The Insane Complexity Of The Semiconductor Global Supply Chain

TITLE: Gj5liYnpTeM CHANNEL: Unknown DATE: ---TRANSCRIPT--- On the outskirts of the city of Tai Chong, just  across the road from some rice farms, sits one   of the most important buildings in the world,  Micron’s 30,000 square meter DRAM mega campus.   Within this facility are billions of dollars worth  of cuttingedge lithography machines working 24/7   to produce the memory modules that in turn go  into military equipment, network infrastructure,   and most lucratively enterprise GPUs to power  data centers. Micron also used to manufacture   memory for home computers, but they don’t do that  anymore. There is not much that hasn’t already   been said about AI data centers, the chips that  power them, the companies that manufacture those   chips, and the machines they rely on to do it. But  all of this is really only scratching the surface   of what is very likely to be the most complicated  supply chain in the world. The chips that come   out of this particular fab are not even a fully  finished product yet. But even here, they’re the   result of over 6,000 individual suppliers from  over 40 countries, making the machines that make   the machines that make the machines that power  the world’s most advanced technologies. This is a   supply chain that’s simultaneously far more robust  than you’ve probably been told, but also full of   far more bottlenecks, which I know sounds like  an oxymoron, but I promise it will make sense.   Yes, TSMC is largely irreplaceable when it comes  to advanced chip production, and only ASML can   build the $400 million lithography machines they  use to do it. But even with this highly simplified   web of suppliers, there are dozens of other  companies where only they can do what they do.   And probably the best way to see that is to  follow the supply chains all the way from   their raw elemental components and explore  how the industry’s current production boom   is making your toilet more expensive. All  of those chips start in a place like this. The only machines in the world capable  of printing nanoscopic blueprints. Hey,   this is national security. I need to make it in  America. We don’t want China to win the AI race.   There’s a revolutionary machine that the whole  world has come to rely on. Black Blackwell is in   production. Incredible amounts of technology. In  a small factory on the southern Japanese island   of Kyushu, one of the most important tools in this  entire process gets pressed into shape. A ceramics   manufacturer takes ultra pure aluminium oxide  alongside a proprietary blend of other classified   materials and presses them into one of the most  uniform ceramic plates that anyone on Earth   currently knows how to make. The exact process is  kept very deliberately secret for obvious reasons,   but we know the basics from a US design patent  that the company filed back in 2008 and from   public engineering papers that have come out  since. The defining feature of this material   is that it barely expands when it’s heated.  And it’s also diamagnetic, which are some of   the very technical specifications that  allow it to serve one specific purpose.   You’ve probably seen videos before of computer  chips being manufactured where a polished silicon   wafer is being thrown back and forth in a sealed  chamber hundreds of times a minute as different   processes get applied to it. Well, it turns out  that actually holding that sheet of silicon in   place during all of this is its own engineering  nightmare. You can’t clamp it because atomic   scales, mechanical clamps would warp the wafer.  You can’t glue it down because the chemicals   required to dissolve that adhesive afterwards  would destroy the chips. And whatever does hold   it in place also has to survive massive swings in  temperature, pressure, and force without imparting   a single bit of stress onto the delicate silicon  wafer itself. It’s a brutal engineering problem,   and these ceramic plates are pretty much the only  thing on the planet that can really solve it.   They use electrostatic forces to keep the wafer  in place, and their atomically uniform surface   means that the force of being thrown back and  forth during production is spread evenly across   the entire piece of silicon. Within each ceramic  chuck, there are also perfectly machined channels   for flowing liquid helium through, so the silicon  stays cool even when it’s pushed into a plasma   etching chamber where it’ll be exposed to extreme  temperatures in the other direction. Again, the   specifics are highly classified and even if they  weren’t, there would be far more qualified people   on the internet to explain it. But the point is  that even at the first step of this tech tree,   the tools are hyper specialized to a point where  only a single company has been able to stay at the   cutting edge. These chucks might only be the  size of a dinner plate, but they can cost six   figures each before you’ve even included any of  the multi-million dollar machinery that they plug   into. The current world leader in producing this  component and the patent holder for this exact   style of chuck is the Japanese company Toto,  which is a business that didn’t start out as a   high-tech semiconductor supplier. Their expertise  in ultra pure ceramics actually comes from their   core business, which up until very recently was  producing the world’s finest toilets and still is.   