The Promise and Peril of Computer Chips with Tufts Professor Chris Miller

What this episode is about

The discussion focuses on the pivotal role of semiconductors in shaping modern business and everyday life, emphasizing both the opportunities and challenges faced by the U.S. in maintaining its competitive edge in the global chip market. Chris Miller highlights the complexities of international supply chains and the pressures from emerging markets, especially China, while asserting that the U.S. still holds a significant position if it strategically invests in innovation and talent.

Key takeaways

• Tech's best-kept secret? Designing chips isn't just for engineers anymore; it's for strategic thinkers too.
• China's market allure is like a siren; every foreign firm risks getting lured into a complex trap.
• Owning a semiconductor fab? It's like having a shark pool; exciting but expensive, and you might get bitten.
• Cars now house over 1,000 chips—proof that we're driving a tech revolution, not just a manufacturing one.
• Tesla blurs lines: Is it a car or a tech company? Welcome to the age of automotive AI.

Who speaks in this episode?

• Kara Miller (Host) · 17 min speaking
• Chris Miller (Guest) · 26 min speaking

What is discussed, and when

1. 0:00 – 5:00: Artificial Intelligence, Chip War, MIT, AMD
2. 10:00 – 15:00: Apple Inc., Alphabet Inc., Chelungfu Fault, San Andreas Fault
3. 15:00 – 20:00: Artificial Intelligence
4. 25:00 – 30:00: Artificial Intelligence, GPUs, CPUs
5. 40:00 – 43:35: Artificial Intelligence, Chip War

Notable quotes from "The Promise and Peril of Computer Chips with Tufts Professor Chris Miller"

• But first, CSAIL Alliances is the industry gateway to MIT's Computer Science and Artificial Intelligence labs.

  about MIT · at 1:33

• If Facebook's fancy headquarters were to sink into the San Andreas Fault, the company might barely notice if TSMC's fabs, which make chips were to slip into the Chelungfu fault, whose movement caused Taiwan's last big earthquake in 1999,.

  about San Andreas Fault · at 10:09

• And today, almost all of America's largest tech companies-- NVIDIA Apple Microsoft, Meta Alphabet they couldn't function without chips from TSMC.

  about Apple Inc. · at 11:54

• Intel was the world's dominant producer of microprocessors that went in PCs and in data centers, the CPU chips, as they're called.

  about CPUs · at 23:59

• And the types of workloads that you need on AI, especially LLMs, are much, much faster and more efficient on GPU chips, not CPU chips.

  about GPUs · at 25:54

• I think the other thing that I'm struck by when you look at the development of AI is that the biggest block by far in the deployment of AI is going to be not computers or technologies or models.

  about Artificial Intelligence · at 41:42

• [KARA] Chris Miller is author of Chip War, The Fight for the World's Most Critical Technology.

  about Chip War · at 42:28

Who and what is mentioned in this episode?

Books & papers

• Chip War — Book · Primary recommendation · 2 mentions · first at 0:13
  Book discussed in the episode, providing insights into the semiconductor industry.

Companies & orgs

• MIT — Company · Host · 1 mention · first at 1:46
  Host of the podcast and a key institution in technology research.
• AMD — Company · Business focus · 1 mention · first at 2:05
  Competitor in the semiconductor industry, known for its CPUs and GPUs.
  In 1969, Jerry Sanders co-founded a company called AMD, Advanced Micro Devices.
• Apple Inc. — Company · Business focus · 1 mention · first at 13:26
  Major tech company relying on advanced semiconductors for its products.
• Alphabet Inc. — Company · Business focus · 1 mention · first at 11:58
  Parent company of Google, involved in AI and technology.
  And today, almost all of America's largest tech companies-- NVIDIA Apple Microsoft, Meta Alphabet they couldn't function without chips from TSMC.

Concepts & topics

• Artificial Intelligence — Concept · Core topic · 5 mentions · first at 0:00
  Central theme of the episode, discussing the impact and future of AI technology.
• GPUs — Concept · Key argument · 1 mention · first at 26:03
  Critical for AI and data center operations, especially those produced by NVIDIA.
• CPUs — Concept · Key argument · 1 mention · first at 25:42
  Discussed as a fundamental component in computing and data centers.

Places & events

• San Andreas Fault — Place · Setting · 1 mention · first at 10:11
  Mentioned as a hypothetical scenario to illustrate the importance of TSMC.
• Chelungfu Fault — Place · Setting · 1 mention · first at 10:19
  Mentioned as a hypothetical scenario to illustrate the importance of TSMC.
  To sink into the San Andreas Fault, the company might barely notice if TSMC's fabs, which make chips were to slip into the Chelungfu fault, whose movement caused Taiwan's last…

Full transcript with timestamps

0:00 [KARA] Welcome to MIT's Computer Science and Artificial Intelligence Labs Alliances podcast. I'm Kara Miller. On today's show, Chris Miller, the author of the best-selling book, Chip War, talks about how semiconductors have come to shape our lives, and why they're in the process of remaking business as we know it. [CHRIS] One of the problems of being successful is that you focus too much on defending your legacy industry, rather than investing in what might come next. [KARA] Miller is a professor of international history at Tufts University, and he notes that his tech becomes more complicated and supply chains become more complex.

