- 18 hours ago
From chips everywhere to AI everywhere, the world is entering a new phase of digital transformation. In this keynote, Christophe Fouquet, CEO of ASML, explores how rapid advances in artificial intelligence are driving unprecedented demand for computing power, and what it takes to scale it. From semiconductor innovation to global supply chains and collaborative ecosystems, he will outline how the technology behind microchips is enabling the AI era and shaping critical applications across society, from healthcare and energy to mobility and connectivity.
Category
🤖
TechTranscript
00:00Hello everyone, and thank you very much for joining me.
00:05I hope that a few of you at least had the right answer to the question about what is a
00:12wafer, what is a EUV machines,
00:15because if not, I'm afraid it's going to be a very long presentation for some of you.
00:20So I'm Christophe Fouquet, I'm the CEO of ISML,
00:24and recently, you know, ISML, that used to be a pretty unknown company, got a bit more on the news,
00:33and one of the reasons for that is because ISML is today the sole supplier of EUV lithography.
00:43And EUV lithography is basically the tool that is being used to print the most advanced chips.
00:52So, of course, at VivaTech, I'm sure you talk a lot about AI, and I'm going to do that as
00:58well.
00:59But AI requires chips, and chips require EUV technology.
01:06So this is why the Wall Street Journal, back in December 2024,
01:11called EUV the most indispensable machine in the world.
01:14So I'm going to try to educate all of you a bit about that.
01:20So we'll go back to the wafer, we'll go back to EUV lithography.
01:23Before I do that, I'm going to most probably echo some of the messages you got from many speakers at
01:31the conference.
01:33AI is going to be everywhere.
01:36We used to talk about semiconductor everywhere in ISML, we start to talk about AI everywhere.
01:43In fact, we believe that AI is most probably the next industrial revolution.
01:51AI is going to enable any serious, any new revolutionary application anyone will invent in the years to come.
02:00And I think that in the next few years, any country, any company that does not embrace AI fully
02:11take a huge risk to fall behind for quite a long time when it comes to innovation and competitiveness.
02:20So AI is going to be everywhere, and as a result of that, we see a huge investment.
02:29We're talking about trillion of dollars being invested in the AI infrastructure in the next two, three years.
02:38By many companies, most of them in the United States, it's about 80% of the investment.
02:45China, of course, China, of course, Europe, a bit.
02:48And this is basically the first round of investment to build the infrastructure so that AI basically
02:57can be used, distributed to everyone and every single industry.
03:04What does it mean?
03:06It means that if we look at the overall ecosystem, the overall supply chain for AI,
03:15data center, accelerators, wifers, you're just one of them, the need for this is going to more than double
03:26in the next two, three years.
03:30And this puts a huge amount of challenges, I will say a huge amount of stress on the supply chain.
03:39And in fact, we believe that for the next few years, we're going to experience a market that will be
03:47limited by supply.
03:49And when you hear all the actors talking about AI, NVIDIA, who, of course, make the hardware for the hyperscaler,
03:58TSMC, who make the wafer for NVIDIA, all the equipment supplier, everyone is talking about a very huge backlog
04:06and the need basically to really accelerate capacity in order to make sure that we can basically deliver on that.
04:13So this is already quite spectacular, but as you know, some people have planned to even make that bigger.
04:24The TerraFab is, of course, a good example, but that's not the only example of huge infrastructure being built
04:32to, I would say, create a real boost in capacity basically to be able to support that.
04:40This is very important because, you know, you'll talk about AI, you'll talk about model, and there's a lot of
04:45focus on that.
04:46But if you don't get the entire hardware, the entire supply chain for that, then at some point, this opportunity
04:55will slow down.
04:59So, I'm going to talk now a bit more about our own technology.
05:03This gentleman here is Dr. Gordon Moore, is one of the founders of Intel, and he told the entire world
05:13that when it comes to chips, the number of transistors per chip had to double every two years.
