#1 Semiconductor Expert Gavin Baker: "China's Not Getting Taiwan's Fabs." | Invested

Thank you. Thank you. So excited to be here.

Oh, yeah. It's awesome.

Absolutely.

It was more than 10 years ago.

You still think that's true? Well, so I think the exact line was, well, first we should give credit to Bill's partner Eric Vishria, who I ruthlessly stole that line from, but I did modify it and say foundation models without access to unique data and internet scale distribution are the fastest appreciating assets in history. I would say my opinion has changed a little bit and I do think xAI, Anthropic, OpenAI, DeepMind, they've opened up a lead and to varying degrees they all have, with probably Google and xAI having the most unique data. I think Anthropic very smartly focused on code early and then OpenAI got to internet scale distribution just because they were there first. But I do think fundamentally reasoning models have made the business of foundation models dramatically better. And the reason is that for every trillion-dollar-plus internet company, every great internet company...

100%. Yeah, for every dominant internet company, there's the same flywheel that underpins it. And that flywheel is that as the product gets better, you get more users, the users create more data, the incremental data makes the product better, which attracts more users. And that flywheel spins at Google every day in the search business, Netflix every day in the recommendation, and Amazon every day in their recommendation engines, at Facebook every day both in their ad targeting algorithm and the algorithm they use to select what content to show you. That loop was not present at all in foundation model companies, LLMs, until very recently. And so the reason that LLMs with unique data had an advantage was one, it would presumably lead to over time an advantage in training, if you could train on data that no one else could. And then secondarily, if you had access to something like X or everything Google knows about the world, you can use that to ground the LLM's answers more firmly in reality. But reasoning fundamentally changed this with the introduction of kind of reinforcement learning with verifiable rewards during post-training and test-time compute. All of a sudden having a giant user base becomes a big advantage and feeds back into the quality of the product in a way that was simply not present until OpenAI came out with o1 in late '24. DeepSeek quickly followed with R1 and then xAI, Google DeepMind, Anthropic quickly followed with their own reasoning models. So I think that these scaled foundation model companies are much better investments post-reasoning because this flywheel where the more users you have, the more data you have, the better the product is, now exists in AI where it simply did not exist before, at least for LLMs.

Microsoft, Google, am I missing anyone?

Broadcom now.

Like it's all of a sudden become a club, the trillion-dollar club. And you look at these Anthropic, OpenAI, and you go, 'Oh, those are obviously going to be trillion-dollar companies.' I sat today with somebody, by the way, who told two people who told me they put up over $100 million in the recent Anthropic round. And I said, 'Boy, betting on the trillion-dollar outcome.' I stopped myself for a second and said, 'Betting on the trillion-dollar outcome? I've been doing this 30 years. There haven't been that many of those.' Yeah. Well, I will just say I do think it's been fascinating to see Grok Code release and take overwhelmingly dominant share of the coding tokens on OpenRouter. Grok Code is a great product, but yeah, these are big numbers. But it's just funny, you know, Anthropic, what was that, $180 billion?

