THE ECONOMIST: Race on to build complex machines producing AI chips in industry already dominated by ASML

Few would expect the future of artificial intelligence to depend on Eindhoven, a quiet Dutch town.

Yet just beyond its borders sits the headquarters of ASML, the only company that makes the machines, known as lithography tools, needed to produce cutting-edge AI chips.

ASML’s latest creation is a 150-tonne colossus, around the size of two shipping containers and priced at around $US350m ($553m). It is also the most advanced machine for sale.

The firm’s expertise has placed it at the centre of a global technology battle.

To prevent China from building whizzy AI chips, America has barred ASML from selling its most advanced gear to Chinese chipmakers. In response, China is pouring billions of dollars into building homegrown alternatives.

Meanwhile, Canon, a Japanese rival, is betting on a simpler, cheaper technology to loosen ASML’s grip.

Yet unlike software, where industry leadership can shift in a matter of months, success in lithography is a slow-moving race measured in decades.

Overtaking ASML won’t be easy. At stake is control of the machine that will shape the future of computing, AI and technology itself.

ASML’s most advanced machine is mind-boggling. It works by firing 50,000 droplets of molten tin into a vacuum chamber. Each droplet takes a double hit — first from a weak laser pulse that flattens it into a tiny pancake, then from a powerful laser that vaporises it.

The process turns each droplet into hot plasma, reaching nearly 220,000°C, roughly 40 times hotter than the surface of the Sun, and emits light of extremely short wavelength (extreme ultraviolet, or EUV).

This light is then reflected by a series of mirrors so smooth that imperfections are measured in trillionths of a metre.

The mirrors focus the light onto a mask or template that contains blueprints of the chip’s circuits. Finally the rays bounce from the mask onto a silicon wafer coated with light-sensitive chemicals, imprinting the design onto the chip.

High stakes

ASML’s tools are indispensable to modern chipmaking. Firms like TSMC, Samsung and Intel rely on them to produce cutting-edge processors, from AI accelerators to smartphone chips.

No other company makes machines that can reliably print chips that are called “7 nanometres” (billionths of a metre) and below (though these terms once related to physical resolution, they are now primarily used for marketing).

Even for more mature technologies (“14nm” and higher), the firm’s tools account for over 90 per cent of the market.

A microchip is an electronic lasagne: a base of transistors topped with layers of copper wiring shuttling data and power.

A leading-edge processor can pack over 100 billion transistors, contain more than 70 layers and have more than 100 kilometres of wiring, all on a piece of silicon around one-and-half times the size of a standard postage stamp.

To build these tiny features, a lithography machine works in stages by etching patterns of transistors and metal wires on a wafer, layer by layer. A single wafer can contain hundreds of chips.

ASML’s tool is complex, yet its basic principle is much like that of an old slide projector: light passes through a stencil to project an image onto a surface.

The smallest feature an optical lithography tool can print depends mainly on two factors. The first is the wavelength of light. Just as a finer paintbrush allows for more detailed strokes, shorter wavelengths enable smaller patterns.

ASML’s older systems used deep ultraviolet (DUV) light, with wavelengths between 248nm and 193nm, producing features as small as 38nm.

To shrink chip features even more, ASML turned to EUV light, with a wavelength of 13.5nm. Whereas EUV is naturally emitted in space by the solar corona, producing it on Earth is far trickier.

EUV light is also completely absorbed by air, glass and most materials, so the process must be enclosed in a vacuum, using special mirrors to reflect and guide the light. ASML spent two decades perfecting the method that fires lasers at molten-tin droplets to create and generate this elusive beam.

The other dial that sets the smallest feature size is the numerical aperture (NA) of the mirrors, a measure of how much light they can collect and focus.

ASML’s latest systems, called high-NA EUV, use mirrors with an aperture of 0.55, allowing it to print features on chips as small as 8nm.

To go smaller still, the firm is studying what it calls hyper-NA by cranking the aperture up to more than 0.75 while still using existing EUV light. A higher NA means that the mirrors collect and focus light coming in from a broader range of angles, improving precision.