05:22Transistors are important on chips because they define the computing power of a chip,
05:28so hyperscaler needs a lot, a lot of transistors, I'll come back to that, or define how much information you
05:34can store.
05:35This is basically for memory.
05:38So this is something that has been with us for many, many years, but more recently, Jensen and Wong said,
05:45well, Moore's law is dead.
05:48And the reason why you say that, it's because as we enter the AI age, the two times more transistors
05:58every two years,
06:00which is already a huge challenge for the entire industry, is not good enough.
06:05In fact, what's happening in the last few years, is that on top of continuing to drive Moore's law,
06:14to increase the number of transistors, to also reduce the cost of the transistor so that the new applications are
06:22still affordable,
06:23the emergence of first accelerated compute and then AI system and factories have increased the need for that.
06:36So we're not talking about two times every two years increase in transistor.
06:40We're talking about ten times every two years.
06:44So the amount of chips that has to be made in order to fill that is enormous.
06:52And what's happening then, so for many, many years, a wafer could just provide the more transistors you needed.
07:01Today, that's not enough, because people cannot scale anymore enough transistors on one wafer to provide a whole capacity.
07:09So what the ten times every two years requirement is making is basically a need for a lot more wafer
07:16capacity.
07:17So if you cannot put more information on the wafer, you're going to end up exposing, printing a lot more
07:25wafer.
07:26So a good example of that is, of course, NVIDIA.
07:29So, you know, in the past, when you had a PC, when you had a phone, your phone or your
07:35PC would have had one CPU, right?
07:38One chip that would do the job.
07:40If you look at NVIDIA product today, Blackwell is already built out of 50 wafers.
07:47So you're talking basically about today already 130 trillion transistors, which cannot be fit in one wafer.
07:57Now, if you look at their product planned for 2027, the Rubin Ultra, there will be an increase of 10
08:04in two years in the number of transistors needed,
08:09which will translate in an increase of a factor of five into the number of wafer needed to build that
08:16product.
08:16So you're looking at 250 wafers for one single product.
08:19So you see basically this explosion, if you want, again, of volume of demand on the wafer.
08:29So that's a bit the big picture.
08:32And since I talk about NVIDIA, I'm going now to bring you a bit inside what's happening.
08:38So this is, of course, a data center.
08:41You may have seen that already.
08:42And if you go inside the data center, you open one of the rack, you're going to see different electronic
08:48boards.
08:48And for those electronic boards, you're going to see chips.
08:52And those chips are being exposed on the wafer you just saw, right?
08:57So this is integrated electronics.
08:59This is all built up out of a silicon stack using many, many different masks.
09:05So this is the wafer LA was showing you before.
09:08And that wafer itself, if we zoom in a bit, we are going to see basically a very complex structure,
09:16which is the chips.
09:17At the bottom of the structure, you're going to find a transistor.
09:21And on top of the transistor, you're going to see basically a lot of what we call interconnect metal layer
09:28that basically build up the circuit.
09:31Now, if we really go to the bottom of the transistor, I was talking about size before.
09:35Therefore, today, the latest and greatest is being built out of 8 nanometer line.
09:44Well, no one knows exactly what is 8 nanometer.
09:47So to give you an idea, if you take one of your hair, it's 80,000 nanometer.
09:53So meaning if you cut your hair in 10,000 pieces, you're going to end up with a dimension of
10:00one of those lines.
10:02So this is a kind of stuff our machine are going to expose.
10:08And this is a kind of precision the chips you're going to be using without knowing it, maybe with AI,
10:16are going to use.
10:18How do we do that?
10:19So now I have to zoom out a bit more.
10:22And you are going to end up basically to the machine, the very important machine.
10:29This is the ASML EUV scanner that is printing the information on the wafer.
10:35We do that at very high speed, 20 seconds per wafer, very high precision.
10:43How do we do that?
10:44It's all about size.
