WEBVTT - Chris Miller Talks AI Demand, Chips

0:00:00.280 --> 0:00:07.240
<v Speaker 1>Bloomberg Audio Studios, podcasts, radio News.

0:00:08.080 --> 0:00:11.000
<v Speaker 2>Hey, I got to say that there are moments in

0:00:11.039 --> 0:00:13.640
<v Speaker 2>time like today, with all the spend and build out

0:00:13.680 --> 0:00:18.120
<v Speaker 2>and companies doing things related to AI and questions around

0:00:18.160 --> 0:00:20.520
<v Speaker 2>the return on investment, the ROI and the impact of

0:00:20.680 --> 0:00:23.720
<v Speaker 2>artificial intelligence, that we kind of wish we could pull

0:00:23.760 --> 0:00:25.480
<v Speaker 2>out that crystal ball and see a little bit into

0:00:25.480 --> 0:00:25.880
<v Speaker 2>the future.

0:00:26.000 --> 0:00:27.880
<v Speaker 1>I think it's fair to say Chris Miller saw something

0:00:28.080 --> 0:00:31.320
<v Speaker 1>years ago. He's the author of the twenty twenty two

0:00:31.480 --> 0:00:34.640
<v Speaker 1>bestseller Chip War, The Fight for the World's Most Critical Technology.

0:00:34.840 --> 0:00:38.320
<v Speaker 1>He joins us from Pittsburgh this afternoon. Chris, it's good

0:00:38.320 --> 0:00:39.760
<v Speaker 1>to have you back with us. Carol and I were

0:00:39.760 --> 0:00:43.040
<v Speaker 1>just talking, and you know, you wrote this book close

0:00:43.080 --> 0:00:47.400
<v Speaker 1>to four years ago. At this point, how different is

0:00:47.440 --> 0:00:52.280
<v Speaker 1>the world of semiconductors and the world's reliance on semiconductors

0:00:52.640 --> 0:00:54.720
<v Speaker 1>now than when you wrote the book.

0:00:57.480 --> 0:00:59.680
<v Speaker 3>Well, I think the key difference is just the scale

0:00:59.720 --> 0:01:02.640
<v Speaker 3>of spending that we've got right now on building data

0:01:02.680 --> 0:01:06.199
<v Speaker 3>centers and buying all of the chips that are required

0:01:06.400 --> 0:01:09.160
<v Speaker 3>inside of them, and Nvidia first and foremost, but not only,

0:01:09.200 --> 0:01:12.440
<v Speaker 3>we've seen companies transformed by the data center build out

0:01:12.600 --> 0:01:14.440
<v Speaker 3>but in a lot of ways, not much just changed

0:01:14.480 --> 0:01:18.360
<v Speaker 3>other than that we're still sourcing our key semi conductors

0:01:18.400 --> 0:01:22.160
<v Speaker 3>from Asia and above all from Taiwan, and so the

0:01:22.200 --> 0:01:26.560
<v Speaker 3>industry is supercharged, its size, growing faster than ever, but

0:01:26.600 --> 0:01:29.000
<v Speaker 3>dealing with some of the same supply chain choke points

0:01:29.000 --> 0:01:30.280
<v Speaker 3>that it had a decade ago.

0:01:30.680 --> 0:01:34.600
<v Speaker 2>You know, we talk a lot with one of our

0:01:34.600 --> 0:01:37.600
<v Speaker 2>brilliant voices on the chip world, and that is our

0:01:37.640 --> 0:01:39.800
<v Speaker 2>own Ian King, who has seen a lot of cycles,

0:01:39.840 --> 0:01:41.559
<v Speaker 2>the ups, the downs, and we talk a lot about

0:01:41.560 --> 0:01:44.720
<v Speaker 2>that supply demand imbalance and what happens when there's a

0:01:44.760 --> 0:01:47.760
<v Speaker 2>lot of demand, there's not enough supply. The buildout continues,

0:01:47.800 --> 0:01:50.600
<v Speaker 2>the investment supply goes up, and then there's a glut.