Although their priorities have visibly shifted in  the last few years as data center customers have   started outbidding the hospitality industry for  the company’s best ceramicists. Toto is also not   the only unexpected company holding a bottleneck  on this incredibly important global industry. It’s   not even the only one based in Japan. Ainom is a  Japanese company most famous for introducing the   world to monodium glutamate or MSG. They’ve  also used their organic chemistry expertise   to produce ABF, a thin insulating film that sits  underneath almost every advanced computer chip on   Earth. ABF is short for agenom buildup film, and  it has a very specific combination of properties   that nobody’s been able to credibly match. It’s  a stable dialectric. It can be patented at very   fine pitches, and it’s extremely receptive to  copper plating. The substrate underneath your CPU   or GPU is essentially layers of ABF with copper  traces patented into them. And as chip designs   have gotten more complex, the number of layers  has crept up from four to five to well over a   dozen on a modern AI accelerator. Those substrates  themselves are then assembled by a different small   group of companies, mostly Ebiden and Uni Micron,  which take the ABF film, sandwich it with copper,   drill thousands of microscopic v through it, and  turn it into the green PCB looking square that   the actual silicon die gets solded onto. So  the next time your unemployed ass is enjoying   some cup ramen, just remember that the same  company bringing you that concentrated sodium   goodness is also keeping the modern world from  literally shortcircuiting. So far, that is two   companies with an effective monopoly on their  own particular branch of this incredibly tall   supply chain. And the reason they exist is simple  practical economics that make strange things like   this almost an inevitability. This video is  sponsored by GenSpark. Pal just wrapped up his   final press conference last month. And like many  of you watching, I wanted to know what a surprise   rate hike would actually mean for my portfolio  before markets opened. So, I asked GenSpark.   GenSpark isn’t just an alternative to chatbased  LLMs. It engages agents with access to all the top   AI models to actually produce finished work. And  they’ve gone from concept to 200 million ARR in   just 11 months. They’ve just launched AI workspace  3.0. And three features have completely changed   how I work. GenClaw works like an AI teammate that  you can message to handle research, write updates,   manage emails, and schedule things, delivering  the finished result back to you. Their super   agent turns any macro event into a full visual  report instead of a wall of text. So with that Fed   scenario, it instantly built historical back tests  and scenario simulations showing how the S&P 500,   NASDAQ, longduration treasuries, and gold  have moved across past tightening cycles. and   meeting bots join your calls, take structured  notes, and send clean summaries afterwards,   so you stay focused on the conversation instead  of documenting everything. None of it tries to   replace how you work. It just handles the parts  that slow you down. That’s why you’ve got to   check out GenSpark. Every new user gets free  credits just by signing up. So, I’ll leave a   link in the description below. Okay, so inside  the chip fabrication facilities themselves,   the lithography machines tend to get all the  attention. And there’s one company in particular   that’s become the center of global attention here.  ASML. ASML is a Dutch company that was originally   spun out of a 1984 joint venture with Philips  and today they make the machines that basically   laser print chip designs onto silicon wafers.  