What was the fourth?

Jasjit Singh

Position sizing.

Position sizing is very important. You just have to pick your game and you have to play it. You can either be a slugging-percentage player or you can be a batting-average player. You have to know what your game is, stick to it, and be consistent.

Jas

Appropriately rated, sounds like.

The production function of Gavin Baker: when you actually look at where your edge came from, maybe this year or last year, what fraction is reading? What fraction is your network? Maybe not so much, given the scuttlebutt. What fraction is just being early to 1 or 2 correct frameworks?

I think reading is overwhelmingly the most important part of it. I will admit I rarely meet with public companies. I only meet with them if they want to meet with me, and they're very well trained. They never say anything that's not in a transcript or 10-Q. I can read much faster than they can speak, so I read a vast amount of transcripts and primary-source material.

I do think these expert transcripts are very good and a great use of AI. Reading is overwhelmingly the most important part, and then pattern recognition is also important. Being early to frameworks is helpful.

2. Memory Cycles Shape the Call

Jas

You were so early to it, and that's a good segue because of that idea of a framework and your mental model. Today, memory prices are up 60%. Micron's margins are probably in the high 60s, and the historical average was closer to 16%. You've been talking about the compute shortage broadly, the shortage stacking into data centers and power, and extending into leading-edge wafer capacity.

What is your mental model for how this evolves? You've also said that whenever there's a shortage, there's eventually a glut. How does this evolve? Talk us through that.

That has been true throughout history, and I'm sure there will eventually be a glut. But “eventually” is doing all the work in that sentence, not “glut.”

I would say, based on every memory cycle we've had for the last 25 years, this is the time to be selling memory—100%. I was actually the Micron analyst in 2000. I remember going to their analyst day in Sun Valley. I'm a veteran of many, many memory cycles, and based on history, this is the time to sell.

However, there's 1 cycle where you absolutely do not want to sell, and that's the cycle we had in the mid-90s, which is the last true capacity cycle that I would argue we've had in memory. Based on that cycle, we may still be very early.

I listened to my friends Alex and Leon, who I thought did a great job, as did Leslie earlier. I would just say I take the over on every number that they gave—every single number. I think they're conservative guys. I bet they would take the over, too. Nobody wants to be wrong a year later here at Sohn. So we may still be early.

This may be the first true capacity cycle, and I do think that these fundamental shortages are good for us as investors. The last thing anyone should want is a bubble. Bubbles are terrible. They're awful to invest through, and the aftermath of them is even worse. We don't want a bubble.

Unfortunately, the entire history of financial markets suggests that whenever you have a profound new technology—whether it's AI, the internet, the PC, railroads, or canals—you almost always get a bubble because markets are efficient. Investors understandably become excited about this new technology. Michael Mauboussin frames it as a breakdown in diversity. Everyone comes to believe in this, you get a bubble, and then that bubble funds the build-out that the new technology required. That's exactly what happened with the internet.

I am optimistic that we may avoid a bubble this time. Smoother for longer is what we all want. The reason we're going to avoid it is that we have fundamental shortages of watts and wafers. We're going to address the watt shortages with orbital compute, for sure, in the next 5 to 7 years. But I think the wafer shortage is going to persist for a long time.

The reason it's going to persist for a long time is that Taiwan Semi is run by flinty old men and women in their 70s. Not to say that 70 is old—it's the new 50.

Jasjit Singh

I'm 50; it's the new 30. [laughter]

But they're the most important people in Taiwan. The president of Taiwan is irrelevant. They are Taiwan. And they view themselves as the guardians of Morris Chang's legacy.

I remember going to Science Park in Taiwan more than 20 years ago, asking, “Do they think they could ever catch Intel?” And they said, “It's a beautiful dream, but it's probably for our grandchildren.” And they did it in one lifetime.

So they're the custodians of this legacy. They need to preserve Taiwan. Taiwan's bubble and bust is a disaster for Taiwan Semi in Taiwan. And so they're simply not expanding capacity as fast as Jensen wants.

Jensen goes there every 3 months, and maybe they expand 5%. He wants them to double or triple. And if they double or triple the capacity, Nvidia could probably sell $1 trillion, $1.5 trillion, or $2 trillion worth of chips next year. I really believe that.

But the other side of that might be very painful for everyone. And so I think these flinty old men and women who are safeguarding Morris Chang's legacy are helping us all avoid a bubble by enforcing a real-world physical constraint that simply has not been present with previous technologies.

Jas

Yeah, no, it's fascinating. As a monopolist provider, they're rate-limiting supply at some fundamental level.