0:46 Companies are under a lot of pressure. [CHRIS] Every foreign firm looks at the Chinese market and simultaneously says, it's very large. We've got to sell there. But also in the long run, the Chinese state wants to squeeze us out and replace us with a domestic firm. That's a very hard dynamic for almost every company to manage. [KARA] So, can the US maintain its competitive advantage when it comes to the chips that are in our coffee makers, our cars, and our phones? [CHRIS] The US is still in a pretty strong position, which doesn't mean that if the US squanders it by chasing away the world's smartest people, we're going to stay in the strongest position.

1:22 But it does mean if we play our cards wisely, I think US researchers and US firms are likely to retain their lead in many of the key segments. [KARA] That conversation is coming right up. But first, CSAIL Alliances is the industry gateway to MIT's Computer Science and Artificial Intelligence labs. Members get access to CSAIL research, talent, conferences, and courses. If you want to learn more, visit csail.mit.edu/podcast. That's CSAIL C-S-A-I-L.mit.edu/podcast.

2:05 In 1969, Jerry Sanders co-founded a company called AMD, Advanced Micro Devices. And Sanders had a very specific style. He wore a Rolex, he sported Gucci loafers, he had a beach house in Malibu and a Rolls-Royce. In fact, by his own account, he was maybe a little too flamboyant. [CHRIS] Jerry Sanders was representative of the old era, this kind of macho style in which you had to have your big, expensive, complicated manufacturing plant in order to be a real chip company. [KARA] Chris Miller says that for an older generation of chip makers, you couldn't really separate designing and manufacturing chips.

2:47 The process was pretty integrated. Jerry Sanders loved to tell people that real men have fabs, which is to say, the facilities that fabricate chips. But then pretty quickly, technology pivoted. [CHRIS] There was a new set of software tools that emerged in the 1970s and 1980s, came largely out of MIT, actually. And it enabled the abstraction of the chip design process away from manufacturing. And by the '80s and '90s, you could begin to be a chip designer and know basically nothing about manufacturing or be a chip manufacturer know very little about chip design.

3:25 [KARA] All of a sudden, business models changed. The companies that design chips and those that made chips were often two different groups of companies. Eventually, as AI came on the scene, lots of non-tech companies started to think about how their chips are designed and how that design can actually help optimize their processes. But manufacturing that mostly got outsourced, which no surprise, has had upsides and downsides. [KARA] Sanders also had a quote comparing owning a fab to having a shark in your swimming pool.

4:01 The logic being that sharks were very expensive to maintain, and in the end, it might end up eating you anyway. And I think that exposed the challenge of running these very expensive and complex manufacturing facilities even for someone like him who grew up with them. [KARA] The danger now, though, is that much of the chip manufacturing in the world is done in Asia, with many of the world's most complex chips made by one single corporation, which happens to be Taiwan's largest company, TSMC. And that danger, which both Wall Street and Washington have taken note of becomes more and more notable as our reliance on chips in farm equipment and refrigerators, everywhere, just keeps increasing.

4:46 As an example of that proliferation, Miller says, consider cars. [CHRIS] It's not been 100 years since cars have been a widely used consumer product, and for most of the history of cars, you can make a car without making a chip to go inside of it. But today, if you buy a new car, it could often have 1,000 semiconductors inside, and they do a number of different things. They manage the fuel injection into your engine. If you have a gasoline powered car, they make your windshield wipers go back and forth. But the key driver for putting more chips in cars is as you have increasingly semi-autonomous driving systems.

5:23 You not only need some pretty high powered processors and memory to actually undertake the calculations that enable that semi-autonomy, you also need a whole lot of sensors and communications capabilities to bring the data in that eventually gets processed. And so we've seen an explosion in the deployment of semiconductors in cars over the past decade, and it's expected to grow a lot over the next decade as we have more and more semi-autonomous operations. So all of the LiDAR, all of the audio sensors, all of the visual sensors, all of the location sensors and GPS, all that's enabled by chips.

5:56 And that's happening everywhere. It's happening in tractors. It's happening in construction equipment. It's happening in pretty much every type of machine. [KARA] And it's interesting that you pick cars because there's always this Wall Street debate about Tesla. Is Tesla a car company and should it be priced that way, or is it a tech company and should it be priced that way? And in some ways what you're saying is car companies are tech companies in the new world. [CHRIS] Yeah, I think that's right. And it's a challenge for a lot of the legacy car companies to have to update themselves.