0:01:50.960 --> 0:01:54.000
<v Speaker 2>So I'm just curious what your view is when it

0:01:54.040 --> 0:01:58.160
<v Speaker 2>comes to the cyclical nature of the semiconductor world. Has

0:01:58.240 --> 0:02:02.080
<v Speaker 2>AI changed it or we we'll see also and over

0:02:02.160 --> 0:02:04.600
<v Speaker 2>build and oversupply and then a glut.

0:02:06.480 --> 0:02:09.400
<v Speaker 3>I think if you look historically, certainly you see cycles up,

0:02:09.440 --> 0:02:12.840
<v Speaker 3>cycles down, but there are moments when you see step

0:02:12.960 --> 0:02:16.680
<v Speaker 3>changes in terms of demand for certain types of chips.

0:02:16.720 --> 0:02:19.320
<v Speaker 3>We saw that with smartphones, for example, there was no

0:02:19.360 --> 0:02:22.120
<v Speaker 3>demand for smartphone chips and then now everyone needs a

0:02:22.120 --> 0:02:24.800
<v Speaker 3>new smartphone every couple of years, and what we're seeing

0:02:24.800 --> 0:02:26.840
<v Speaker 3>in AI right now is that type of step change,

0:02:26.880 --> 0:02:29.840
<v Speaker 3>a huge increase in just the baseline amount of chips

0:02:29.840 --> 0:02:32.680
<v Speaker 3>that we're going to need for data centers driven by AI,

0:02:32.840 --> 0:02:34.720
<v Speaker 3>and so I think we shouldn't expect cycles to be

0:02:34.760 --> 0:02:37.560
<v Speaker 3>over Certainly there'll be ups and downs in the future,

0:02:37.680 --> 0:02:40.359
<v Speaker 3>but it's now very clear. I think that we're just

0:02:40.400 --> 0:02:43.359
<v Speaker 3>going to need a lot more compute for AI purposes

0:02:43.400 --> 0:02:45.040
<v Speaker 3>in the future, and as a result, will need a

0:02:45.040 --> 0:02:46.880
<v Speaker 3>lot more of the AI chips that go inside of

0:02:46.919 --> 0:02:47.480
<v Speaker 3>data centers.

0:02:47.919 --> 0:02:51.640
<v Speaker 1>In your view, is the promise that many people think

0:02:52.080 --> 0:02:57.800
<v Speaker 1>AI is supposed to deliver, Will it actually be delivered? Like,

0:02:57.880 --> 0:03:02.000
<v Speaker 1>what's the future that you envision after all this capex

0:03:02.080 --> 0:03:02.440
<v Speaker 1>is spent?

0:03:04.600 --> 0:03:06.880
<v Speaker 3>Why? I think it's in some ways funny when we

0:03:06.919 --> 0:03:09.480
<v Speaker 3>ask will AI deliver? Because if you look just three

0:03:09.560 --> 0:03:12.400
<v Speaker 3>years ago, even after chat schipt was released, there are

0:03:12.440 --> 0:03:16.960
<v Speaker 3>so many things that were not possible that are possible today,

0:03:17.280 --> 0:03:20.920
<v Speaker 3>whether it's the number or the scale of hallucinations and

0:03:21.400 --> 0:03:24.360
<v Speaker 3>answers that chatbots give you, or the scale of work

0:03:24.400 --> 0:03:27.520
<v Speaker 3>that you can put together today that just wasn't possible

0:03:28.200 --> 0:03:30.800
<v Speaker 3>in the past. We've already had so much new capability

0:03:30.840 --> 0:03:34.200
<v Speaker 3>generated just in the couple of years since Chat schipt

0:03:34.360 --> 0:03:37.000
<v Speaker 3>that in some ways, I think it's an absurd question

0:03:37.080 --> 0:03:38.920
<v Speaker 3>to ask will AI deliver it already has in a