I mean, they went from $1 billion to $6 billion faster than any company in history in essentially six months. So all of a sudden you're 30 times current month's revenue. I think there are real corporate structure concerns there. Like in 2018 when the governing documents were written and Microsoft made its investment, people thought AGI was a fantasy and as soon as AGI is declared, all of kind of the future technology and future technology from OpenAI goes to the nonprofit. But nonetheless, I mean, OpenAI has, you know, 10-plus billion dollars revenue. xAI, it's not public, but they're scaling rapidly. So these companies now all of a sudden you can value them on revenue multiples. Now that you have that core product loop driven by reasoning in place, I think it's pretty clear that it's going to be an oligopoly that's going to be tough to break into. But I think maybe the heart of your question is to what degree is this a sustaining or disruptive innovation for the existing trillion-dollar-plus companies? And I think the answer has to be, I mean not to be clever or avoid the question, is it's both. So in I think maybe in a prior interaction I said, you know, pre-AI each one of these mega-cap tech companies was kind of in its own swim lane. Like Netflix, they competed a little bit with Amazon and Google and Apple, but you know, Netflix was pretty dominant. Google dominant in search, Meta dominant in social, Microsoft dominant in productivity. And then you have this cloud computing oligopoly of Google, Microsoft, and Amazon. And you know, nobody's really competing with Amazon in e-commerce. Post-AI, post-LLMs, post-transformers, at some level, they're now all in the same race because I think it's very clear that the transformer is going to be able to do essentially everything that lies at the core of each of those companies. You're going to be able to make personalized movies, personalized songs, you're going to have an agent that can do all of your shopping. You know, the fulfillment infrastructure will still be valuable. They can obviously do search. And then I think it's an open question what is the value of human-created content in a world where AI can create perfect content that looks like it's from humans. And so, you know, they're kind of all in the same game, but at the same time the core ingredients to success are compute and data, and these companies have the most compute and the most data. So while it's very disruptive and very scary, it's also sustaining in that it plays to some of their advantages if you can attract the right talent. Google was able to do that. Amazon not really in the game, but they bet on Anthropic, which for sure has the right talent. xAI has been able to do that, but Apple has not even tried. And Meta has until now essentially failed. We'll see if, you know, V3 or V4 of the strategy works.

Free agency market. You know, it's just very funny. Like I worked for a long time at a large financial institution and I think most big companies they eventually learn that it's pretty hard if you have a culture of growing your own talent to bring in expensive external talent because it really disenfranchises and upsets the loyal internal employees. And you have for sure seen that at Meta. They have a lot of advantages, but it's also kind of the fight of their lives. And that's why I do think they're all in a prisoner's dilemma where, you know, the amount of capex is going to an unimaginable place. Like I think, you know, I think you could see some of these companies, you know, first issue debt, then cancel their dividends, then stop the buybacks because they all feel that winning in this, we'll say winning in LLM, is existential and they feel like the prize on the other side is artificial super intelligence which they all have a near-religious belief is going to create tens of trillions of dollars of economic value.

Maybe somewhere in between. I mean, you know, the current trajectory, the current trajectory I think is supportive of getting to artificial super intelligence. But I think one of the most interesting parts of all of this is we don't know what the economic returns to super intelligence will be definitionally because we have never seen it. And if humans, we have pushed the limits with, you know, people like Albert Einstein or whoever else pushed the limits of the natural laws of the universe, if we've pushed the limits of physics and biology and chemistry and the math that underpins all of it, then the economic returns to super intelligence, they may not be that high. On the other hand, if we haven't and super intelligences are curing cancer, inventing warp drives, and we're colonizing multiple solar systems, the returns to super intelligence are going to be really high, but it's very unknowable. So I think there's a, I would say there's a quasi-religious belief that we're going to get to super intelligence that I think is well-founded in data, but the economic returns to that super intelligence is to me definitionally unknowable. And that is where I think there is, you know, for sure religious belief.

He tried to weaponize capital. When you weaponize capital you end up with this fight to the death like you had between Uber and Lyft for a while until Uber won it. But they really won it in the private markets. You know, they raised untold amounts of capital in the private markets to do it. Now, we're in this incredible race where you've got multiple public companies chasing the same bacon. Kosher bacon, of course, but chasing the same bacon and a couple of private companies doing the same thing and they're all fighting over what's I guess notionally the same prize. Yes, Anthropic kind of specializes in code and OpenAI is the consumer, but foundationally they're after the same prize. How does that end? And to your point, they all have this religious belief they're going to raise hundreds of billions of dollars maybe each, right? If you believe OpenAI's deck, which I'm not sure I do, to get to this. And has that ended a return? I mean, well, so to date, there has actually been a storied venture capital firm whose name we will not mention on this podcast, a storied American venture capital firm. One of their partners wrote an essay called the $200 billion question and then the $600 billion question. And the point of the essay was the first $200 billion question is where's the economic return on this $200 billion in GPU spend going to be and then where's the return on the $600 billion? And the funny thing to me about that, you know, sometimes I think, you know, VCs, they're very in their own world of venture. The funny thing to me about that was all the biggest spenders on GPUs with the exception of xAI are public companies.