6:28 And in the past, you differentiate yourself based on the mechanical engineering in your motor and your drivetrain, for example. And today, that's less the case. There are some people who buy their cars because they want it to accelerate really rapidly. But actually, I think you increasingly see, it's the semi-autonomous driving systems. It's the infotainment systems. These are key differentiating features, and these are all features that are fundamentally enabled by compute. [KARA] So what you've just talked about is this huge ramp up in terms of our dependence on chips, our need for them, the way that they power the economy.

7:04 But we also see that the US, at least starting in about 2000, started making-- or maybe starting in 1990, started making fewer of them. So, you were at about 37% that the US made in 1990, 19% in 2000. By 2010, we were down to about 13% in terms of what we manufactured. Just give me a sense of what was going on there and what's happening now? [CHRIS] Well, I think there were two broad trends that were driving that. One was the construction of these extraordinarily complex supply chains and value chains that are needed to produce smartphones or PCs, or even automobiles today, where you have to assemble hundreds or thousands of different parts from many different types of industries, and that entire process has been increasingly centered in Asia, in Taiwan, in Korea, Japan, and in China.

8:02 And that's a process that has taken decades to develop, but is now impacted to varying degrees, every sector of the industrial base. And its impact on the chip industry, too, because most of the new investment that's come in chip making plants over the last couple of decades has come in Asia, and especially over the last decade or so in Taiwan, which has become the world's leading chip maker. The second thing that's happened simultaneously is that the chip industry itself has become separated into different parts.

8:33 So if you rewind the clock 40 or 50 years ago, you'd find that chip companies designed their own chips, they manufactured their own chips. They produced in-house, the chemicals and materials, you need to make chips. They produce the machines, you need to make chips. And as each of those steps got harder, more complex, more R&D-intensive, more capital-intensive, that process broke apart. And so now you have individual firms that design chips, and then they contract with other firms that manufacture, who themselves contract with other firms that sell them the materials and the chemicals and the gases that are needed.

9:08 And that specialization has been, on the one hand, enormously important for enabling the improvement of technology and enabling us to produce devices at low cost. But it's also led to regional areas of expertise. And so the US has found itself really focusing on the design of semiconductors, which has been a great part of the supply chain to be in. Companies like NVIDIA, AMD, Qualcomm, they do nothing but design chips. They don't manufacture anything today. But that's also caused them to rely on suppliers in Asia, which is focused on the manufacturing of semiconductors.

9:45 And so you've had this specialization that has been differentiated by region, which has been, in my view, both a source of efficiency, but also a source of risk, because you find yourself in situations in which one company or one country is a critical supplier of certain-- of the components or capabilities in the supply chain. [KARA] Yeah, so to that point, you write that you have this great line. If Facebook's fancy headquarters were to sink into the San Andreas Fault, the company might barely notice if TSMC's fabs, which make chips were to slip into the Chelungfu fault, whose movement caused Taiwan's last big earthquake in 1999, the reverberations would shake the global economy.

10:30 That is such a stark contrast between Silicon Valley and Taiwan. What is so crucial about this tiny country and very disputed country off the Coast of China? [CHRIS] Well, I think, Taiwan today is really critical because it's invested over the last 40 years in becoming the world's best, most capable, most technologically advanced and largest chip maker. So around 40 years ago, in 1987, TSMC, the Taiwan Semiconductor Manufacturing Corporation was founded with a radically new business model. Before, as I mentioned, you had companies that both designed and manufactured their own chips, and TSMC decided to do something different.

11:14 It wouldn't design anything, it would only manufacture. And the key insight behind its founding was that if you manufactured for multiple different customers, you could build a lot more scale than any of your competitors. And if you built that scale, you'd both have lower costs because you could buy materials in bulk. But more importantly, you'd have better technology because you'd produce more chips. You'd harvest data from every chip you produced, and you'd use that to hone your own production process. And that's exactly the strategy that TSMC is executed on the last 40 years.

11:46 And so it's reached the position, which is held for now over a decade of being both the most advanced and the biggest producer of processor chips. And today, almost all of America's largest tech companies-- NVIDIA Apple Microsoft, Meta Alphabet they couldn't function without chips from TSMC. And for many of the key processors they purchased, there really is no real alternative. [KARA] So, I mean, in some ways, if we go back to Jerry Sanders view, which we talked about at the beginning, one of the co-founders of AMD.

12:18 In some ways, it feels like it has a lot in common with the current approach of the Trump administration. Andy Grove, as you write about, one of the founders of Intel, basically said, "Levy attacks on offshore labor. And if the result is a trade war, treat it like other wars. Fight to win." That sounds kind of like it could come out of Trump 2.0. Do you feel like there's some sort of comparison between what people used to think about in terms of like, we need to be the manufacturers and what you see now in the current political environment?