10:47We need to be able to print the smallest possible size.
10:51The size, the resolution of a feature is defined by the wavelength of a lithograph imaging divided by the optical
11:01numerical aperture.
11:03EUV is about the wavelength.
11:05And NA is another parameter we can grow.
11:07So over time, basically, what ASML has done is develop different machines, different tools that could print smaller lines.
11:18We started with about 400 nanometer lines.
11:21And today, again, we are down to 8 nanometers using many, many innovations, which I'm going to describe later on.
11:29Before I do that, I'll tell you a bit about the ASML story.
11:33It started in 1984 in the Netherlands, in Eindhoven, spin-off from Philips.
11:38So in case you don't know where are the Netherlands, we'll help you a bit here.
11:41And this is the old campus of Philips.
11:46And like every great technology company history, we started basically a bit more than 40 years ago in this very
11:56nice facility with 31 employees.
12:01Things have changed, of course, but today we are still located, basically, a few kilometers away from that campus.
12:07This is the current facility of ASML.
12:12And this is the place, basically, where all the machines, all the EUV machines are being built today.
12:18You're going to see a few pictures.
12:19So this is our clean room.
12:21These are all people working on the machine.
12:22So you see already a bit of a glimpse of the complexity of those tools.
12:30It takes us about four to six months to build one machine in Veldhoven and then, basically, ship it.
12:38This one, INA, is the latest version of it.
12:46And here, again, you have an idea of the complexity.
12:49So there's a lot of physics.
12:51I'll come back to that.
12:53There's a lot of high-precision mechanics.
12:57And there is a lot, basically, of software in order to be able to run this machine at very high
13:04speed, very high accuracy.
13:06And I will say up time, which is more than 95%, because this is a machine that is basically built
13:17for the industry.
13:18So these are some pictures, a bit of what we do.
13:21So, again, you see here a wafer being exposed with the chips coming out of the machine.
13:26And this is the latest.
13:28Now, to give you an idea also of the complexity of the operation, I'm going to show you how we
13:34brought this very first machine, now, a couple of years ago, to Intel in the US.
13:39So this is one of the seven plans we need to carry the tool from our facility to our customer.
13:48Spatial containers, very sensitive technology.
13:50And once the machine has been assembled, tested in ASML, it's going to be disassembled, transported, and reassembled at our
14:00customer.
14:00And this whole, basically, cycle takes about a year between the time we start building a machine and the time
14:06it's going to be released to our customer.
14:08So this gives you a bit of an idea of, again, the complexity.
14:12You see, the machine is being built up.
14:14Well, you know, you have all seen a London bus, I guess.
14:18I think the INA machine is slightly bigger than that.
14:21And if you look at all the facility needed to support it, you can add another three or four London
14:28buses.
14:29So this is, of course, significant.
14:32And this is the resource of many, many, many years of work, of innovation at ASML to bring, basically, this
14:40technology together.
14:41So now I'm going to talk a bit about some of this technology.
14:46The first thing, EUV, EUV light.
14:50It doesn't exist.
14:52There is no EUV light around us.
14:55You can find EUV light on the sun.
14:59And the reason why the sun has EUV light, naturally, is because of the very high temperature and the plasma
15:07on the sun.
15:08So since we couldn't have it, we had to artificially create that light, create the extreme environment.
15:16And this is done, basically, as you will see here, to create EUV light.
15:22And we are using a very, very small droplet of tin, which we are going to hit with tens of
15:32kilowatt powerful laser.
15:35And this creates a plasma, 220,000 degrees Celsius.
15:40So this is, basically, the temperature of the environment.
15:42And in order to get enough light, well, you could say it's pretty hard to hit a droplet.
15:48You have to realize the droplet is a few micron big.
15:51But we have to do that today 60,000 times a second.
15:56And we have to hit every single time the droplet in the right way.
16:02Now, if I really wanted to show off, I would tell you, in fact, that we don't hit it once,
16:08we hit it three times.