0:03:38.960 --> 0:03:41.720
<v Speaker 3>lot of ways. But I think I understand why there's

0:03:41.760 --> 0:03:44.280
<v Speaker 3>plenty of questions about what about the investment is happening

0:03:44.400 --> 0:03:48.040
<v Speaker 3>right now, will that investment pay off? And I think

0:03:48.040 --> 0:03:52.600
<v Speaker 3>there are reasons to think carefully about how much each

0:03:52.640 --> 0:03:56.080
<v Speaker 3>company is putting in over what time horizon. But I

0:03:56.080 --> 0:03:58.840
<v Speaker 3>guess when I zoom out, I ask myself, do I

0:03:58.840 --> 0:04:00.360
<v Speaker 3>want a world in which there is more access to

0:04:00.400 --> 0:04:04.440
<v Speaker 3>compute or last? And it seems to me that we should,

0:04:04.720 --> 0:04:07.080
<v Speaker 3>rather than being concerned that there's too much compute being

0:04:07.520 --> 0:04:10.720
<v Speaker 3>put in the ground, be excited that there's actually companies

0:04:10.760 --> 0:04:12.960
<v Speaker 3>willing to invest in the infrastructure that's going to deliver

0:04:13.400 --> 0:04:14.360
<v Speaker 3>all these capabilities.

0:04:14.480 --> 0:04:16.120
<v Speaker 1>Well, I think it's an important question to ask for

0:04:16.160 --> 0:04:18.360
<v Speaker 1>a few reasons, and one of those is because if

0:04:18.400 --> 0:04:20.839
<v Speaker 1>you look at just the capex that these companies are spending,

0:04:20.880 --> 0:04:23.800
<v Speaker 1>you know, one company two hundred billion dollars in a

0:04:23.839 --> 0:04:27.440
<v Speaker 1>single year. That's a big commitment, and that's something that

0:04:27.440 --> 0:04:30.479
<v Speaker 1>they have to convince investors that is money that is

0:04:30.520 --> 0:04:34.160
<v Speaker 1>going to the right place, that's money that has to

0:04:34.200 --> 0:04:38.120
<v Speaker 1>be earned back. Plus on companies coming to a hyper

0:04:38.120 --> 0:04:41.599
<v Speaker 1>scaler and saying, yes, we think that this money is

0:04:41.640 --> 0:04:43.640
<v Speaker 1>not only being well spent, but then we can use

0:04:43.680 --> 0:04:47.200
<v Speaker 1>this compute to actually create a product that will provide

0:04:47.200 --> 0:04:49.599
<v Speaker 1>a return on investments. So, yes, we've seen a lot

0:04:49.640 --> 0:04:52.320
<v Speaker 1>of the hyperscalers benefit the you know, the anthropics and

0:04:52.360 --> 0:04:55.960
<v Speaker 1>the open ais and the mistraws like that's amazing stuff.

0:04:56.000 --> 0:04:57.920
<v Speaker 1>But at the end of the day, there are a

0:04:58.040 --> 0:05:01.760
<v Speaker 1>lot of companies that are not necessarily technology companies that

0:05:02.080 --> 0:05:06.920
<v Speaker 1>maybe aren't yet necessarily seeing an increase in productivity as

0:05:06.960 --> 0:05:10.400
<v Speaker 1>a result of these tools. So my question is do

0:05:10.520 --> 0:05:12.800
<v Speaker 1>those companies start to see that do we live in

0:05:12.839 --> 0:05:15.160
<v Speaker 1>a world where this is a layer, just like the

0:05:15.160 --> 0:05:18.839
<v Speaker 1>Internet was a layer of technology.