12:50 [CHRIS] Yeah, I think it is a really interesting parallel to draw between Andy Grove and some of the rhetoric we see today. I think the interesting dynamic that I always find is when you're in the US looking at the relative decline of the US's chip manufacturing sector, it's easy to lament. If you go to Taiwan, what you hear about is how lucky the United States is to have its extraordinary, vibrant chip design sector. And although TSMC is the world's most important chip manufacturer by a very wide margin, it's less than a third of the value of its most important customers, Apple and NVIDIA.

13:28 And so if you ask yourself what was the smart investment to make if you were going to buy a stock 30 years ago, by a chip manufacturer or by a chip designer? And the answer is pretty clear. The chip designers did much, much better even than the leading chip manufacturer, TSMC. And so the rest of the world looks at our chip design companies with envy because they've done so well financially. And so I have extraordinary respect for Andy Grove, who, as you mentioned, was one of the early figures who really built the chip industry.

14:00 But I think we've got to be balanced in our view of what scale of manufacturing one needs to have in a given country. And you can go too far. It's certainly the case that if you always underestimate manufacturing, if you don't provide the right incentives, if you don't provide a level playing field, if you don't have the workforce that knows how to do it, you'll never build the scale you need to have effective manufacturing businesses. I believe that fully, but I also think we shouldn't understate the importance of having very capable design and software and R&D capabilities because these are also really fundamentally important, economically highly valuable.

14:37 And I think the really important part to me is if you look at high value manufacturing today, it's really inseparable from the R&D and the design and the software capabilities. And that's going to be-- it's true today, but it's going to be even more true as we apply more and more AI to manufacturing capabilities. So to sum it up, I'm sympathetic to the Andy Grove argument, but I'm also hesitant to fully embrace it because I think it doesn't give enough weight to the economic value that's been created by R&D and design focused firms.

15:11 [KARA] Talk to me a little bit about the Trump 2.0 approach and the Biden approach, because both of them work-- seemed to have concerns about A, or lack of manufacturing. But B, you mentioned NVIDIA, the ability of NVIDIA to sell even not its most cutting edge chips to a place like China. So how do you think about the approach that the current administration and maybe to some other degree-- the previous administration had when it comes to chips? [CHRIS] Yeah, I think with the exception of tariffs, where President Trump is somewhat unique, I think there has been a lot of continuity between Trump and Biden.

15:53 And then to be honest, between Biden and Trump's first term. And it's not really about the two presidents. It's actually a bipartisan consensus that's also present in Congress and elsewhere in the US government. I think there are a couple of pillars of this new consensus. I think pillar one is that there's real concern about US reliance on chip source from Taiwan, not because Taiwan itself is not a reliable partner, but because China continues to threaten Taiwan and build up the military capabilities that you would use to disrupt Taiwan's ability to export chips.

16:27 And so there's been a real desire to reduce the instances where the US is solely sourcing chips from Taiwan and encourage Taiwanese companies like TSMC to build capacity elsewhere, including the United States. So that's one. I think the second is really deep concern about the implications of China's very aggressive effort to become a major player in semiconductors. Yeah, right when the made in China 2025 policy was released around the middle of the 2010's, it was clear that semiconductors were a core focus of China's efforts.

17:00 And we've seen China spend tens of billions of dollars a year trying to build its capabilities. And I think the US government realized that just like in almost every other segment of manufacturing, just like in solar panels, just like we've seen in EVs, China was trying to do that in chips, too, and it was going to be harder than in any other segment, because making chips is just very technologically challenging, very capital intensive. But China wasn't to be underestimated. And if the US government didn't do something, one should expect China's market share to grow rapidly at the expense of US and other Western firms.

17:33 So that was a second area of focus. And I think third was that-- well, before the release of ChatGPT, there was a realization in parts of the US government that AI was going to be a really transformative force and that chips were a critical ingredient to AI. And that's why you have six weeks before ChatGPT was released. The US imposed a new set of restrictions on who you could sell AI chips to and who you couldn't on the theory that AI would really matter, and that the US wanted to make sure that its companies and companies and allied countries had better access to AI chips than adversaries did.

18:13 And so, I think those three pillars have been roughly constant between the last three presidencies. [KARA] How far ahead do you think China is of American companies when it comes to chips? We talked about NVIDIA being essentially restricted and seems like increasingly restricted in the current Trump administration from being able to sell even not the cutting edge stuff that they've got, but really kind of older stuff. Do you feel like we're six months ahead they've pulled even, we're years ahead? What's going on? [CHRIS] Well, I think, it really depends on what you're measuring, because depending on which segment of the stack you measure, you get different answers.

