LongCut logo

How ASML Makes Chips Faster With Its New $400 Million High NA Machine

By CNBC

Summary

Topics Covered

  • A single Dutch company controls the future of AI
  • EUV took 20 years and nearly failed
  • Monopoly pricing is constrained by Moore's Law
  • AI could collapse under its own energy demand
  • China is scaling 7nm chips without EUV

Full Transcript

All right. We're about to enter the High NA lab.

Behind this highly secured door in an ASML cleanroom.

There's a giant $400 million machine.

It's bigger than a double-decker bus.

Made with such closely protected technical specificity.

Very complex. The most advanced machine tool in history.

That no filming's ever been allowed, even by ASML's own team.

Until we flew across the world to see it.

We're here at ASML's headquarters in the Netherlands, where 100% of the world's extreme ultraviolet lithography machines are made, and now the next generation machines called High NA.

We had to wear these bunny suits to protect these extremely precise machines from getting contaminated by our own hair and skin cells.

This is high numerical aperture.

High NA, the new generation of EUV machines, the only machines in the world capable of printing nanoscopic blueprints onto the most advanced chips.

Nvidia, Apple, Taiwan Semiconductor, Samsung Intel.

None of their advanced chips can be made without EUV.

And this company has that market completely cornered.

The first High NA machine was installed in the U.S.

at Intel's Oregon chip fabrication plant, or Fab, last year.

Only five have ever been shipped.

Indeed, the machine is so big you can't ship them in ones, so we break them into parts.

Now they're being ramped up to make millions of chips on the factory floors of the few companies that can afford them Intel, TSMC and Samsung.

ASML says eventually China will be used by all its EUV customers. Other advanced chipmakers like Micron,

customers. Other advanced chipmakers like Micron, SK Hynix and Rapidus.

Without EUV, without very advanced logic manufacturing.

Most of the AI cannot happen.

CNB went from ASML's labs in California all the way to its Veldhoven, Netherlands, headquarters to bring you this never-before-seen look at the world's most advanced chipmaking machine and to ask its makers exactly how it works and what could go wrong, amidst Trump's tariffs, export controls and a new training facility coming to the

U.S.

U.S.

Suited up and throuhj an air shower.

It's basically to blow off all the contamination from your suit and your clothes.

Down several long hallways where humongous parts are being meticulously moved.

How many people would you say are working here? 24 /7

.

Inside the EUV factory alone, it's like 2,000 people.

It's a 24 /7 operation.

Ayssia Haddou took us to the restricted lab space where 40 years of ASML's technology research all comes together.

It takes approximately a year and a half to create this system.

Haddou leads the team that tests and qualifies the High NA machines here.

It still does blow my mind.

I think even if you learn more about it, it blows your mind even more.

The big machine at the end of the day is used to transfer a pattern, which we store on a mask onto a wafer, and then the end of the day, you cut them in pieces, you package them, and then you end up in your phone.

Jos Benschop joined Dutch electronics giant Philips in 1984. Two weeks later, ASML was founded as a

1984. Two weeks later, ASML was founded as a subsidiary, conducting research out of a leaky shed nearby in Eindhoven.

It broke off from Philips with an IPO in 1995, and two years later, Benschop helped spearhead ASML's first foray into EUV.

Three or four years later, we had enough confidence to put all our eggs in one basket called EUV.

With hindsight, we can say we placed the right bet.

On a yearly basis, I think we file several hundreds of patents on EUV.

Christophe Fouquet began as CEO just over a year ago, but he's been with ASML more than 17 years.

We barely made it.

I think sometimes people forget that.

But EUV took us practically more than 20 years.

I would say it's been a very risky investment because when we started, well, there was no guarantee the technology would work.

By 2018, ASML had shown that EUV was possible, and it's the same process used in High NA.

Here's how it works.

Chips are built in a grid formation on thinly sliced wafers of silicon, each with up to 100 thin layers that make up billions of transistors that determine what the chips can do.

These layers are printed using lithography, extremely precise rays of light projected through a mask of the chip design until they hit the surface of the wafer that's been coated with photoresist chemicals. ASML's older generation DUV machines,

chemicals. ASML's older generation DUV machines, which it still makes, use rays of deep ultraviolet light with a wavelength of 193 nanometers.

Nikon and Canon in Japan are still competitors in DUV, but ASML's real breakthrough came with that bold bet on a technology many thought was impossible using even shorter wavelength extreme ultraviolet light, EUV, something that to this day no other company does.

A few more people tried for many years, but the size of the investment, the commitment you have to have, I think, discouraged quite a few people.

Transistors have gotten so small, down to 10,000 times thinner than a human hair, that ASML decided to create its own source of smaller, more precise EUV light.

At a wavelength of 13.5 nanometers, it's the width of just five DNA strands laid side by side.