ASML itself has over 5,000 individual suppliers   feeding into a single machine. There are trump  lasers from Germany, Zeiss optics polished to a   smoothness measured in single atoms, industrial  power supplies that have to be more stable than   the ones running a hospital, and vacuum pumps that  operate by knocking individual atoms out of the   system one at a time. Each and every one of these  examples is again producing components with no   viable alternative. The most advanced machines are  physically so large that ASML has to ship them in   pieces using three Boeing 747s, plus a small army  of engineers who travel along with them and spend   weeks reassembling the thing on the factory floor  before it can be plugged in. As much as possible,   all these suppliers are European or North American  because of the geopolitical concerns around this   technology. But some pieces of this sub supply  chain like the rare earth magnets in optical   assemblies really can’t avoid China entirely. Now  you might have heard that ASML is the only company   on Earth that can produce these photoiththography  machines which is not strictly true. Both Nikon   and Canon yes the camera makers also produce  lithography machines for semiconductors.   The reason ASML is distinct is because it is so  far the only company on the planet that can build   the twin scan NXE3800E in particular, which is  currently the only machine capable of cuttingedge   singledigit nanometer resolutions. To get there,  the inside of one of these machines has to fire a   high-powered carbon dioxide laser at a falling  droplet of molten tin 50,000 times a second,   vaporizing each droplet into a plasma at around  220,000° centigrade. And the resulting flash of   extreme ultraviolet light is then bounced off  a stack of Zeiss SMT mirrors to project a chip   pattern onto the wafer below. Each one of these  mirrors is the smoothest manufactured surface in   the world. According to their own manufacturer,  if you scaled them up to the size of Germany, the   largest bump on the surface would still be smaller  than a millimeter tall. The Canon and Nikon models   are at best a generation behind on resolution,  but that actually doesn’t matter all that much.   Yes, high-end chips on the bleeding edge are  a massive market and one that gets the most   financial and geopolitical attention. But most of  the chips being produced today are not going into   AI data centers or fighter jets to toasters, Wi-Fi  routters, hospital monitors, washing machines,   and your latest vaccine. These mass market chips  are perfectly happy to use those slightly older   Canon and Nickon machines alongside the more basic  models that ASML still builds themselves at the   bottom of their own product range. ASML is also  one of the only companies in the world that the   American government has put export controls  on despite not actually being American. So   for any customer that doesn’t strictly need  the bleeding edge, it’s often a lot easier   to just buy a more basic machine from another  lithography manufacturer and skip the diplomatic   paperwork entirely. Even you, if you were so  inclined, could purchase one of these more basic   lithography machines, and they are surprisingly  affordable. This one from this uh very respectable   looking website is barely more than a fully loaded  modern pickup truck. Now, lithography is important   and it does grab a lot of attention, but it’s far  from the only worldclass tool inside one of these   facilities with no other global alternatives.  Despite the fact that you’ve probably never   heard of it, Applied Materials is currently the  40th most valuable company on the planet because   they make a huge chunk of the other machines  needed to actually build a microprocessor. We   are talking about deposition tools that lay  down layers of metal and insulation one atom   at a time. Iron implanters that put the semi in  semiconductors, chemical mechanical polishers that   flatten everything back down between layers, and  metrology tools that constantly check whether any   of the previous steps did what they were supposed  to do. None of these get anywhere near the press   coverage of an ASML EUV scanner. But without  a fab full of them, even a TwinCan NX 3800E is   just a very expensive laser pointer. And Applied  Materials doesn’t sit alone in this tier either.   Lamb Research is another American company  that completely dominates plasma etching,   the step where patterns drawn by the lithography  machine get carved into the actual silicon. KLA,   also American, makes most of the world’s wafer  inspection and process control equipment.   Essentially, the very expensive microscopes and  metrology systems that tell the fab whether the   chips being produced are actually any good before  too many of them get scrapped. And Tokyo Electron,   usually just called Teel, is a Japanese company  that quietly holds near monopolies on the Kota   developer track tools that handle every wafer  before and after lithography, as well as on a   long list of thermal processing and cleaning  steps that nobody outside of Fab Engineering   really thinks about. Between ASML, Applied  Materials, Lamb Research, KLA, and Tokyo Electron,   you have the so-called big five of semiconductor  equipment. And they collectively account for the   overwhelming majority of every dollar spent inside  a leading edge fab with each one of them being   the undisputed world leader in their particular  place on the production line. Although only one   of them ever really gets that much attention.  