Today we're diving into something absolutely critical to the future of artificial intelligence that you might never have thought about before: the machines that make AI chips possible.
When we talk about AI advancements, we often focus on models, algorithms, and software breakthroughs. But there's a fascinating hardware story unfolding that will determine the trajectory of AI for decades to come. At the center of this story is a Dutch company called ASML, headquartered just outside the quiet town of Eindhoven in the Netherlands. (ASML once stood for “Advanced Semiconductor Materials Lithography” but they now consider it to not stand for anything.)
ASML makes lithography machines – the incredibly complex tools that print microscopic circuit patterns onto semiconductor wafers. Their latest creation is simply mind-boggling: a 150-tonne colossus, roughly the size of two shipping containers, with a price tag of around $350 million. And here's the kicker – ASML is the only company in the world capable of making the most advanced versions of these machines that produce cutting-edge AI chips.
The technology behind these machines is almost beyond comprehension. ASML's extreme ultraviolet (or EUV) lithography machine fires 50,000 droplets of molten tin into a vacuum chamber. Each droplet gets hit twice by lasers – first to flatten it into a tiny pancake, then to vaporize it completely. This creates a plasma that reaches temperatures of nearly 220,000 degrees Celsius – that's about 40 times hotter than the surface of the sun! This process generates extreme ultraviolet light with incredibly short wavelengths, which is then reflected by a series of mirrors so smooth that their imperfections are measured in trillionths of a meter. The light is focused onto a template containing the chip's circuit blueprints and finally projected onto a silicon wafer, imprinting the design. (If you’d like to hear more about silicon wafers, check out Episode #875 of this podcast.)
This technological marvel is what allows companies like TSMC, Samsung, and Intel to produce the cutting-edge processors that power everything from AI accelerators to smartphones. No other company makes machines that can reliably print chips with the smallest features possible today, i.e., smaller than 7 nanometers. Even for more mature technologies, ASML's tools dominate over 90% of the market.
To put this in perspective, a modern microchip is like an electronic lasagna – a base layer of transistors topped with layers of copper wiring that shuttle data and power. A leading-edge processor can pack over 100 billion transistors, contain more than 70 layers, and have more than 100 kilometers of wiring – all on a piece of silicon about one-and-a-half times the size of a standard postage stamp. A single silicon wafer processed by these machines can contain hundreds of individual chips.
The complexity of this technology has placed ASML at the center of a global technology battle. To prevent China from building advanced AI chips, the United States has barred ASML from selling its most advanced equipment to Chinese chipmakers. In response, China is investing billions to develop homegrown alternatives. Meanwhile, Canon, a Japanese competitor, is betting on a different, potentially cheaper technology called nanoimprint lithography to challenge ASML's dominance.
But here's the key insight: unlike software, where industry leadership can shift in a matter of months, success in lithography is measured in decades. ASML spent two decades perfecting its method of producing EUV light. Replicating this achievement is not something that happens quickly, regardless of how much money you throw at the problem.
ASML isn't standing still either. Their latest systems, called high-numerical-aperture EUV, use mirrors with an aperture of 0.55, allowing them to print features as small as 8 nanometers. To go even smaller, they're working on what they call hyper-numerical-aperture, which would crank the aperture up to more than 0.75. This comes with significant engineering challenges – when ASML increased the numerical aperture from 0.33 to 0.55, the mirrors doubled in size and became ten times heavier, now weighing several hundred kilograms. Increasing it again will only add more bulk and power consumption concerns.
Some researchers are already planning to go beyond EUV light, aiming for wavelengths of around 6 nanometers. This would require breakthroughs in light sources, optics, and the light-sensitive coating on wafers. But many see this as a "plan B" if the hyper-numerical-aperture approach fails to deliver.
China, cut off from the most advanced tools, is trying to extract more from the older ASML machines it can still import. One approach is multi-patterning, which breaks a pattern into multiple etching stages, allowing a machine to print details twice or four times as small. While effective, this adds complexity and slows production. China is also trying to build its own lithography tools, with a state-owned firm reportedly making progress on a machine capable of producing 28-nanometer chips. But developing an EUV system would be an entirely different challenge, requiring China to replicate ASML's vast supply chain of more than 5,000 specialized suppliers.
Canon's alternative approach, nanoimprint lithography, stamps circuit patterns directly onto wafers, much like a printing press. In theory, this could create features with nanometer accuracy at about 40% lower cost than ASML's machines. However, it faces significant challenges with defects, alignment precision, and production speed. So far, it's found more success outside semiconductor manufacturing, particularly in making smartphone displays and memory chips where higher defect rates are more tolerable.
The outcome of this technological race will ultimately shape the future of AI. More advanced lithography tools enable the production of faster, more energy-efficient chips capable of powering new generations of AI models. While ASML currently holds the crown for the world's most important machine, the battle to control the technology that will shape computing's future is far from over.
If you're interested in the nuts and bolts of computer hardware enabling AI advancements, this is definitely a space to watch. The innovations happening in this field are just as crucial to AI's future as the software and algorithmic breakthroughs I more frequently discuss on this show.
The SuperDataScience podcast is available on all major podcasting platforms, YouTube, and at SuperDataScience.com.