And public companies report something called quarterly financials, and they file a document called the 10-Q, which has to be accurate. And you can use those documents to calculate something called return on invested capital. It's really easy to do and to date the ROI on all of this spending has been strongly positive.

And some of that you could argue was that they got religion about expenses at around the same time. But in a world of AI fundamentally the economic return a lot of it will come from either replacing or augmenting humans with GPUs.

We will see. But the maximalist I think for sure would say labor replacement. So, I think that's a fair way to calculate it. And then I would just say it's become really clear over the last six months in particular, you started to see big accelerations in the top line. And a lot of that has been in using LLM and AI and GPUs to improve the recommender AI algorithms that power a lot of these companies. But nonetheless, I mean the returns have been good. But just the question is the quantum of spend with Blackwell, which is the next generation Nvidia GPU, is so big. You know, these $50 to $100 billion data centers, just will the ROI continue to be good? So while the answer to the $200 billion question was yes, the ROI was awesome. The answer to the $600 billion question was yes, the ROI has been awesome. Now we're at the trillion-dollar question and the ROI has actually still been good, but we're rapidly heading for two to three trillion. And will the ROI continue to be good? Unknown. And if it's not, do economics dictate or does the prisoner's dilemma dictate?

That is a great question. I wanted to think, you know, you are, I would argue, one of the world's premier experts on semiconductors and the semiconductor supply chain and understand this incredibly deeply. So the dominant architecture today is the GPU. Does that continue?

Absolutely. Comes from Annapurna.

Yeah, so that's an incredible team but I think everybody else is going to struggle to out-GPU the GPU. And but then if you can pick a niche, you know, because at some level all designs are about trade-offs. Like in tank warfare, they talk about the iron triangle. You have to make a trade-off between offense, defense, and mobility. And because everybody is dealing with the same raw materials, the stronger you make the defense, the less mobile it is because the more the tank weighs. And so you have to make tradeoffs. And it's the same thing in semiconductors. And so as these niches become so big, I think there is an opportunity to make trade-offs that are different from what the GPU is doing to do something different. And then you can maybe start with a niche that's 0.5% but that's, you know, billions in revenue and tens of billions in market value and then you can go from there. And so recently you've started to see, my firm I should say is an investor in Cerebras and Cerebras and Grok and I think a couple of others have an SRAM architecture and that can do something fundamentally different from a GPU.

On a big chip. Cerebras in particular, Cerebras did something very radical. Instead of cutting up a 300-millimeter wafer into lots of little die and making that the basis for a chip, they used the whole wafer. It's a really radical approach and it was really hard for them. It took them their third generation of their chip before they saw success. But now they and Grok, Grok 2 was another SRAM company, maybe less radical approach but I think they're both seeing real commercial traction because they can do something different and that they can generate extremely high throughput on tokens. So tokens per second really matters. Many of you use a reasoning model, you know, when it's thinking it's literally generating tokens and depending on who you use you can kind of see its chain of thought and you know an answer that might take 30 seconds takes 3 seconds on Cerebras. So that's a niche. Next Silicon, mutual investment of ours.

Mutual investment of ours. They're also doing something fundamentally different. I don't know how much is public about Next Silicon, but they're doing something different. And I would, you know, and I saw a lot and I pick a niche, you know, pick a niche. And you know, and I think it's a truism that in a lot of semiconductors it takes three generations to really, really be successful. Three generations just to be clear to everyone. Three generations of tape-outs, you need the third generation chip and you finally make it. People don't remember that Nvidia is a 30-year overnight success.