18:57 So, if you look at AI model quality, I think what you find is that per most of the benchmarks, Chinese AI models, whether it's DeepSeek or Qwen, they're very close to leading US models and by some benchmarks ahead, some behind. But there's not a lot of differentiation. So, in terms of ability to train an AI model, I think we're at parity or pretty close to it. But if you go a level deeper in the stack and look at the chip level, what you find here is that Chinese firms can design pretty impressive chips. And Huawei's the leading designer of AI accelerator chips in China.

19:34 If you compare them to NVIDIA, they're quite impressive, a bit less energy-efficient, but highly impressive chips. The real challenge that China has is manufacturing chips at scale. And the reason that's a challenge is because in all of China's chip fabs, as well as all of the chip fabs in the rest of the world, you need to acquire these ultra specialized tools, which today are produced basically by five companies. One from the Netherlands, one from Japan, and three from the United States. And unless you can get these tools, you can't make chips at scale, at least at the cutting edge.

20:07 And that's where China's been unable to buy the tools for the last six years now. And because of that, it's been unable to produce chips at large scale. And so if you look at the situation now with AI chips, you mentioned NVIDIA is not able to sell not just it's most advanced, but also second generation chips to China. And the reason that's impactful is because although China can design chips that are roughly comparable, it can't manufacture them at large scale, which is why last year, NVIDIA was the largest supplier of chips to China, even though it was restricted from selling its high end.

20:40 China was buying the second-generation because it couldn't produce enough domestically. [KARA] What is the level of concern? I'm sure you talked to a lot of folks in business. I wonder what the level of concern that you hear is. I certainly have talked to people in tech, I was just at a big biotech convention. All anybody was talking about, it felt like-- or maybe every third thing, people were talking about was China, and how they're really giving American business a run for their money in these technical biomanufacturing.

21:15 But really moving up the chain, much as what we've been talking about. What do you hear when you talk to folks in business? [CHRIS] Yeah, I think in the chip industry, it's a huge focus. On the one hand, China is the largest market for most non-Chinese chip firms. They sell into China in very large volumes because most of the end products that are manufactured, like smartphones or PCs, those are also assembled in China. And so China's ends up being often a third or more of many companies global sales. So they've got to sell to the Chinese market also not viable.

21:49 But they also realize that for most of them, in the long run, at many segments of the supply chain, the competitors that are most worried about are also Chinese because they've watched Chinese firms rapidly move up the value chain, and they've realized that the Chinese government is pushing local firms very hard to buy domestic. So, for example, just in the headlines yesterday, it was reported that China is now encouraging auto companies in China to source 100% of their chips from domestic providers. There's always been some pressure to buy domestic, but that's an acceleration, I think, of the pressure to buy local.

22:27 And that pressure is present throughout the Chinese tech sector, and it's been a key driver of success of Chinese tech firms because they have this big built-in domestic market. And every foreign firm looks at the Chinese market and simultaneously says, it's very large. We've got to sell there. But also in the long run, the Chinese state wants to squeeze us out and replace us with a domestic firm. And that's a very hard dynamic for almost every company to manage. To come back a second to manufacturing in the US, which we touched on.

23:01 And NVIDIA, I should say, makes GPUs. But if we come back to CPUs, you have a chapter with a great name, “How Intel Forgot Innovation.” Now, I would venture to say no company wants to forget innovation. How did that happen? And especially think of this in light of the fact that huge amounts of money have been plowed into Intel in recent years by the federal government to make sure to kind of assure that the US has a strong position in terms of manufacturing. [CHRIS] Well, I think Intel's challenges really predate the last couple of years.

23:37 They were developing throughout the 2010's. And for the last couple of years, Intel's been trying to turn itself around when I think some success, but a whole lot of challenges that persist. I think the key problem that Intel had is that it was too successful for too long, from the middle of the 1980s, all the way up until the middle of the 2010's. Intel was the world's dominant producer of microprocessors that went in PCs and in data centers, the CPU chips, as they're called. And this was the best place to be during those decades, because it was when everyone got a PC and when data centers boomed with CPUs at the center.

24:17 And one of the problems of being successful is that you focus too much on defending your legacy industry, rather than investing in what might come next. And so Intel missed a series of technological shifts and missed the smartphone. It missed AI. And it found itself increasingly being run not by engineers but by the CFO'S office. And as a result, it underinvested. It took too little technical risk and it fell behind its competitors. And so now the companies, they've identified this as the problem. For the last couple of years, they've been trying to invest in new technologies-- invest in new capabilities to turn the ship around.

24:57 But it's a big ship and it turns very slowly. And I think Intel has found itself in a very complicated position where its business model is no longer fit for purpose. And the leadership team has got a really tough task ahead of themselves, which they well know of trying to turn what was once the most important US ship producer around. [KARA] There was a time, I think, when Google Cloud and Microsoft Azure and AWS bought tons from Intel. Is that still the case? [CHRIS] Well, it still is the case, but a couple of things have changed.

