Is China Entering the Global Memory Supply Chain?

Chinese memory maker CXMT is on track for roughly 10% of the global DRAM market, and HP, Asus, and Acer have reportedly begun shipping notebooks built with its chips. Adam Goodwin and Greg explain why memory behaves like a commodity, why bit density determines both performance and cost, and how a three-firm market led by Samsung, SK Hynix, and Micron ended up short on supply. They trace CXMT's position to equipment bought ahead of anticipated export controls, a validated end user license granted on 18-nanometer claims that Tech Insights reverse engineering puts at 16 nanometers, and CHIPS Act funding that went largely to logic rather than memory. 

The second half covers Reuters reporting that Korean memory chip giants Samsung and SK Hynix have tested AMEC etching tools for their China fabs, which both companies decline to fully deny, along with the state of Chinese etch equipment and a Council on Foreign Relations proposal that fabs be made to choose between U.S. and Chinese tools.

Chapters

1. DRAM and memory Overview (00:10 - 09:18)

2. Existing memory market and CXMT’s rise (09:18 - 13:05)

3. Why firms are turning to CXMT, why is Chinese memory an option, and export controls (13:05 - 19:03)

4. Chinese subsidies, and Congress and Apple are pulling in opposite directions (19:03 - 22:26)

5. Furnishing Korean fabs in China and tension with U.S. government (22:26 - 30:35)

6. Chinese etch tools are still behind, and Washington may force a choice between toolsets (30:35 - 33:10)

Transcript

SPEAKER_01 0:09

Welcome back to the AI Power Podcast. Today we'll focus on recent reports that China is entering the global chip supply chain with memory providers using Chinese equipment and hardware companies using Chinese memory chips. HP, ACES, and Acer have started shipping notebooks built with memory from a Chinese company. Before we get to why, set the table for us, Greg. Who actually makes memory? What is DRAM? And when we hear a number attached to a memory chip, what are we actually measuring?