Nobody realizes it's a 30-year overnight success. Trillion dollars or $4.3 trillion did not come to Jensen overnight. It was 30 years of a life's mission. Absolutely. And just in case, each chip is kind of like a two-to-three-year cycle and it's really hard. Tape the chip out. You send off this big digital file to Taiwan Semi in Taiwan. You anxiously wait for three or four months while they make the chip.

No, I do think it's very interesting. Tesla has made this giant partnership with Samsung in Texas. And one of the conditions of that partnership was that Elon would be able to have influence over the design of the fab and the way it's laid out. And in the same way he kind of reconceptualized electric vehicles and then rockets and then data centers from scratch. I think it's very underappreciated that the consensus belief of essentially every AI architect, data center architect was with the Hopper generation you could not get more than 30,000 coherent in a training cluster. He looked at a data center from first principles, did something very different and got 200,000 coherent. He's going to look at semiconductor manufacturing from first principles so I actually think that is something interesting may come of that but no it's kind of wild to make one little semiconductor, you know, which in the case of a GPU is not that small. GPU is what's called a full reticle chip, and now it's two of them. Think of it as being about half the size of an iPhone. That takes four months to make it. And then you get the chip back and it's, you know, an incredibly anxious moment. You literally plug it in and you hope it works. It's called power-on. And sometimes it doesn't. And then you have to make some tweaks or if you really screwed up, you have to start again. So, I don't know how we got on that though.

Yeah.

So I think it is a little overstated. So Taiwan Semi, so Taiwan Semi is 18 months ahead of kind of Intel and Samsung who are the other competitors, maybe two years. Now two years is a lot, that's Moore's law. So you are getting a chip, you know, however you want to call it, you know, 50% more performance, 30% less power. It matters. But if the world had to roll back to two years ago, that would be a big shock to the world, but the world would not stop spinning. There's a lot of capacity that's essentially two years behind Taiwan Semi.

Yep.

In larger, in larger sizes.

There wouldn't be... what generation of iPhone just came out?

Yeah, I think, okay, so if that happened there'd be the iPhone 17 production would stop.

And you'd have to take, I don't know, I mean semiconductor guys get mad at me but you know it'd be six months of really hard work to port the design of the chips that are in that to the old process.

But a generation back.

Have you been there?

If I was Taiwan Semi I'd have you come.

For 50 years, Intel was ahead. And I've been to Science Park. And Morris, the legendary chairman and CEO, he described catching Intel as a dream, a beautiful dream. And he wasn't sure it would happen in his lifetime, but it was a beautiful dream. And the reason it's very hard to catch the leader is semiconductor manufacturing is a little bit like baking where you can give two people the same recipe and it doesn't taste quite the same and everybody's using the same equipment. But if you are ahead and the leading-edge customers are using your recipe, it gives you a start on the next one and how you kind of combine all the ingredients, you know, the flour, the sugar, the salt, and in this case, you know, maybe it's the scanners and the metrology machines and the etching, deposition. Yeah. But you're fundamentally using the same machines. You have a recipe and because you developed the last recipe and figured out what worked 18 months before anyone else that helps you figure out the next recipe. The reason Taiwan Semi caught Intel is Intel made two crucial errors. One, Intel for a while was run essentially by three CFOs. The chairman, the CFO, and the CEO. A well-known semiconductor analyst called them the three idiots. And they made a really big mistake because this is a game of trial and error and it's almost as much art and science and you've been baking these recipes for a long time. They did not understand that the most important people in the manufacturing division were what were called the gray beards. The people in their 50s and 60s who had been, you know, sticking their finger in cakes for a long time, licking it and being like, 'We need more sugar.' Okay? And they're like, 'Oh, this is so smart. We got these really these 30-year-old guys, and we can nuke these 55-year-old guys and save some money, and these guys are energetic.' And that was disaster number one. Okay, Taiwan Semi is never going to make that mistake. Nobody's going to make that mistake again. The second mistake they made, I think most bad decisions in business and life come down to arrogance. And they believed that they were so good that they could move to what's called EUV, which is a kind of lithography, one node after everyone else. And this was despite being the largest shareholder of ASML, but they were outraged at the prices that ASML was charging. Thought we're so good that Taiwan Semi, they have to buy those, but we can do it without them. Double, triple patterning with something called DUV.

