The Insane Underground Machine That Changes Chips Forever

本站添加: 2026-05-30

[00:00:00]Right now, the future of technology depends on  one of the strangest machines humanity have ever built. A machine so extreme it fires lasers at  droplets of molten tin 50,000 times a second, just to create a flash of invisible light. It is  the only reason the most advanced chips on Earth even exist. For years, EUV lithography was the  absolute peak of semiconductor engineering. And now it's hitting a hard physics wall. And this  new machine they're building to break through that wall is the strangest chipmaking tool ever  attempted. I'm a chip design engineer and honestly I never thought that the future of computing will  drift so closely to the particle physics. The chips keep shrinking but the machines required to  build them keep growing. Subscribe to the channel and let me explain. Most people think AI race is  about GPUs. It's not. The real bottleneck is light because modern chips only exist if humanity  can reliably print nanoscopic structures onto silicon. And the smaller the transistor becomes,  the harder it gets. We are basically etching rocks with invisible light at that point. Lithography  is essentially photography for computer chips. You generate an incredibly precise beam of light.  That light travels through a maze of mirrors, reflects of a photo mask containing the chip  design and eventually reaches the silicon wafer below. The wafer is coated with a special  chemical layer called photoresist. And when the photons heat that surface, they chemically change  the material in extremely precise locations. At that moment, the design is effectively printed  onto the wafer. And then comes the rest of the manufacturing process. etching, deposition,  doping, stacking layers again and again until billions of transistors emerge from an ultra pure  slice of silicon, at least when everything goes right. Because the foundation of all lithography  is just one simple problem, your wavelength.

[00:01:59]It defines how small your features can be. And  eventually, the industry reached a point where the light itself became too large, way larger than  the patterns it was trying to print. For years, chipmakers relied on deep ultraviolet lithography,  193 nanometer light, but they kept trying to print smaller and smaller features with it anyway. So,  engineers started to use different tricks. You take one complex pattern and split it into several  simpler ones. Print the first part of the pattern, etch it, expose the second half, then etch it  again. Suddenly, you can create features smaller than the original light should physically allow.  This became known as Multi-Patterning. Well, we are currently doing Multi-Patterning in the  studio. And this is where cheap manufacturing started turning into an engineering maze. The  industry was forcing older lithography tools to print features they were never meant to  handle. Every extra pattern meant more masks, more alignment steps, more chances for things to  go wrong and destroy multi-million dollar wafers.

[00:03:08]And the industry kept forcing the old light to  print features. It was never designed to print.

[00:03:13]And eventually this workaround became harder than  the scaring itself. So ASML and the semiconductor industry did something extraordinary. They  jumped to an entirely new type of light. Extreme ultraviolet light with 13.5 nanometer  wavelength. Small enough to print features only a few nanometers wide. It was one of the  biggest technological leaps in history. And the problem was that just a generation of this light  pushed humanity to the edge of what was physically possible. Inside the machine, microscopic droplets  of molten tin are fired through a vacuum chamber.

[00:03:51]A first laser pulse hits the droplet and reshapes  it for the main blast. Then comes the real heat. A giant CO2 laser slams into the tin droplet with  enormous energy and the tin instantly explodes into plasma. And for a tiny fraction of a second,  that explosion emits the rare 13.5 nanometer EUV light needed to print advanced transistors. And  that is how modern microchips begin. But the strange thing is of how little of usable light  this whole complex process actually produces.

[00:04:25]After all of this complexity, only tiny fraction  of the original energy ever reaches a wafer. Some estimates place the wall plug efficiency below.1%.  Which is astonishingly inefficient way to generate light and controlling that light turns out to  be even harder. It is so extreme that even air and glass immediately absorbs it, meaning blocks  it completely. So the entire machine operates in vacuum. The EUV light bounces across a maze of  mirrors, reflects of the mask carrying the chip design and finally reaches the wafer and that  tiny flash of light is now the foundation of modern computing. But the real problem starts  at 3 nm and below shot noise. At these scales, fabs start approaching stochastic limits,  meaning there are literally not enough photons hitting the wafer consistently enough to print  perfect features. Physics becomes statistical, random. Imagine trying to spray paint a perfect  nanoscopic line. But some of the paint particles never arrive. At normal scales that doesn't  matter, but at 3 nanometers, it becomes a problem.