SPEAKER_00 0:40

Okay. So in the semiconductor debate, so much of the conversation in the United States has been about logic chips, the kind of chips made by Intel, the kind of chips made by NVIDIA. The surge in memory prices over the past year have made everybody wake up and remember oh, right, memory chips are super important too. Now, the big difference between memory chips and logic chips is that logic chips are much more specialized and much more diverse. And memory chips are much more like a commodity. And so you can think about it in terms of oil is a commodity or copper is a commodity or steel is a commodity. If you don't buy somebody's copper, you can probably buy somebody else's copper. Whereas if you don't buy Intel's chips and you want to switch to NVIDIA's chips, well, those chips are for totally different things, right? You can't just put an NVIDIA chip in your laptop and have it work as though it was an Intel chip and run Windows on it and stuff. So there's a lot more diversity to logic chips. That's one big thing about the difference between logic and memory. In memory, the market is much more concentrated and consolidated. And that is because the economics are very different. Again, you know, tied to it being a commodity. In some ways, it's kind of a scary business to be in because if your competitor comes in and undercuts you on price, they can steal all of your customers very quickly, right? As long as their product is equivalently good. Now, Moore's Law does not work the same way in memory as it does in logic. It still is the case that the chips today are radically, radically better than the chips of a long time ago, but it's not precisely the same sort of Moore's Law type exponential growth trajectory. And in the past decade or so, the pace of memory improvement has slowed considerably. We're still making exponential growth in terms of performance per dollar, but not to the same extent as logic. It's really, really expensive and hard to get denser. So, what is density? Well, recall that we start in a chip factory with a blank silicon wafer. It's just a big circle, looks like a record plate, and it's basically looks like a mirror at the start, because it's it's just blank. And then we print the computer chips on that using all of this fancy semiconductor manufacturing equipment. So, what is density? Density is how many gigabytes of memory you can fit in each chip on a per area basis, on a per square millimeter or square inch, if you're uh here in the good US of A type of basis. And because so much of your cost structure is tied to how many wafers you process. If you want to make twice as many wafers, you probably need twice as many machines. You probably need twice as much factory floor space, you might need twice as many employees, and so on and so forth. So if you have higher density, if you have higher gigabytes per square millimeter, that means you have a greater amount of memory per chip wafer, which means not only is your product better, it's also cheaper because your cost structure is less, right? If I can fit 100 gigabytes on one 10 millimeter chip and my competitor can only fit one gigabyte on a 10 millimeter chip, then that means they have to use way, way, way more wafers to produce the same amount of this commodity, this commodity being gigabytes of DRAM. So the reason why I mention all of this is it's very hard to compete in memory if you're not at the frontier of technological performance. In logic chips, if you make old logic chips, there's people who want to buy old logic chips because their specific application requires old logic chips. But in memory, if you're not making the latest and greatest chips, then you're probably inferior in bit density. And if you're inferior in bit density, not only is your product worse performing, it's probably more expensive. So it's a commodity, but it's a viciously competitive market to be in, where it really, really matters to be in the frontier. Okay, that's the background of all of this stuff. Then there's the market context. Three companies, two Korean, one American, control more than 90% of the world's DRAM. Those companies are Samsung, SK Heinex, and Micron. Well, recently, China has sought to get into this market. And they've done that multiple times. Multiple times, it's been something of a disaster. Fujin Jinhua was like a train wreck. Qinghua Unit Group was kind of a train wreck in China. But CXMT, this latest and greatest one, has really got a lot of momentum behind them to the point where it looks totally plausible that they might capture something like 10% of the world's DRAM market share by the end of next year. How did this happen? Well, it's actually not because CXMT is at the technological frontier, their stuff is worse. It's not because they're the cheapest, their stuff is actually more expensive because they're not at the technological frontier. I mean, they get some cost reduction from the Chinese government subsidies, but they have all these cost disadvantages from not being at the technological frontier. So you might ask yourself, how in the world are they about to capture 10% market share? And it all has to deal with the market timing dynamics of the memory industry. Building a new memory fab is kind of like building a cruise ship. So for those of you who don't know about cruise ship economics, there's rooms on a cruise ship. So the total market, you can measure it in rooms, like the number of passengers who are going to stay in a number of rooms. And if there's more passengers than there are rooms, then there's a shortage of cruise ship capacity. If there's more available rooms than there are passengers, then there's a surplus of capacity. Well, when there is a shortage, it looks like a great business to be in. Everybody starts building cruise ships because they're like, wow, these rooms are going for a good price and people are making money hand over fist. What a great industry. But it takes a long time to build a cruise ship. These things are freaking huge. It's like a skyscraper turned sideways, and it has to have an engine, it has to float, blah, blah, blah, blah, blah. So when there's a shortage of capacity, room prices are really high, everybody starts building cruise ships. And then two years later, a bunch of cruise ships come off the dockyards, and now there's a surplus. The price per room crashes. And then everybody says, Oh my gosh, why would you ever build a cruise ship? What a terrible business. Everybody's losing money like crazy. So nobody builds a cruise ship. And then a few years later, there's a crazy shortage of capacity. So everybody starts building again. And on and on and on the cycle repeats. Very similar to the memory cycle. When memory prices are very high, everybody wants to build memory fabs, the factories that make memory chips. When memory prices are low, nobody wants to build memory fabs. And three years ago, it takes between three and six years to build a memory fab. You can retrofit an old campus in two years, maybe if you have to. But three years ago, memory chip prices were in the toilet. SK Heinex was at one point called a quote, zombie company, right? Well, last year, memory prices were so sky high, and now they're so sky high, that over the past year, Micron stock price has appreciated 700%. It was the best performing stock in the entire American stock exchange.

SPEAKER_01 9:18

They're doing good business this year.