0:05:19.000 --> 0:05:20.640
<v Speaker 3>I think there's no doubt that we're going to have

0:05:20.720 --> 0:05:24.320
<v Speaker 3>AI as a layer that's embedded into all technology that

0:05:24.480 --> 0:05:27.039
<v Speaker 3>we use. And when I look at the economic question,

0:05:27.240 --> 0:05:30.039
<v Speaker 3>I say, first off, is it profitable to serve AI

0:05:30.120 --> 0:05:32.880
<v Speaker 3>systems today, not train the next generation, but serve today.

0:05:33.040 --> 0:05:36.520
<v Speaker 3>And if you listen to what's publicly reported about OpenAI

0:05:36.720 --> 0:05:40.480
<v Speaker 3>or anthropic, their margins on their inference business are not

0:05:40.560 --> 0:05:43.440
<v Speaker 3>just positive but quite good, and so that I think

0:05:43.520 --> 0:05:46.360
<v Speaker 3>is pretty strong evidence that the delivery of already existing

0:05:46.400 --> 0:05:50.479
<v Speaker 3>AI services is a pretty profitable business. The next question

0:05:50.560 --> 0:05:52.320
<v Speaker 3>is should we be investing in R and D in

0:05:52.360 --> 0:05:56.200
<v Speaker 3>the next generation, which is of course very very expensive.

0:05:56.320 --> 0:05:59.160
<v Speaker 3>But I think it's hard to argue we should dramatically

0:05:59.200 --> 0:06:01.160
<v Speaker 3>slow down R and D and AI just given all

0:06:01.200 --> 0:06:04.080
<v Speaker 3>of the extraordinary improvements that we've seen over the past

0:06:04.160 --> 0:06:06.680
<v Speaker 3>couple of years in terms of capability. And so when

0:06:06.720 --> 0:06:08.760
<v Speaker 3>you start breaking it down and ask, well, which specific

0:06:08.800 --> 0:06:12.400
<v Speaker 3>investment do we think is excessive? Which specific investment would

0:06:12.480 --> 0:06:14.680
<v Speaker 3>rather not be doing? Would you really like to be

0:06:14.760 --> 0:06:17.400
<v Speaker 3>the CEO of the only big tech company that's not

0:06:17.480 --> 0:06:19.640
<v Speaker 3>investing in AI. That doesn't seem like a very comfortable

0:06:20.160 --> 0:06:22.159
<v Speaker 3>place to be. And I'd rather have a situation in

0:06:22.200 --> 0:06:25.520
<v Speaker 3>which the big technology companies are investing more rather than

0:06:26.160 --> 0:06:28.480
<v Speaker 3>putting money in the bank or buying back their stock

0:06:28.480 --> 0:06:31.040
<v Speaker 3>because they didn't have any comfitable by investment opportunity.

0:06:31.120 --> 0:06:33.159
<v Speaker 2>Well, you know, I would say Apple's doing it differently,

0:06:33.200 --> 0:06:35.200
<v Speaker 2>and I guess time will tell whether or not their

0:06:35.240 --> 0:06:39.280
<v Speaker 2>approach works out. Chris, I am curious to you know,

0:06:40.160 --> 0:06:42.200
<v Speaker 2>who do you talk to, what research do you follow,

0:06:42.240 --> 0:06:44.480
<v Speaker 2>what are the leading voices, what are the leading companies

0:06:44.480 --> 0:06:49.200
<v Speaker 2>that you watch to figure out kind of how the

0:06:49.200 --> 0:06:52.920
<v Speaker 2>semiconductor space is evolving. I mean, we still know TSMC

0:06:53.320 --> 0:06:56.120
<v Speaker 2>is still the big manufacturer of all chips in the world,

0:06:56.160 --> 0:06:58.880
<v Speaker 2>and there's geopolitical attentions to that. But give us an

0:06:58.920 --> 0:07:02.839
<v Speaker 2>idea of where do you think kind of investors and

0:07:02.920 --> 0:07:05.960
<v Speaker 2>just the world at large need to be focusing their

0:07:06.040 --> 0:07:09.520
<v Speaker 2>attention on. When it comes to the semiconductor world. Is

0:07:09.520 --> 0:07:12.000
<v Speaker 2>it Nvidia, is it China? Is it something else?