[00:05:36]And at some point, you start fighting probability  itself. And that's how ASML became one of the most important companies in the world because  generating this light reliably and at scale became one of the defining engineering problems  of modern computing. But now for the first time in many years the semiconductor industry is seriously  considering a different approach. Instead of exploding molten tin with giant lasers, they want  to generate EUV using electrons moving close to the speed of light inside particle accelerators.  And one of the companies pursuing this approach is the American xLight. Their idea is to build  something called free electron laser or FEL which sounds way cheaper than it will actually be  because it will cost close to a billion dollars.

[00:06:29]Free electron lasers were never invented for  chip manufacturing. They are now mostly used as scientific instruments. They are like world's most  powerful slow motion cameras for atoms, molecules, proteins, and materials. And scientists use them  to watch matter change at absurdly small scales and absurdly fast speeds. Basically, FELs were  built to study nature, but now the semiconductor industry may need them to manufacture nature's  smallest devices. FELs generate light directly from fast moving electrons. And if it works,  it will not just challenge ASML. It could push the semiconductor industry toward a  completely different kind of factory. Now, this is where the story takes an unexpected turn  because a free electron laser doesn't work like a normal laser at all. Traditional lasers use atoms.  You excite a material, the atoms release photons, and you get a beam of light. While a free  electron lasers skips the atoms entirely. Instead, it fires a beam of electrons close to the speed  of light through a long magnetic structure.

[00:07:43]The magnets force the electrons to rapidly zigzag  as they move forward. And when particles suddenly change direction at these speeds, they emit  light. At first, that light is weak. But then something interesting starts to happen inside the  beam. The electrons start grouping together into synchronized bunches. And once that happened,  the emitted light suddenly becomes dramatically more powerful and focused. By carefully tuning  the electron beam and magnetic structure, we can make that beam produce the EUV light. And this  is the moment where semiconductor manufacturing starts drifting into particle physics. So why  does the industry suddenly paying attention to this? Because traditional EUV systems are becoming  harder and harder to scale. They still work very well actually. But every increase in performance  now requires more power, more complexity and more engineering effort just to keep progress  moving forward. And at the same time, AI demand exploding. Data centers are eating up the world  and require enormous numbers of advanced chips, which means fabs suddenly need to process far  more wafers than ever before. And this is also an economics problem because the faster you can print  wafers, the more chips you can produce from a chip factory that already costs you tens of billions  of dollars to build. And this is where a FEL's technology becomes really interesting because the  single FEL machine can generate way more light than a single EUV machine. And in lithography,  more light means faster printing. potentially enough light that one centralized source could fit  multiple lithography scanners simultaneously and this would dramatically change the architecture of  modern chip factories. But there is another reason why this matters. Today's entire EUV ecosystem  is built around one specific wavelength 13.5 nm and that single number underpins much of modern  computing. But FELs are tunable. You can change the electron beam and you can change the magnetic  structure and in principle you can push toward even smaller wavelength. Some FEL facilities  already operate around 6 nanometer wavelength and some even below 1 nanometer entering the  X-ray range. And that means that in principle they could eventually print transistor features  even smaller than what today EUV systems can realistically achieve with any clever tricks.  which is extraordinary if you think about it and suddenly it's not only about improving the  current EUV systems but the question is whether there is a path for the semiconductor industry  beyond them altogether and the strange part is machines like this already exist with some of the  most interesting developments now happening in the United States Japan and China and interestingly  each of them is taking this idea into completely different direction and we will have a closer  look at that in a moment. But before that, if you're paying separate subscriptions to Chat  GPT, Claude presentation tool, image generator, you're basically bleeding money. That's exactly  why I started using Genspark. It's all in one AI workspace. And honestly, their growth has been  insane. They have reached a $250 million annual run rate in just 12 months. Traditional AI just  chats with you. Genspark actually executes the work for deep semiconductor and AI research. I  use their super agent. What's impressive is that it combines all the top AI models like Claude Opus  and Gemini together which gives you much stronger research results and cross validation than relying  on a single model. Then it can even turn research directly into slides. You can use it for AI  sheets, presentations, meeting notes, emails, and even building websites. You still decide what  matters. While Genspark removes the repetitive busy work, my favorite feature by far is their new  feature Speakly because I manage multiple teams and I need to write lots of emails and messages.  Instead of slowly typing it down, I just stake all my thoughts and it then automatically turns  them into a polished email about four times faster than typing. I genuinely love this feature. So  instead of paying separately for Chat GPT, Claude, Canva and Gamma, Genspark puts everything in  one place. one $20 subscription gives you access to the top AI models plus AI chat plus unlimited  image generation for old paid users in 2026. Check it out through the link below to try it for free  with sign up credits available. Now back to the future of chip factories. So what happens when you  actually build one of these things? To find out, we first have to look at the European XFEL in  Germany. ASML EUV machines are famous for being gigantic, the size of a double decker bus. So, how  do we solve the next bottleneck? By considering even larger machines, larger than a football  stadium. A gigantic free electron laser stretching more than 3 kilometers underground across  Hamburg. And it was never designed to print chips.