25:31 So one, Intel's historic competitor, AMD, which produces the same type of chips, has gotten a lot better over the last half decade at Intel's expense. So the two of them essentially divide up the market for the chips that go in PCs and the CPU chips that go in data centers. That's first. Second is that AI has changed the way data centers are designed fundamentally. And the types of workloads that you need on AI, especially LLMs, are much, much faster and more efficient on GPU chips, not CPU chips. And Intel hasn't had a competitive GPU.

26:05 It's been NVIDIA that's dominated that market. And so, you've seen a shift in data center spending away from CPUs towards GPUs. So, NVIDIA has been by far the biggest beneficiary of the boom in AI data center construction. Not because there aren't any Intel chips inside those data centers there are, but because the most important of those chips are now produced by NVIDIA, not by Intel. [KARA] When you see the capital expenditure budgets of big tech companies, which are things like $60, $70, $80 billion a year, and a lot of that is going towards these cutting edge NVIDIA chips like Hopper in 2022 or Blackwell in 2024, do you ever think to yourself this kind of money and this kind of race to get cutting edge chips this cannot be sustained?

27:00 [CHRIS] Well, I think it all depends on how much AI we all use and how much money companies can make off it. Right now they're all betting. The hyperscalers are betting the AI labs are betting that if they build better and better AI systems, we will find more and more economically valuable ways to use them. And therefore, we'll be willing to pay for those services. It's true there's an extraordinary amount of investment, but I think it's also true that there's an extraordinary set of new capabilities that you don't have to be a believer in the imminence of artificial super intelligence to think that there's going to be thousands of dollars per person or productivity gains created by the types of AI systems that we're seeing developed right now.

27:43 And if that materializes, then all this investment will, I think, look pretty reasonable. So, we got to be careful. I think there's certainly been a boom and bust cycles in the chip industry in the past and in the data center industry in the past. But when I look at AI uptake, when I look at the growth rates of both paid and unpaid users for AI services like ChatGPT, when I look at all of the ways in which AI is being deployed to industries in which you might not expect AI to be central, but in fact, it's proving quite productive.

28:19 To me, it all seems like there's an extraordinary amount of focus and investment in speed in building out these capabilities. And so I'm not sure that I would see all this investment as a bubble even though the scale is quite large. [KARA] I assume you started working on this book before AI was on everybody's lips, right? [CHRIS] Well before, yes, I started in the middle of the 2010's. So, half decade before ChatGPT was released. [KARA] And what have you thought as this has happened and unfolded? Well, it's certainly been surprising the speed over the past three years at which AI has become something that is front and center in every industry that you now have AI companies valued in the hundreds of billions of dollars that I think has surprised everyone, certainly surprised me.

29:13 I think some of the longer term trend lines you could have visualized or seen them developing. One would have been the shift from CPU to GPU, which was already certainly happening before the last couple of years and the investment boom. I think the second-- the intuition behind all this investment that if you want better AI systems, you'll need to have bigger AI systems, the scaling laws, which thus far still seem to be very much in play as the reasoning models that we're using more and more of-- require more and more compute, which means more and more semiconductors.

29:50 [KARA] And do you think your thinking on how things will go as we move forward, diverges from the average person and maybe in business watching this happen because of what you know about how we got to now? [CHRIS] I think one of the things you do see historically in the computing industry is cycles of centralization and decentralization of computing. And I think right now we're seeing just tons and tons of investment in these vast data centers, which is representing a centralization of computing resources. And I think as we discussed, that's certainly needed.

30:23 But I wouldn't be surprised to see the pendulum at some point swing the other direction and also see a wave of investment in decentralized computing for AI and devices, having many specialized AI processors in your phone or in your PC, or in your car, or in other devices. And there will be a need for both centralized and decentralized. The benefit of centralized is the efficiency. The benefit of decentralized is that you'll have lower latency, more privacy if it's all kept on your device, less risk of disruptions if your connection with the center of the network is disrupted.

30:59 So in the end, you'll want a bit of both centralization and decentralization. I think thus far, we've seen the vast data centers. We've seen less of a change so far on the edges of networks like in our phones or in our PCs. [KARA] One thing that you write about is that many large tech companies, while they are pursuing this path of spending billions of dollars buying cutting edge chips from places like NVIDIA and AMD, they are also places like Amazon, for example, are working on making their own custom chips. Do you think that threatens the places that their whole business is making chips?

31:38 [CHRIS] Well, it certainly is a challenge, a competitive challenge. I think, for a company like NVIDIA, they were not surprised to see their customers also beginning to put resources into developing their own chips. No one wants to have a sole supplier. And for some time, NVIDIA felt almost sole supplier of the GPUs that AI requires. And so when you've got firms valued at $1 or $2 or $3 trillion, like the big hyperscalers not shocking that they're going to try to see if they can build up some of their own internal capabilities to-- at the very least, give them an alternative to NVIDIA, even if they still end up buying a large number of NVIDIA's chips, which I suspect that they will.