So small that it gets absorbed by almost everything. And that's why this whole system needs to

everything. And that's why this whole system needs to be in a vacuum.

This tiny thread may look like the strand of a spider web, but it's actually molten tin being shot out at 50,000 droplets a second.

The tin droplets are vaporized with a laser, and those tiny explosions are what emit photons of EUV light.

This is called a CO2 laser.

It's been used in the car industry for many years.

They would cut through steel with one of these.

We use four of them to shoot at one of these tiny tin droplets to generate a plasma, which is hotter than the sun.

That plasma generates EUV, and that EUV is captured by this collector mirror that comes from Carl Zeiss and brings the light into our scanner.

The light bounces off of mirrors that aim it through the lens, much like how a camera works.

But because EUV gets absorbed by typical mirrors.

German optics company Zeiss made specific mirrors just for ASML that are the flattest man made surfaces in the world.

And then the light is emitted through the projection optics onto the wafer, which is about here.

So this is the wafer table.

When the EUV finally hits photoresist chemicals on the surface of the silicon wafer, it prints the minuscule blueprint that makes up the chip. The aim needs to be so precise,

chip. The aim needs to be so precise, TSMC says it's equivalent to shining a laser from the moon to hit a coin on the Earth.

EUV made up less than 8% of ASML's total number of machines sold in 2024: 44 EUV machines versus 374 cheaper DUV machines, but with a price tag of $220 million. EUV made up 38% of ASML's system sales revenue.

million. EUV made up 38% of ASML's system sales revenue.

But a cost of the chips the tool is going to create is going down.

Because yo u're t he o nly c ompany in the world that makes these m achines, w hat's stopping you from setting the price even higher?

Well, what stops us is Moore's Law says that we need to continue to drive costs down.

This allows you to go to the next node and then the next node, and then again the next node.

If this was to stop, then everyone gets stuck.

ASML began developing its latest $400 million High NA machines around 2016.

Inside, all the same processes still happen with the same EUV light source.

But there's one key difference.

We increase the aperture of the opening angle of the lens.

High NA stands for high numerical aperture, meaning it has a larger lens opening, increasing the angle at which the light is captured by the mirrors. More light coming in from steeper angles allows High NA machines to capture increasingly small designs on chip masks, so they can be projected in one exposure step, whereas original lower NA machines reach higher resolution to project these tiny designs using multiple exposures through multiple masks.

One become two, then two become three, three become four.

And when the number increases, it gets very complex process wise.

And the yield goes down.

But with a higher NA, resolution improves and the small designs can be printed with fewer masks, saving time and money.

The larger the number, the smaller the thing you can print.

But this higher NA also comes with a much higher price tag for one main reason.

The higher the NA of an optic, the bigger the mirror you have to use, and therefore the bigger the system.

These machines also take up a huge amount of power.

How much power do these machines take?

I think they take too much.

If we don't improve the power efficiency of our AI chips over time, the training of the models could consume the entire worldwide energy, and that could happen around 2035.

Since the moment we started EUV, 2018, we have reduced the power the energy needed per wafer exposure by more than 60%.

ASML says its machines print about 200 wafers per hour, 24/7, 365. The goal is several hundred per hour.

24/7, 365. The goal is several hundred per hour.

You can print at the end of the day, millions of these wafers per year.

At a conference in February, Intel said it started making chips with High NA using it for some 30,000 wafers so far, and that it's about twice as reliable as previous models.

And Samsung said High NA could reduce its cycle time by 60%.

High NA means two things.

First and foremost, shrink.

So there's more devices on a single wafer.

And secondly, by avoiding multiple patterning, you can make them faster and you can make them with higher yield.

In reality, every chip is made using multiple different lithography techniques at different resolutions.

Every chip consists out of like 100 different layers.

Most of the layers are actually made with deep UV, which is a much smaller system.

Legacy DUV machines sell for less 5$ million to $90 million, but still make up over half of ASML's sales.

The older machines are in high demand.

Many of the chips we use in our daily life do not require DEUV. If you press the button to open the

require DEUV. If you press the button to open the window in your car, unlikely an EUV-based chip is needed.

Turns out ASML sells a lot of its legacy machines to China, although U.S.

export controls prevent it from selling EUV there.

It's a ban that started under the first Trump administration.

The Trump administration is very focused on ensuring that China does not get access to the most advanced tools and chips that are required to power AI.

So without access to ASML's EUV machines, could China develop its own?

I feel like it's a real long shot that China could enter this space. But right now, China's really figuring out

this space. But right now, China's really figuring out ways to take things like seven nanometer, the most advanced nodes that can be done without EUV, and figuring out ways to scale them.

And they've been able to do that successfully in some areas, like in smartphones and personal devices.

Still, export controls haven't stopped China from stockpiling ASML's less advanced machines.

At one point recently, u p to 49% of your business was in China. What happened there and what is that number at now?