[00:13:10]Scientists use it to study molecules, proteins,  chemical reactions, and matter. Basically, it's one of the most advanced light generation systems  humanity have ever built. But then semiconductor engineers started paying attention because unlike  most FEL systems which pulse relatively slowly, this XFEL fires up to 27 times per second. And  future chip factories may require exactly this kind of industrial scale photon output powerful  enough to feed manufacturing lines around the clock. And then there is the infrastructure  itself. Parts of the machine are cooled close to absolute zero using superconducting systems. The  accelerator stretches for kilometers underground and suddenly semiconductor industries start  drifting into territory that looks nothing like a modern chip factory. Not rows of manufacturing  tools but giant infrastructure, vacuum tunnels, superconducting systems, particle accelerators  underneath the fab itself. And this is where the story gets very interesting because if this  approach actually works then the future chip factories may look more like CERN. Today every  ASML EUV scanner generates its own light locally.

[00:14:32]Each machine operates mostly independently and  even if one scanner fails the rest of the fab keeps running. But a free electron laser changes  this model completely because instead of each EUV scanner having its own light source, multiple  scanners can be connected to a single powerful light generator. And this is where economics start  to be very interesting. A sufficiently powerful FEL could generate several kilowatts of EUV  light, enough to feed up to 12 scanners at once.

[00:15:04]One particle accelerator powering an entire chip  factory. But that creates a new problem because if a FEL becomes the heart of the factory, one  failure could turn the entire factory dark. And suddenly the scale of this starts becoming hard  to ignore because shrinking transistors may soon require infrastructure on the scale of national  labs. Basically turns it into fusion between chip factory, particle physics facility and a power  plant. And then a natural question emerges.

[00:15:37]How do you actually commercialize something  like this? Factories don't work like research experiments. They need stable uptime 24 hours  a day, every single day. And this is where the company xLight enters the story. Their goal is  not to build another giant scientific facility, but to shrink this entire idea into something fabs  could realistically deploy. xLight was founded by accelerator physicists coming out of places like  SLAC, which is Stanford linear accelerator center, one of the most advanced accelerator labs in the  world. And their idea is straightforward. Take decades of accelerator research, remove as much  complexity of science project as possible and turn it into factory equipment. Instead of redesigning  the entire fab, xLight wants to replace only the EUV light source itself. They plan to keep  the scanners, keep the mirrors, keep most of the existing lithography infrastructure, just  replace the light generation system entirely. So they get rid of tin plasma and exploding droplets  using only electrons accelerated through a linear machine generating EUV light directly. And their  targets are ambitious. Around four times more EUV power than current systems with 50% lower  operating cost. But when they say compact, don't even dare to imagine something small. This is  still a giant linear accelerator sitting next to the fab rooting EUV light into the scanners. But  compared to kilometer scale facilities like XFEL we just discussed, this suddenly start looking  practical. And this is where the American strategy becomes clear. Shrink accelerator physics enough  that semiconductor facilities can actually live with it. But Japan looked at exactly same problem  and came to a completely different conclusion.

[00:17:38]At Japanese KEK they believe that size is not the  biggest problem. Power is. So they are completely focused on efficiency. That is the core idea  behind something called an energy recovery linac.

[00:17:52]Normally electrons get accelerated once, generate  light and then get dumped away as waste energy.

[00:18:00]Instead, this system tries to loop those electrons  back through the machine. And this is a clever part because as the electrons slow down, they give  a large portion of their remaining energy back into the accelerator itself. And that recovered  energy then helps power the next electron beam.