0:07:13.240 --> 0:07:15.360
<v Speaker 3>I think the hard problem is that it's all of

0:07:15.400 --> 0:07:18.520
<v Speaker 3>the above, and we've seen this play out in the

0:07:19.000 --> 0:07:21.760
<v Speaker 3>GPU supply chain over the past couple of years, where

0:07:21.760 --> 0:07:26.040
<v Speaker 3>we've had shortages and different types of components, different materials,

0:07:26.120 --> 0:07:29.720
<v Speaker 3>and the memory chips that go next to GPUs and

0:07:29.760 --> 0:07:33.560
<v Speaker 3>AI servers. Each part of the supply chain has had

0:07:33.600 --> 0:07:37.960
<v Speaker 3>to dramatically ramp up its production capacity to respond to

0:07:38.000 --> 0:07:39.680
<v Speaker 3>the surge and demand. And so if you only look

0:07:39.680 --> 0:07:43.000
<v Speaker 3>at one part of the supply chain, you miss the

0:07:43.120 --> 0:07:47.160
<v Speaker 3>challenges that other parts are often facing. And so it's

0:07:47.240 --> 0:07:49.600
<v Speaker 3>the chip designers. It's the chip makers, but it's also

0:07:49.640 --> 0:07:53.400
<v Speaker 3>the materials suppliers, which are often not even thought of

0:07:53.440 --> 0:07:56.800
<v Speaker 3>as being semiconductor firms, but produce many of the capabilities

0:07:56.800 --> 0:07:59.240
<v Speaker 3>that are critical to actually manufacture the AI chips and

0:07:59.280 --> 0:08:00.840
<v Speaker 3>servers that we need.

0:08:01.040 --> 0:08:02.680
<v Speaker 2>Do you think the world do you think the US

0:08:02.760 --> 0:08:05.280
<v Speaker 2>specifically does need to be restricting sales of its most

0:08:05.280 --> 0:08:09.200
<v Speaker 2>sophisticated chips to China or on others? I mean, we

0:08:09.200 --> 0:08:12.600
<v Speaker 2>did see Nvidia. They still face the uncertainty in China,

0:08:12.720 --> 0:08:16.120
<v Speaker 2>that's their largest market, or which it is the largest

0:08:16.120 --> 0:08:20.000
<v Speaker 2>market for chips. The government granting some licenses to ship

0:08:20.000 --> 0:08:22.720
<v Speaker 2>a small amount of some of its processors to customers there,

0:08:22.760 --> 0:08:25.080
<v Speaker 2>but and Vidia is not sure if the Chinese government

0:08:25.120 --> 0:08:27.120
<v Speaker 2>will give its approval. So there's still some back and

0:08:27.160 --> 0:08:32.320
<v Speaker 2>forth here. But is it still an arms race of source.

0:08:34.040 --> 0:08:35.240
<v Speaker 3>Arms race of sorts?

0:08:35.240 --> 0:08:35.720
<v Speaker 2>Excuse me?

0:08:35.840 --> 0:08:39.199
<v Speaker 3>Yeah, I think armed races is not a bad analogy.

0:08:39.280 --> 0:08:42.200
<v Speaker 3>When you listen to tech CEOs, they'll speak in the

0:08:42.200 --> 0:08:44.840
<v Speaker 3>same language. They're struggling to get access to all the

0:08:44.840 --> 0:08:48.400
<v Speaker 3>computing power that they need, that they envision needing more

0:08:48.440 --> 0:08:50.960
<v Speaker 3>tomorrow than they've got today. And if you listen to

0:08:51.040 --> 0:08:54.160
<v Speaker 3>Chinese technology leaders. What you hear from them is challenges

0:08:54.200 --> 0:08:58.520
<v Speaker 3>and getting access to computing power. And the primary reason

0:08:58.559 --> 0:09:03.080
<v Speaker 3>they've struggled to deploy AI products at scale is because

0:09:03.080 --> 0:09:05.720
<v Speaker 3>they've struggled to get access to all of the computing capabilities.