32:20 They want a bit of optionality, and their in-house chip design capabilities give them that. [KARA] For companies that don't make their own chips, let's say large financial companies, large health care, retail, how much concern is there about escalating costs, the pressure to buy cutting edge technology if you don't do it, but then your competitor does? I wonder if that's an anxiety that companies have you think? [CHRIS] Well, I think it is certainly anxiety. Because of that, we've seen actually in a number of industries companies not designing their own chips, but getting much closer to the chip design process so that they can most efficiently integrate semiconductors into their overall system.

33:05 So the auto industry is a great example. Tesla is kind of unique in designing its own chips, but pretty much every car company has done a lot more systematic thinking over the past couple of years, and how the chips that they buy integrate into their software, integrate into their systems. In drones, you see a similar example. You've got to really optimize your data interconnect. You got to really optimize your sensors, your battery life, all that is shaped by the chips that you buy. In certain types of industrial equipment as well, we see much more serious thinking going into the types of chips you buy and the way you integrate them.

33:44 And so I think across complex industrial systems, we see more than ever before serious thinking going into if you choose X or Y chip, how does that impact? How your system functions? And I think a lot of companies are discovering that. Although in the past they thought if they just ignored the chip level, just bought something off the shelf and focused on the software, focused on the system, it would be more efficient because they wouldn't have to understand the chip level. Now, they realize they get performance gains from thinking about the entire tech stack.

34:15 And so that's driven a range of sectors that in some ways are surprising to get involved in thinking about chip design, and chips as an important part of systems design more than ever before. [KARA] It recalls a little bit the Marc Andreessen line of Software's Eating the World, it's like, no matter what business you're in, you're in the software business, or in this case, no matter what business you're in, you're in the chip business even though that wasn't the business you thought you were in. [CHRIS] Yeah, I think that's exactly right.

34:44 And for a lot of companies, it's a bit of a culture shock and a change that's stressed their existing talent pool because they've hired the software engineers, they've hired the other engineers they need, and they now need to find people who understand the chip level, too. [KARA] We were talking before about competitiveness vis a vis China, and you quote a China focused think tank as saying about 30% of leading AI researchers are from China versus about 20% from the US. But I think this is like the key thing. The US employs nearly 60% of the top researchers.

35:20 When you see a crackdown-- an increasing crackdown on, like a fewer foreign students in higher education in the US, foreign workers in general. I mean, it's even more acute at a place like Harvard, for example. Does that concern you? Do you worry that that's an important part of our tech edge? [CHRIS] I think it's a central part of our technology edge. And it's been that way since the origin, certainly the chip industry. And if you look at the first chip startup, Fairchild Semiconductor, found in the 1950s. Of the eight founders, two of them were not born in the United States, and that same ratio was more or less present at all of the key chip companies.

36:05 We talked about Andy Grove. Andy Grove was born in Budapest, Hungary, fled as a refugee from Communist oppression there. And I could give you about 1,000 other examples of key researchers or executives who have helped build the US chip industry as immigrants. So, I think it is hugely counterproductive. I think if you talk to anyone in the chip industry, it's hard to find anyone who disagrees with that statement. And I think what's really striking is when I go to other countries and talk to their chip industries and ask them, what are your challenges.

36:41 I repeatedly hear whether I'm in India or in East Asia or in Europe, one of the biggest challenges is that our best engineers want to go to the US. Our best startup founders want to raise their money in the US. And to take any steps to call that into question or to make that more difficult seems to me to be just completely self-defeating. [KARA] Is your sense that there will be a shift? I mean, we have seen it depends on what the technology is or what domain of science. But France, Ireland, China, they've all kind of done some version of rolling out the red carpet and saying, please, scientists and tech folks, come here.

37:24 So I wonder if you think that's in the process of shifting. [CHRIS] Well, I think I'm not sure how big the magnitude is. I think certainly the steps we've seen from the administration on student visas, on just calling into question the visa status of many different people. That's certainly been a real shock, I think, to both the academic community, but also the AI research community is such a large proportion of people born outside of the United States there, too. It's been impactful in the chip industry more broadly.

37:56 I'm not sure yet of the magnitude. I think in the first Trump administration, we saw some pretty harsh rhetoric on immigration and some moves on visas that I think ended up not being as impactful for these sectors as one might have expected at the time. So, I'm going to hold judgment on just the magnitude of impact as I see how it develops. But whether it's low impact or high impact, it's negative impact either way. And so, I think, either way, I'm opposed to it. [KARA] In 2021, the former Google CEO, Eric Schmidt wrote, "China could surpass the US as the world's AI superpower."