SPEAKER_00 9:19

Exactly. Because they were making money hand over fist because they were able to sell memory chips, and there was so much demand. Now, why is there so much demand? We've talked about this on a previous episode. Because high bandwidth memory, the kind of memory that you want for AI chips, is just stacked DRAM. There's some like fancy ways that you can stack these chips on top of each other, give them a really high bandwidth. That's exactly what you want, high capacity and high bandwidth for AI chips. So the shortage in HBM directly translates to a shortage in DRAM. Now that shortage was all baked in a few years ago in 2023, when prices were in the toilet and nobody wanted to build memory fabs. And some people saw the AI revolution and the future of data centers coming, but the memory chip companies evidently didn't. They didn't believe it enough to start investing massively, with one very important exception. CXMT. And the irony here is CXMT, they're not like some super geniuses who saw the AI memory shortage coming and decided to invest like crazy in 2023. Actually, what it was is they were terrified of export controls. They're like, oh my gosh, the Americans just banned all these AI chips. They just banned all this equipment, but there's all this other equipment that we're still allowed to buy, but they might ban that next year. So let's buy everything. Let's buy everything during this period of time when we're allowed to buy stuff. So at the same time that Samsung, SK Heinex, and Micron were dramatically scaling back their capital investment. That's the exact same time that CXMT is almost alone among DRAM providers in ramping up their capacity. And they just happened to be catastrophically right, at least from the US perspective, catastrophically right. Because now there's this crazy shortage of DRAM because everybody wants to not make DRAM. They want to make HBM because they can make higher profits selling HBM to data centers than they can selling DRAM to computer makers. And so those computer makers and those smartphone makers that you mentioned in your opening, they're going to CXMT and they're like, hey, is your memory very good? And CXMT's like, no. And they're saying, hey, is your memory very cheap? And CXMT's like, no. And they're saying, but do you have any memory at all? And CXMT is like, yes. And that is what is going on right now. That is why CXMT was the best performing IPO in the history of the Chinese stock market, because they are out there making chips. And again, to make the sort of comparison to oil, for Saudi Arabia, they can make a profit when oil is only $30 a barrel, because it only costs them $20 a barrel to pull it out of the ground. Meanwhile, like most of the oil fields in Texas, if oil is below $60 a barrel, they all get killed, right? So CXMT, they're the high cost producer, but the price windfall is so insane. The supply shortage is so insane that they are miraculously well positioned. And it's so intense that you've got these companies like HP, an American company, ACES and Acer, which are Taiwanese companies, but they're very, very important in the laptop computer market. They're now sourcing from CXMT, although they're trying to keep it a secret. And as we discussed a while back, Apple is trying to get permission to source from CXMT.

SPEAKER_01 13:05

So we've kind of talked a little bit about the market economics of why this is occurring, why it would be attractive for these companies to reach out to CXMT kind of on this basis. How did we get to a point where Chinese memory is actually an option for US brands at all? Right. Can you take us through how the US export controls have shaped this industry and why we are where we are right now?

SPEAKER_00 13:28

Yeah. So I think folks on this show who listen to this show will recall our conversation about DUV lithography versus EUV lithography. And memory is different than logic in that EUV became essential later in the story. Basically, you had to get EUV machines after 2022 if you wanted to compete at the frontier of logic chip technology. In memory, it was an option, but it was nowhere near as essential. Samsung and SK Heinrich started adopting EUV to a greater or lesser extent in like 2020, 2021. Micron didn't adopt EUV until 2025. And they remained competitive. There were ways that you could remain competitive in terms of chip quality and price without using EUV technology. And EUV is the big category of technology that has been export controlled. There's other important categories of technology in inspection and etching and metrology that have been export controlled. But the long story short here is that CXMT miraculously got permission to buy machines from America, even though I would argue there was a very clear case to export control them and prohibit them from buying such machines. But the Dutch and the Japanese were basically saying, like, hey, the market is in the toilet in 2023. These guys want to buy, let us sell to them, please. We desperately need somebody to buy our stuff in the midst of this dearth of purchases. And so the government has multiple types of export controls. They have countrywide export controls, they have end use export controls, and they have end user export controls. Now, the end use export controls for semiconductor manufacturing equipment were based on how sophisticated the memory you're going to make was. Specifically, we talked about an 18 nanometer feature size or half pitch in DRAM. And CXMT basically says, well, our memory is going to be worse than that. And so the US government grants them a validated end user license, and then the floodgates open. They start buying all the equipment they're allowed to buy. Now, there's been some reporting that suggests that number one, this was actually in violation of export controls because now they're making memory that, according to Tech Insights reverse engineering, is actually 16 nanometer memory. In other words, they lied about what kind of memory they were going to make in order to get permission to buy the machines. And then there's other reasons where you where you could point and say, like, how exactly are they getting this status? But the Biden administration didn't block it. The Trump administration hasn't blocked it. And so they've been ramping up capacity in a massive way. And even though their product's not the best, even though it's not the cheapest, they have it when almost no one else does. And everyone is desperate right now. The laptop computer market is expected to shrink worldwide by 10%. And it's not because of some big economic recession. It's literally because they can't make the laptops. They don't have the DRAM to make the laptops. It's causing this overall market shrink. So that should give you an indication of just how desperate these companies are and why they might be willing to go for a Chinese supplier when normally they would never consider such a thing, especially outside. So CXMT's previous customers were all Chinese companies. That has been true until extremely recently in a past handful of months. All their customers were not just Chinese companies, but like the low-end smartphone producers or tablet producers or server makers in China. But the shortage is so desperate that suddenly they have this opening window. And like, if you go back to the CHIPS Act for a moment, we wanted to invest $52 billion in chip manufacturing in this country. The overwhelming amount of investment went to logic chip companies. And if you could get in a time machine and go back and say, cancel everything and just give a ton of money to Micron and tell them to build two memory fabs. You can build one of these for like $25 billion. The CHIPS Act would have returned hundreds of billions of dollars in profit to the economy. Now, of course, knowing the future is always very lucrative, but it's kind of amazing just to look back at how the American export controls, I would argue, missed the boat in terms of an opportunity to block CXMT, which sooner or later is going to be a potent force in China's AI ecosystem. And then secondarily, the CHIPS Act missed an opportunity to really head off this memory crisis, which it had time to do if it had moved quickly and with laser precision, which is basically what China did. I mean, this is a mostly state-owned company. Who do you think gave them all the money to buy those machines? It was the Chinese government which was anticipating the future export controls.