0:09:05.760 --> 0:09:07.960
<v Speaker 3>And it's on a regular basis we see new Chinese

0:09:08.000 --> 0:09:12.120
<v Speaker 3>models launched that can actually be deployed at scale because

0:09:12.160 --> 0:09:15.000
<v Speaker 3>they don't have access to the advanced chips that they need.

0:09:15.120 --> 0:09:17.080
<v Speaker 3>So this does, I think, seem to me that this

0:09:17.160 --> 0:09:20.360
<v Speaker 3>is still a very powerful card that the US has

0:09:20.440 --> 0:09:21.959
<v Speaker 3>to play, and so I think we should be very

0:09:21.960 --> 0:09:26.000
<v Speaker 3>careful or on any decisions to give China more access

0:09:26.080 --> 0:09:28.160
<v Speaker 3>or at least make sure that we're getting something in

0:09:28.200 --> 0:09:28.840
<v Speaker 3>exchange for that.

0:09:29.280 --> 0:09:34.120
<v Speaker 1>On that, Chris, the US support of TSMC and the

0:09:34.280 --> 0:09:38.520
<v Speaker 1>US support of Intel different types of support, And I

0:09:38.559 --> 0:09:41.520
<v Speaker 1>guess you could call the TSMC one encouragement to build

0:09:41.600 --> 0:09:44.280
<v Speaker 1>here in the United States. What is the right industrial

0:09:44.320 --> 0:09:48.960
<v Speaker 1>policy to reduce reliance on companies outside of the US.

0:09:51.240 --> 0:09:54.120
<v Speaker 3>It's a really hard problem because TSMC is such a

0:09:54.160 --> 0:09:58.800
<v Speaker 3>capable and efficient manufacturer in their home base in Taiwan.

0:09:58.880 --> 0:10:01.400
<v Speaker 3>But I think the US government is right to say

0:10:01.400 --> 0:10:04.680
<v Speaker 3>that we need a more diversified manufacturing base for the

0:10:04.679 --> 0:10:07.800
<v Speaker 3>world's most important semiconductors. I think we've learned over the

0:10:07.840 --> 0:10:10.679
<v Speaker 3>last couple of years there's no silver bullet. President Trump's

0:10:10.679 --> 0:10:14.640
<v Speaker 3>tried tariffs that comes with obvious downside. President Biden tried

0:10:15.240 --> 0:10:19.600
<v Speaker 3>subsidies that worked to a degree, but only to a degree.

0:10:19.800 --> 0:10:23.400
<v Speaker 3>Here's the reality. The chip industry has involved hundreds of

0:10:23.400 --> 0:10:27.120
<v Speaker 3>billions of dollars of cappax over the last several decades,

0:10:27.120 --> 0:10:29.320
<v Speaker 3>and so it's just not going to move fast. And

0:10:29.360 --> 0:10:32.240
<v Speaker 3>so if you want to slowly change the structure of

0:10:32.240 --> 0:10:35.840
<v Speaker 3>where chips are produced, you've got a plan for years

0:10:35.880 --> 0:10:39.959
<v Speaker 3>of years of implementation of measures designed at the shift

0:10:39.960 --> 0:10:41.200
<v Speaker 3>economics of the GIP industry.

0:10:41.440 --> 0:10:43.439
<v Speaker 2>All right, Kenly with there, Chris Miller, thank you so much.

0:10:43.480 --> 0:10:45.800
<v Speaker 2>Good to check in with you. Professor of International History

0:10:45.800 --> 0:10:49.600
<v Speaker 2>at the Fletcher School at Tufts University, author of Chipward

0:10:49.720 --> 0:10:51.839
<v Speaker 2>joining us on this Thursday.