We had a huge backlog in China because we got orders in 2022 at the peak of the market.

And in 2023, 24, the rest of the world kind of went down. So basically we could use our capacity to fulfill this backlog.

End of 24, last quarter, the part of China in our business went down to 30%.

U.S. concern over advanced tech making its way to China has accelerated amidst the generative AI race.

But that boom has also sent demand for chips and the machines that make them soaring.

And with it, chip stocks including ASML's, which hit an all time high in July.

Still, it's declined some since as the chip industry faces uncertainties, chief among them Trump's tariffs.

Any idea how tariffs are going to impact ASML?

Well the short answer is I don't know.

We we don't know.

And if anybody would know it could make a lot of money.

But we don't.

In its first quarter earnings reported last month. ASML missed on order expectations amid tariff

month. ASML missed on order expectations amid tariff uncertainty.

I suspect ASML will come in and advocate for a very low to minimal to no tariff on their machines.

The impact of tariffs is complicated by ASML's complex web of 800 global suppliers and the winding worldwide path. ASML's machines take before

worldwide path. ASML's machines take before arriving in a fab.

Each High NA machine is made up of four subsystems manufactured in Connecticut, Germany, the Netherlands and California.

Each module is shipped separately to ASML's headquarters in Veldhoven, where they are put together, tested, then disassembled again, bound for fabs in places like the U.S.

or more frequently, Asia.

How do you ship something this massive?

Yeah, it takes about seven Boeing 747s to ship this whole thing, and it's about 25 to 30 trucks.

For years, Asia has made up more than 80% of ASML's business, but the U.S.

share, around 17% in 2024, is growing fast, especially as Intel builds new chip fabs in Ohio and Arizona. And TSMC is now in volume production in

Arizona. And TSMC is now in volume production in Arizona too.

So there's a lot of chip customers in the U.S.

So there is a, I would say, a good reason to have cheap manufacturing.

When we first got a look at EUV in 2022, ASML had about 32,000 employees.

Today, that number has grown to 44,000, and 8,500 of them are in the US.

Like Eric Ma, who's been with ASML in San Jose, California, for 18 plus years.

Have you gotten to see the High NA machines?

Yes. What is it like?

Oh, it's huge.

Of ASML's 60 global offices, 18 of them are in the U.S., including a manufacturing site in Connecticut and R&D sites in San Diego and here in San Jose.

We're here at ASML, Silicon Valley, California location.

And behind me is a metrology machine.

This one has 25 beams of electrons that are used to test the different microchips, to make sure that the yield is going to be good.

For High NA in the U.S., it's likely that Arizona could be a major hub. Many of ASML's first High NA shipments went to Intel last year, and the struggling U.S.-based chipmaker is building a giant new chip fab in Chandler, Arizona.

Intel has been for many years a formidable partner for ASML, they were the first company to believe in EUV and the first company to work with us to develop EUV.

They have done the same on High NA, and I think that moving forward, the role of Intel for the U.S.

autonomy, independence on semiconductor is still very critical.

But TSMC is far ahead of Intel in advancing chip nodes. We recently visited TSMC's new fab north of

nodes. We recently visited TSMC's new fab north of Phoenix, the most advanced chip fab on U.S.

soil. The need for High NA there will likely come soon, with ASML's original EUV machines in use now.

In order to hook one up requires something over 2,000 electrical connections, over two kilometers of cabling, 100,000 parts, 40,000 bolts.

We're talking about massive.

With so many EUV machines in Arizona, ASML is for the first time bringing a training facility to the US. Its only other training centers are in the Netherlands, Taiwan and Korea.

We're planning to open that in the next few months in Arizona, with the idea that we could train 1,200 people every year on EUV dpof.

As chipmaking shows no signs of slowing in the U.S.

or elsewhere, ASML is already hard at work on its next generation machine, Hyper NA.

With an even larger NA that can print even smaller features.

I heard it could be up to $700 million.

Do you have any price tag?

Oh, it's way too early to mention the number, but typically price goes up.

So no, I don't think it's going to be that high.

Despite market headwinds, Fouquet told CNBC that customers will need Hyper NA within ten years.

Anything you can tell us about Hyper NA?

How far away are we from seeing that?

For us, it's not necessarily a difficult product.

We need our customers to need it as well.

And if you look at the roadmap of our customer, I think the need for Hyper NA will be somewhere between 2032 and 2035.

But for now, ASML is focused on meeting demand for High NA. It plans to ship at least five more systems this

NA. It plans to ship at least five more systems this year, ramping to a production capacity of 20 machines in a few years.

While this is a monopoly right here and today, in the era of AI, there will be new vehicles for building semiconductors.

And if I'm ASML, if I'm from any other company in this space, I'm being careful to make sure that I'm protecting my position, not just in the near term, but in the long term as well.

Loading...

Loading video analysis...