They dismissed Sam Altman as a podcast bro after they met with him. [laughter]

3. Code Becomes The Killer App

Jas

You and I were talking backstage about this. You would take the over on OpenAI and Anthropic combined at $200 billion of revenue, maybe in 12 to 18 months. I don't want to put a time frame on it, but near term.

It turns out that code generation was the killer app to monetize AI, at least in chapter 2 of this whole journey and beyond. And if we're going to go from where we are today to get to that $200 billion in the next 12 to 18 months, where is that revenue coming from?

Is every S&P 500 company going to miss earnings because of tokens to Anthropic? I don't think that is an edge case. If you are not aggressively token-maxing, Jensen said at GTC that his goal was for his best engineers to spend a minimum of half of their compensation on tokens.

And if you just look at the wage expenses of every S&P 500 company, we simply cannot tolerate that level of token spend without significant adjustments to the labor force, which is the point Leon was making.

But I do think a few things are probably going to lead to us not having widespread misses because of token-maxing. And if you're not token-maxing, you should be. If you don't know what token-maxing is, best of luck.

Point number 1: All of these models are shifting to usage-based pricing. It used to be that everyone in this room could get a good sense of the capabilities of frontier AI if you spent $250 a month on the best subscription from a frontier model provider. That is no longer true.

The best capabilities are locked behind harnesses and reserved for people on enterprise plans who can pay for use. Now, this is wildly bullish for AI. It's incredibly bullish for the pricing of these frontier tokens.

Going back to the cellular industry, the reason cellular was a great industry as a growth investor, and the reason long distance before that was a great industry as a growth investor, is that you bought a fixed amount of minutes and then, if you went over, you paid by the minute. And people really like to talk to their friends and family. That was why telecom was a great growth industry for a long time.

We're just moving from these all-you-can-eat plans to usage-based plans with overage, where those usage tokens cost a lot more. And we're finding out that we're nowhere near people's ceiling price for how much they'll spend. So I think you'll get a lot of productivity.

I think that this fundamental compute shortage we have—I thought what Alex said was so profound. Ten basis points of the world's population is using these models the way they should probably be used, and we're in an insane shortage despite spending cumulatively trillions of dollars. What happens when 5% of the world's population is using these models the way the cutting-edge 10 basis points are? It's unimaginable.

This is why Orbital Compute is a necessity. But to answer your specific question, I do think someone posted on X that coding may be the shortest path to ASI and AGI. Because if you can write code to do something and write it for yourself, that's a pretty fast, elegant path to AGI.

And I think it may end up being that coding is not just the killer app for AI, but it's the ultimate AI app, and it subsumes more and more. I would just encourage you: Claude Code and Code X will get you better answers for investment questions, even if you're not a coder, than you will using the regular model.

4. Nvidia Faces New Rivals

Jas

Yeah, no, I appreciate that. You talked about silicon and what's happening on the chip side. Obviously, you were very early in Nvidia. I mean, you were talking about talking to Jensen in 2000, let alone 2023, when the ChatGPT moment happened.

Alternatives are coming. Competition is coming. He's still going to be the dominant part of the market, but competition is coming. As you think about Trainium, TPUs, and MTIA, which of these is most underestimated by the market? Where does the consensus get it wrong?

Trainium by far. Trainium is going to be to 2026, especially in the second half of this year when Trainium 3 really ramps, what TPUs were to 2025.

And if somebody is wildly bullish on TPUs today, let's go look at their 13F and see if they owned Momentum or Elastica, which were the best ways to invest in TPUs. I owned one of those. So I do feel like I have some credibility to say this.

But I do think Google, for a lot of reasons, made very conservative design choices with the TPU v8. Nvidia and Trainium made very aggressive design choices. So Trainium is for sure the most underestimated, not only because of those design choices, but because all of these frontier models are what are called mixture-of-experts models.

To inference one of these, not to get too technical, you need something called a switched scale-up network. And the only 2 functioning switched scale-up networks in the world today are the ones that power Nvidia's GPUs and Amazon's Trainiums.

Jas

Interesting.

And this is why Google invented the MLPerf benchmark. They will not submit TPUs to their own benchmark, which you can just see is visibly driving Jensen crazy.

But listen, the TPU is a great chip. I'm sure the TPU v9 is going to be amazing. They'll make more aggressive choices. I would never bet against Google. I would never bet against Broadcom. But I do think Trainium is super underestimated right now.

5. Neo Clouds Find Their Edge

Jas

Yeah, I appreciate it. I want to switch gears to a topic that is actually how we first got connected, way back in 2022 and then 2023: neo-clouds.