And partially that low profile is by design.   Applied Materials, Lamb, and KLA are all American,  so they’re naturally less in the geopolitical news   than the more exposed Dutch ASML, and there’s a  good chance that most regulators have also never   heard of them, making it much less likely for them  to interfere in their operations. Any country that   wanted to be fully independent of the globalized  chip supply chain would need to replicate all five   of these businesses, plus their thousands of  individual suppliers, which would take decades   to do, if it were even possible at all. A common  characteristic is that a lot of these companies   had been honing their industrial processes since  before modern silicon semiconductors even existed.   Starting from fundamentals again would involve  a lot more than just reverse engineering some   machinery. Now the best of all these machines all  eventually flow into just four companies. Samsung,   Global Foundaries, Intel and above all TSMC. With  the exception of Intel, these companies do not   design any of the chips themselves. They just take  the designs sent by their clients and manufacture   them on their behalf. Now, this would be a little  bit like Ford designing the new Focus and then   subcontracting out the entire manufacturing to  a completely separate business that also happens   to be building the new Toyota Corolla on the same  factory floor for a direct competitor. It is a bit   unusual, but the technical skill set required to  operate this kind of facility is so concentrated   and the equipment so eyewateringly expensive  that figuring out how to do that in-house just   doesn’t make any sense for any chip designer  that needs to ship new technology this decade.   This has of course caused further geopolitical  concern because TSMC is by far the largest of   these four manufacturers currently responsible for  somewhere north of 90% of the world’s leading edge   chip production. And the T in that name stands  for Taiwan which sits squarely in the crosshairs   of the People’s Republic of China and could  potentially disrupt this very delicate supply   chain if and when a conflict ever broke out.  It would be a bit like the straight of Hormuz,   but if 80% of the world’s oil flowed through  there instead of just 20%. And that oil was   made of glass. And we’d built the world’s most  concentrated pool of investment in history based   on that glass. TSMC has under pressure from the US  tried to mitigate this risk by building new fabs   in Arizona and Japan. But the most advanced node  is still made almost exclusively in Shinchu and   Tynan. And the company has openly admitted that  its overseas facilities are running a generation   or two behind for now with the Arizona ramp up  in particular having faced repeated delays and   cost overruns. This is not because they don’t  have access to all the best tools from the big   five. It’s because even with access to these  machines, actually producing anything useful   with them still requires decades of soft skills  that have been notoriously hard to transfer. Now,   before even the best of these foundaries can do  anything, they need the raw material itself. Bare   300 mm silicon wafers, the round mirror polish  discs that every chip in the world starts its   life on, are themselves, well, you wouldn’t  believe it, made by an extremely concentrated   group of suppliers. Two Japanese companies,  Shinetsu Chemical and Sumo, between them account   for somewhere around half the entire global wafer  market with smaller shares held by Global Wafers   in Taiwan and Silronic in Germany. The process  starts with high purity quartz sand which gets   refined into electronic grade polysilicon. A  material so pure the contamination is measured   in parts per billion which then gets melted  and pulled into a single crystal ingot using   the Trrowski process which then gets sliced,  lapped and polished into the actual wafers.   Each of these steps is its own art form. And a  meaningful disruption at any one of them ripples   through every single fab on the planet, which  is exactly what happened in 2021 when a fire   at one of Shinu’s suppliers contributed to  a global semiconductor shortage that ended   up sitting halffinish cars on dealership  lots for almost 2 years. just as important   to the system as Logic Foundaries are to the  memory manufacturers. DSMC, Global Foundaries,   and Intel really only focus on logic chips, but  computer memory is manufactured using a similar,   though not identical, set of machines, and  has very clearly become just as important to   global business. Without DRAM and increasingly  without the stacks of high bandwidth memory that   get packaged directly next to a GPU, even the  most advanced logic chip would effectively be   useless cuz it would simply have nothing to read  from or write to fast enough to