This comes at a cost. Larger NAs require bigger mirrors to intercept and direct the expanded light paths. When ASML increased the NA of their machines from 0.33 to 0.55, the mirrors doubled in size and became ten times heavier, now weighing several hundred kilograms.

Increasing the NA again will only add bulk, raising concerns about power consumption.

Another obstacle is pricing. ASML does not disclose precise figures, but its latest EUV machine was almost twice as expensive as its predecessor. A hyper-NA system would be dearer still.

Though the company cautions that there are no guarantees of it ever being produced, Jos Benschop, ASML’s head of technology, believes a hyper-NA machine could arrive within the next five to ten years, pending demand.

Some researchers are already planning to go beyond EUV light, aiming for wavelengths of around 6nm. This would require breakthroughs in light sources, optics and photoresist (the light-sensitive coating on wafers).

Shorter wavelengths also bring new challenges, including “shot noise”, or random particle movements that blur patterns. But Yasin Ekinci of the Paul Scherrer Institute, a Swiss research centre, sees this as a “plan B” if hyper-NA fails to deliver.

While ASML pushes the boundaries of optical lithography, China — cut off from the most advanced chipmaking tools — is trying to extract more from the older ASML machines (capable of 28nm and above) it can still import.

One approach is multi-patterning, in which a pattern is broken into multiple etching stages, allowing a machine to print details twice or four times as small. Multi-patterning is effective, but adds complexity and slows production.

China is also trying to build its own lithography tools. SMEE, a state-owned firm, is reportedly making progress on a machine capable of producing 28nm chips using DUV light.

But developing an EUV system is an entirely different challenge. Jeff Koch of SemiAnalysis, a research firm, points out that beyond mastering EUV light itself, China would need to replicate ASML’s vast supply chain, stretching to more than 5,000 specialised suppliers.

ASML’s dominance in high-end lithography, therefore, seems unshakeable. But Canon, once an industry leader, is betting on an alternative.

Nanoimprint lithography (NIL) stamps circuit patterns directly onto wafers, much like a printing press. In theory, NIL could create features with nanometre accuracy, offering a low-cost, compact rival to ASML’s EUV machines.

The NIL process begins with the creation of a master mask which has the template of the circuit etched onto it by an electron beam. Next, droplets of a liquid resin are applied to the wafer before a mask presses the circuit pattern onto the wafer.

Ultraviolet light is then used to solidify the resin and form the circuit patterns, after which the mask is removed.

This step is repeated for every layer of the chip. Canon estimates that its approach costs around 40 per cent less than a comparable machine from ASML.

For NIL to become a mainstream chipmaking technology, it must overcome several challenges. Defects are a big concern — tiny particles or imperfections on the mould can create repeating flaws across entire wafers.

Alignment is another hurdle. Since chips are built in layers, the circuit patterns of every layer must line up precisely.

Any variation in wafer flatness or slight misalignment between the mould and wafer can cause nanoscale errors, disrupting electrical connections. Canon claims its system achieves nanometre precision, but maintaining this consistently during production is difficult.

Then there is throughput, or how many wafers a machine can process per hour.

ASML’s high-NA EUV tools can handle over 180 wafers per hour, with some older models reaching nearly twice that. In contrast, Canon’s latest NIL system manages only 110 wafers per hour, making it less suited for high-volume chip production — at least for now.

So far NIL has found more success outside semiconductor manufacturing, particularly in making smartphone displays and other high-precision components.

The technology is now making inroads into memory-chip production, where higher defect rates are more tolerable than in logic chips. Iwamoto Kazunori, the head of Canon’s optical division, believes that NIL can co-exist with EUV lithography, cheaply performing manufacturing steps where it can and steering clear of finer detail.

Such innovation could help firms design faster and more energy-efficient chips capable of powering a new generation of AI models. If ASML is not careful, the world’s most important machine may not keep its title for ever.