38:34 I wonder when you think back to 2021, if you think that that seems more or less likely now than it did then. [CHRIS] Well, I think China's super impressive in many different ways. It's overperformed many people's estimations repeatedly over the last couple of decades and basically, every industry. But I do think that any sort of net assessment of AI capabilities looking at the software and the hardware stack, looking at which companies have market share, looking at who's doing the fundamental research and the US is still in a pretty strong position, which doesn't mean that if the US squanders it by chasing away the world's smartest people, we're going to stay in the strongest position.

39:19 But it does mean if we play our cards wisely, I think, US researchers and US firms are likely to retain their lead in many of the key segments. But that doesn't mean I'd underestimate China and Chinese firms and the extraordinary amount of talent that is in China, because it certainly is impressive. [KARA] What do you find that people-- when you speak to audiences generally don't understand about chips that you've come to understand by writing this. But maybe in the public or in the business community is there's a gap?

39:52 [CHRIS] I think the biggest thing is just the complexity of manufacturing and cutting-edge chip. I often hear people say, well, we're so reliant on Taiwan, for example, why don't we just build a couple factories here? Each factory costs $20 billion. Each factory is the most expensive factory in human history. And inside, they've got the most expensive and complicated machines that have ever been invented. Manufacturing chips with billions and billions of transistors at microscopic scale. There's really nothing more complex than humans have ever done than produced advanced chips.

40:25 And I think one of the reasons chips have been so impactful is that the rest of us can take it for granted, because they just magically appear in all our devices, magically give us computing capabilities and communications capabilities and sensor capabilities that we don't have to think about. And when we do think about them, we begin to realize just what a miracle it's been that it's possible to create such complicated machines at such a low price point. We can fit one in our pocket. [KARA] But it's not a build it by next week kind of situation?

40:54 [CHRIS] Certainly not. [KARA] And finally, we've talked about a whole bunch-- a whole range of things. I wonder what your own sort of greatest concern is going forward when you think about chips and AI and this whole sector that you know so much about. [CHRIS] Well, I do think we can't underestimate the significance of the supply chain vulnerabilities that we've got, especially in East Asia and never before in world history has an economic revolution been so dependent on one country to produce its critical ingredients.

41:31 And so I think that is something that deserves immediate attention and more attention than we collectively have given it over the last couple of years to reduce those dependencies. I think the other thing that I'm struck by when you look at the development of AI is that the biggest block by far in the deployment of AI is going to be not computers or technologies or models. It's going to be people and making sure we're designing AI systems in a way that people will find them easy to use, and that we're encouraging people not to be afraid, but rather to take advantage.

42:04 Because I think one of the things you find historically is that new technologies often run into lots of stubborn people in companies and in educational institutions, in governments across society, who often delay the adoption of technology, sometimes for good reasons, but I think sometimes just for stubbornness. And I hope we don't see too much of that here. [KARA] Chris Miller is author of Chip War, The Fight for the World's Most Critical Technology. He's a professor of international history at Tufts. Chris, thanks so much.

42:36 This is great. [CHRIS] Well, thank you for having me. [KARA] And before we go today, if you enjoyed the show, we would love for you to leave a comment on Apple podcasts. It helps other listeners discover us. On our website, we've got a lot more about CSAIL and how you can get involved. Just head to csail.mit.edu/podcast. Through CSAIL Alliances companies can access research, talent, conferences, courses, a whole lot more. Again, the website csail.mit.edu/podcast. I'm Kara Miller. Our podcast is produced by Matt Purdy and Andrew Zucosky with help from Audrey Woods.

43:24 Join us again next time and stay ahead of the curve.

Questions about this episode

Who speaks in "The Promise and Peril of Computer Chips with Tufts Professor Chris Miller"?

Kara Miller (Host), Chris Miller (Guest).

What is "The Promise and Peril of Computer Chips with Tufts Professor Chris Miller" about?

The discussion focuses on the pivotal role of semiconductors in shaping modern business and everyday life, emphasizing both the opportunities and challenges faced by the U.S. in maintaining its competitive edge in the global chip market. Chris Miller highlights the complexities of international supply chains and the pressures from emerging markets, especially China, while asserting that the U.S.…

What is mentioned in this MIT CSAIL Alliances episode?

10 people, works and ideas are mentioned, including Artificial Intelligence, Chip War, MIT, AMD, Apple Inc., GPUs. Each one on this page links to its own page with every episode that mentions it.

How long is "The Promise and Peril of Computer Chips with Tufts Professor Chris Miller"?

The episode runs 44 min; the transcript is about 7,274 words in 70 timestamped sections.

More from MIT CSAIL Alliances

• All analyzed episodes
• People mentioned on MIT CSAIL Alliances
• Topics mentioned on MIT CSAIL Alliances