I called you in the summer, when you were on vacation somewhere, to bug you about a company called CoreWeave in the summer of 2023 and get your advice and input. Ultimately, that led to us investing $7.5 billion behind CoreWeave to scale them up at a really pivotal moment. So first of all, thank you for that.

Fast-forward to today: CoreWeave, Crusoe, Nebius, Lambda, and others. Is the category durable today? Is it a transitional arbitrage on hyperscaler CapEx and token friction?

I absolutely think it's durable. And so, first of all, CoreWeave is a little bit of a sad subject for me. The Trade Desk could have invested over $50 million in the round at $1.1 billion, and I was conflicted out because of Crusoe.

I love Crusoe, and I think Crusoe's going to work, and we have a large position in Crusoe. But every time I think of being able to put $50 million into CoreWeave at a billion dollars, I get a little sad. I'm very happy you put $7.5 billion in.

It's absolutely a durable category. And the way to think of running one of these clusters is like driving a Formula 1 race car. You watch Formula 1 races, and it looks easy. It's like anybody could do it.

In the same way, if you watch Tom Brady, it's like, “Well, why did he miss that throw?” Well, because he's in a stadium with 100,000 people yelling. There are a bunch of people who weigh 100 pounds more than him, running 20 miles per hour at him, trying to kill him before he throws the ball.

In Formula 1, it's the same thing. It looks easy, but it's really hard. And that is to say that if I tried to get in a Formula 1 car and drive it in a race, I would die. I would be just a danger to myself and a danger to everyone, including the people in the stands.

And that is running a cluster. It's really hard to do well. You understand this, but the reason a company like CoreWeave can charge a huge premium for its GPU hours is because all GPU hours are not the same.

Those CoreWeave GPUs are being utilized 2 to 3 times more per hour, on average, than the GPUs from maybe a bottom-of-the-barrel provider. And by the way, this all goes for Crusoe, Nebius, and other high-quality neo-clouds.

But I think this is very underappreciated. People think, “Oh, that can't be a durable competitive advantage.” Well, I was a retail analyst in 2005, and I observed that all you had to do to create a $50 billion market cap in America, in any category, was be able to run 1,000 stores in 50 states with different climates and different consumer preferences, staffed by friendly, knowledgeable employees who don't steal from you, stocked with the right inventory at the right time and at the right prices, and have clean, well-lit stores.

That's it. In history, like 10 companies have been able to do that. Running one of these clusters is even harder. So I think it is durable, and the hyperscalers for a long time were stuck in a cost mentality.

The hyperscalers were competing with people running these Formula 1 cars, and they were doing overnight shifts in 18-wheelers, trying to stay awake to deliver the lowest cost. And that's not what AI is about. I think they've made this mental and cultural shift. But I think some of these neoclouds have a very durable business model.

6. Orbital Compute Takes Off

Jas

Yeah, no, we certainly agree. We've talked a little bit—you've actually alluded to it already tonight—which is orbital compute. We're not going to get into the science of it because there are papers you can read online and send through your favorite LLM, Grok, et cetera, to distill. When does it become commercial in a way that it gains meaningful market share?

Orbital compute: when does that commercially gain real share, and what is your view on the most underappreciated short in the market today that is not being priced in? Is it terrestrial data center operators? If that's the case, we're in a lot of trouble with Blackstone. But are there other things that today are clear shorts for you based on that?

I think it becomes clear that it is possible, going to work, and economical in the next 2 years. I think it starts to take meaningful share toward the end of this decade. Mhm. I do think there may come a day when we never build another terrestrial data center. Terrestrial data centers that have been built and are in the ground are always going to be valuable.

You're going to do training and RL in terrestrial data centers. But I can't imagine a day in the next 7 years when we never build another terrestrial data center. The years leading up to that are going to be very painful for a lot of the companies in the power and cooling spaces. These industrial names have massively flexed up capacity to support a build-out that could really come to a screeching halt.

In space, the power comes from the sun, and the cooling comes from the dark side of the satellite. There are certain names that I'm not supposed to say, but if you've seen the illustrations of a prominent potential provider of orbital compute and what their satellites are going to look like, the radiator is 300 or 400 feet long, and it sits behind the satellite in a sun-synchronous orbit.

So, you have these big solar wings and a satellite, and the satellite is a rack. It's not a data center. It's just a rack: 8 feet tall, 2½ feet wide, 4 feet deep. They're stitched together using lasers to make a virtual data center. Then you just have the radiator in the shadow behind the rack.

Jas

Gavin, we're out of time, but thank you. This has been fantastic. I appreciate you doing this.

Thank you.

Jas

Thank you all.