16:10So we have three different lasers, two lasers to prepare the droplet, make it bigger, and the third one, basically,
16:16to get the light.
16:18This is about the power.
16:21This is about the source light.
16:22Now, the mirrors are also very complex.
16:25We talk about 8 nanometer.
16:27Mirrors are being used to bring the light to the wafer.
16:30And in order to do that, the accuracy of those mirrors has to be 20 picometer.
16:37Just so the example, basically, the most accurate mirror before that was Hubble telescope, which is known by all of
16:43you, a factor of 1,000 on the accuracy.
16:46What you see here also is that we are so accurate that we could, from Earth, with those mirrors, project
16:55the size of a coin to the moon.
16:58That's the type of accuracy we create on our mirrors.
17:02This is, of course, all designed from scratch.
17:05Again, requires a lot of innovation and is very unique in what we do in ASML.
17:13A few more examples.
17:15Speed is important.
17:16We want to be able to be economically, I would say, viable for our customer.
17:21So we need to move wafer very, very fast.
17:24In order to do that, what you see on the left is basically the reticle.
17:28This is the part that contains the information you want to print on the chip.
17:33This accelerator at 22G, Formula 1, which is very impressive, is only 5G.
17:40In fact, if any of you were to go and sit on one of our reticle stage, you would be
17:45killed because you would not be able to sustain the acceleration.
17:48So 22G, and I should say, of course, that this happens at nanometer precision.
17:56So not only do we accelerate at 22G, but we are able to move at a nanometer precision, which I
18:02think the Formula 1 is still pretty far from being able to achieve.
18:09Finally, comparing again a bit the resolution, 8 nanometer, I talked about the hair.
18:15This is another example.
18:17This is a microvirus, 18 to 28 nanometer.
18:21So the stuff, again, that is printed on the wafer, as you have seen before, is smaller than some of
18:27the smallest virus known to humans.
18:30So this is a bit what we do.
18:32And you also understand why this machine became so important, because you understand that all this technology put together, 40
18:43years of experience, 40 years basically of learning at ISML,
18:48but also together with our partner being customer, supplier, is very difficult for anyone to do if you start from
18:57scratch.
18:59Therefore, the reliance on almost, I would say, or most probably today the entire world on that technology to create
19:07those AI chips.
19:10So since I talk about AI, you know, AI is also, of course, good for ISML.
19:16I told you before that if companies don't embrace AI, they may fall back for a long time, because they
19:28will lose their competitiveness.
19:29So this is true for ISML.
19:31There is a lot of discussion about models.
19:35We believe that the true value of AI is data, because if you don't have data, you know, there's no
19:42need for any powerful models.
19:44And the machine I just showed you is creating a huge amount of data.
19:49We're talking about, at our customer, 15 terabytes per hour of data creating in any fab.
19:57Altogether, we are looking at a pool of data at ISML of about 120 petabytes.
20:05You can go and look what a petabyte is.
20:07It's a lot of zeros.
20:08So we have this information, and, of course, this information can be used to improve our system, to improve our
20:16design, to improve our processes, to accelerate R&D.
20:20Again, I'm sure you have heard a lot of that in the last few presentations.
20:25And also to help, basically, our customer with their own process.
20:30So this is also the reason why we created this partnership with Mistral.
20:37We wanted to bring a company that could work with us very deep in what we are doing, have their
20:46people with our people.
20:48And Mistral offered that.
20:50This allowed us to strengthen our core competence, as I just explained.
20:55This will allow us to support connected markets, and, over time, also potentially allow us to explore more opportunities.
21:06So, this is the end of my talk.
21:08I hope that, by now, a few more of you know what is a wafer.
21:14A few more of you understand why this entire supply chain is so important.
21:19And maybe why, also, the EUV machine is, indeed, the most indispensable machine in the world.
21:25Thank you very much.
21:26Thank you very much.
Comments