matter. Samsung   is the single company in the world that does both  logic and memory alongside SKH Highix and Micron   which are the world’s leading dedicated memory  manufacturers and collectively supply more or less   100% of the chips going into any business worth  mentioning in the space. SKH Highix in particular   has become the sole qualified supplier of the  most advanced HBM3e memory stacks that Nvidia   bolts onto its top-end accelerators, which is a  fun position to be in until you realize that the   entire AI data center boom is currently sitting  on top of a memory chip designed in Inchon and   packaged with bonding wire processes that almost  nobody outside of South Korea actually knows how   to do at volume. And the bottleneck doesn’t  stop at the memory itself. To actually attach   those HBM stacks to a GPU die, manufacturers  use a process called co-wash, chip on wafer on   substrate and co-wash capacity. In 2025 was almost  entirely owned by a single facility inside TSMC,   which means that for a stretch of about 18 months,  the global supply of Frontier AI accelerators   was effectively rate limited by how fast a single  packaging line in Taiwan could move. Nvidia could   design as many chips as it wanted and SKH Highix  could manufacture as much memory as it wanted. But   if TSMC’s packaging capacity didn’t expand, those  parts simply could not be combined into a finished   product. Now, as a special mention, if you thought  we were done with the bottlenecks, Nvidia, Apple,   AMD, Broadcom, and Qualcomm do not manufacture  any of their own chips. They only design them.   But they also don’t design what they design  with. Just drawing the blueprint for a modern   chip requires highly specialized software that  can simultaneously model out tens of billions of   individual transistors while also calculating how  those transistors will perform at a scale where   quantum uncertainty actually starts to matter as  a real engineering input rather than a theoretical   curiosity. There are again only two companies on  Earth that produce this software competitively.   cadence and synopsis. Together, they sit at the  very top of a software stack that the entire   semiconductor industry uses to verify whether  a design will even work before it gets sent off   to a fab. Now, as someone who personally likes  to complain about paying my Adobe subscription   every year for a product that I used to be  able to buy just once, these companies take   software licensing to a level that I think  most people would find genuinely unhinged.   an annual subscription to use their tools can cost  over half a million dollars US per year per seat,   which means that a relatively small team of  10 engineers will run one of these companies   several million dollars a year just on the  software. And they pay it because there really   is no replacement. If a chip designer tried to  migrate to anything else, they would effectively   be rebuilding their entire design pipeline from  scratch on tools that haven’t even been used in   production by anyone they actually trust. And  just to make this picture slightly clearer,   both Cadence and Synopsis are American companies,  which means that the US government can and has   added their software to its export control list.  Now again, this has been a very narrow look at a   very narrow selection of the companies that this  whole process relies on. We have barely touched   on the photoresist chemistry coming out of JSR and  Tokyo Oka, the specialty gases from Air Liquid and   Lind, the mask blanks from Hoya, the bonding wire  from Tanaka, or the dozens of subsystem suppliers   buried two and three levels down inside the bill  of materials for a single EUV scanner. It’s an   industry that has encouraged and required hyper  specialization into techniques sitting at the   cutting edge of what’s currently scientifically  possible. And that means there’s a lot of single   points of failure scattered along the way, but  also a lot of companies that have become very,   very good at working together because there really  is nowhere else for any of them to go. Sure,   a lot of these companies may be the only supplier  in their market, but the companies they sell to   are also the only buyers. So, their monopoly  power is somewhat offset by the monopsin power   on the other side. And for as long as that remains  true, every new AI data center coming online is   going to keep bidding up the price of ultra pure  aluminium oxide, MSG grade organic chemistry,   Zeiss grade optical glass, Korean memory packaging  expertise, Taiwanese substrate assembly capacity,   and for that matter, the world’s finest Japanese  toilets. But if you think all this sounds a little   bit unsustainable, it’s got nothing on global  ship building. Go and watch this video next to   find out why only three countries still build the  most important tools in the modern global economy.