[00:18:18]It's basically regenerative breaking but  for particle accelerators. And suddenly the economics of the entire system start to look  entirely different because if you can recycle energy efficiently, you can push much higher beam  without exploding the electricity costs. That is why Japan is so interested in this architecture.  Researchers at KEK believes that future free electron laser systems could potentially generate  around 10 kilowatts of EUV power. And that is an astonishing number if we compare it to the modern  EUV systems which operate in the 100 watt to 500 watt range. And importantly, they will be using  it not just for one machine, but potentially it will be enough to power multiple scanners  simultaneously. But the catch is right now this is still experimental technology and they are  nowhere near commercial lithography yet. And this is where the story takes another turn. America is  trying to make accelerator physics compact enough for fabs. Japan is trying to make it efficient  enough to make economic sense. China decided to scale it up instead. And meanwhile, ASML is  still pushing the existing EUV architecture further than almost anyone thought possible.  You see, different countries and completely different strategies. And honestly, the most  fascinating part of this story is happening right now in China. Because for China, this is  not a technological upgrade. This is strategic necessity. Right now, China cannot buy ASML's  most advanced EUV systems because of expert restrictions. And without advanced lithography,  building cutting-edge AI chips becomes extremely difficult. Multi-Patterning with DUV can only  take you that far. So, China started to invest heavily into alternatives and one of their most  important project is so-called SSMB. steadys tate micro bunching led by researchers connected  to Tsinghua University. And unlike the American or Japanese approaches, China is not trying to  squeeze accelerators neatly into existing fabs.

[00:20:33]They want to rebuild entire factory around the  machine itself. In their concept, they're using a massive circular storage ring roughly 150 meters  around where electrons continuously circulate at near light speed. Instead of accelerating  electrons once through a straight line, China keeps them moving continuously inside the  ring. Then another laser organizes those electrons into extremely tight bunches, causing them to  emit powerful EUV light continuously. And that matters because one powerful EUV light source can  fit huge amount of EUV scanners simultaneously.

[00:21:14]So China is focusing on scale massive centralized  EUV infrastructure and at that point this looking less like competition over chips and more like  countries building entirely different industrial systems for controlling light itself. And once  you look at who is investing in this technology, a pattern starts emerging like Japan who wants to  secure its semiconductor future through projects like Rapidus. And I actually made a deep dive on  this interesting chip factory. Subscribe to the channel now so you can watch it afterwards. The  US wants alternatives as AI demand explodes. China wants independence from Western restrictions.  And meanwhile, ASML still controls one of the most important chalk points in modern technology  because whoever controls the light controls the future of chipmaking. And this is where the story  becomes less romantic because even ASML themselves considered this new particle accelerator approach  years ago, but they still stayed with tin plasma.

[00:22:22]And that says a lot because despite all that  insane complexity, today's EUV systems actually work. Not experimentally, but inside real fabs  every day. And in semiconductor manufacturing, that matters more than elegant physics. A  free electron laser may look cleaner on paper, but fabs have to be optimized for uptime. The  semiconductor industry rewards the machine that keeps running at 3 in the morning without  freezing a billion dollar production line.

[00:22:54]And this is where FEL's technology faces its  hardest transition. Scientific demonstrations are exciting. But the challenge is making particle  accelerators reliable enough for semiconductor fabs. And even if engineers solve all of that,  there is still one issue, timing. Because the AI infrastructure buildout is already happening right  now. TSMC and Intel are expanding aggressively in Arizona. In Texas, projects like Terafab are  moving at astonishing speeds and have made deep dives on both. The industry needs more compute and  more advanced manufacturing capacity right now, immediately, not in 10 years from now. Which  means ASML will continue dominating because its machines already exist and the entire  semiconductor ecosystem already knows how to build the manufacturing recipe around them.  History is full of examples where technologies were technically better but commercially late.  And this is what makes this story so interesting because the semiconductor industry may already  be building the next generation of factories before it fully knows what the next generation  of lithography will look like. Because the next era of cheap manufacturing may not be defined only  by smaller transistors, but by who can build the most extreme light source. And that creates a  strange paradox because chips and transistors keep shrinking while machines required to  build them keep growing larger and larger which is fairly unexpected direction from the  industry which is obsessed with making things smaller. Because if this approach actually  works, printing the smallest transistors on Earth may require building the largest factories  in human history. If you enjoyed this episode, make sure to watch this breakdown on  the largest and the most controversial chip factory being built right now in Texas.  Love you guys and I will see you there. Ciao.