It's hilarious. Absolutely. Yes, that's right. Marques, Will Danoff, the great Fidelity portfolio manager and a good friend and mentor of mine who runs I think close to $300 billion with incredible numbers. His parents were in the garment district of I guess New York and would say to him, 'Price is forgotten but quality remains.' Now price is important. Price does matter in our business. But nonetheless sometimes prices are worth paying. So that was a terrible mistake. And then the last thing Taiwan Semi had going for them was the entire industry of semiconductor capital equipment. So ASML, KLA-Tencor, Applied Materials, Lam Research. They had their A teams working at Taiwan Semi because they were unhappy that Intel effectively had this monopoly on the leading edge. So now what's happened? Taiwan Semi has made the bet we're so good that we can do High-NA EUV a node after Intel. Intel went to that node first. Second, the A teams from these, the KLAs and the ASMLs have been at Intel because now Taiwan Semi has a monopoly. And I think you had, you have a team there led by Lip-Bu Tan who's a legendary semiconductor investor, great executive, you know, who understands what needs to be done and I actually think Pat Gelsinger has been very unfairly treated in the press, you know, he embarked on an audacious strategy that was kind of their only chance and now you have the US government who I'm sure in many, many ways is going to politely say, 'Hey, we know that you know, you like the chocolate cakes made in Taiwan more than the chocolate cakes made in Oregon and Arizona by Intel. We'd really like you to buy that American chocolate cake and you know, just find something.'

To bring business to Intel, just to say...

Even if it costs those commercial companies more money and they have to kind of stay a half a generation or a generation behind. Yeah, I think the cutting-edge products will probably still be made in Taiwan for the foreseeable future. I mean, we'll see how Intel 18A and 14A nodes pan out. But there's, you know, Nvidia now has, yeah, I think seven chips that are very important and maybe some of them you can make at Intel. But the other fascinating thing is if Intel can just get its act together, Taiwan Semi, I think they're very surprised. You know, they open up the Arizona fab, sold out in milliseconds.

There's unbelievable demand.

Really?

Yes.

No, I just think the geopolitical fears are real.

Yeah. I just, this administration I think a lot of their policy is about, I think during COVID what the US military realized was, wow, we are too dependent on China to fight a war with them right now.

An extended war.

And so I think, you know, they do believe in peace through strength and to credibly deter China and to keep Taiwan, which is a democracy, you know, a small democracy that, you know, I actually think in a lot of ways it's kind of parallels Israel, you know, maybe four times the population but you know, it's a small democracy and you know, Israel is now essentially geographically surrounded by friends but that was not always the case. And, you know, Taiwan, you know, depending on who's in power in Taiwan, China is more of a friend or maybe a little more adversarial. But I think they believe that if we can effectively decouple from China, then that can bring durable peace and avoid having a giant war in the South Pacific where essentially China faces off. I think people forget it would not just be America. It'd be Japan, it'd be South Korea, it'd be Australia.

Not many people know the geography of that part of the world. I've discovered when you find yourself between New York and Los Angeles.

Yeah. And so I mean just to state the obvious. So my iPhone would cost the American consumer 20% more.

Yeah. But a lot of those are rapidly being relocated to Mexico.

Yeah, absolutely. And you know Nvidia and Apple and Amazon and the hyperscalers are driving a lot of this. Actually, big Chinese companies are opening up American subsidiaries that can be legally separated and they're bringing their best engineers to stand up manufacturing plants in America or Mexico.

100%.