SPEAKER_01 19:03

Yeah. What kind of support has CXNT gotten from the Chinese government, right? How does this fit into the overall Chinese government strategy for this technology?

SPEAKER_00 19:12

So there's like massive subsidies of every kind for everyone in the chipmaking ecosystem of China. And that's been true for a long time now. Sometimes, you know, Chinese subsidies are structured more like equity investments, but they are best understood as subsidies. And I realize that might be confusing, but it has to do with the nuances of Chinese industrial policy. And it's basically just been unlimited money for everybody who can buy all the semiconductor equipment they want before the door slams shut.

SPEAKER_01 19:44

So when we talk about the larger market economics here, would bring Chinese memory companies into the market actually ease this shortage and start to reverse some of the trends we've seen, like the one that you talked about with laptops?

SPEAKER_00 20:00

Yeah. So there's different ways you can think about structuring this question. Number one, does CXMT have enough capacity? Will they have enough capacity to end the current shortage? No, not even close. Samsung, SK Heinrich, and Micron are all talking about doubling or tripling their capacity over the next five years. That's how insane this shortage is. CXMT is not a solution to the shortage by any means. What they are is just any new capacity coming online at this current moment, which everyone is desperate for. And again, they're not a low-cost producer. They're just matching the prices that Samsung, SK, and Micron are charging right now, which is like the going market rate. They're not undercutting, they're just saying we have this capacity.

SPEAKER_01 20:58

The reaction from the US government on the other side has been interesting. We've seen a number of different letters coming from members of Congress about potentially blacklisting or entity listing CXMT. But to the point that you were just mentioning there, we haven't seen much interaction from Congress about energizing supply.

SPEAKER_00 21:18

Yeah. So the House Select Committee on China, the Chinese Communist Party, sent a letter to the Department of Commerce on July 16th, basically saying that this should happen. But at the same time, you've got companies like Apple, which are saying loosen the restrictions. And what's funny is like CXMT memory is not even going to go into any Apple device, even if the government said, yeah, do it right now. It's not going to go into any Apple device this year. It's not going to go into any Apple device next year. I mean, it's going to take a long time before CXMT is qualified as a supplier to be in Apple devices. But that's how far ahead they're thinking. I mean, a lot of people think this memory shortage is going to continue until. 2028 or beyond. And remember, once the implications of Chat GPT and its uh descendants for the memory market became clear, well, now SK, Samsung, Micron, they're all investing like crazy again. It's just it takes a long time for those new fabs to come online and to be built. But when they do, they will dwarf everything that CXMT is building.

SPEAKER_01 22:25

Yeah, certainly understood. Certainly understood. So, Greg, the Korean memory companies are being squeezed by CXMT on one side. And on the other side, their own fabs inside China have also been in the news, right? So Reuters has reported that Samsung and SK Heinex have been testing Chinese-made etching tools for their fabs in China, and both companies deny it. What is the actual allegation here? And what does an etch tool actually do?

SPEAKER_00 22:55

Yeah, so one important thing to point out here is we've been talking about the Korean companies, the American companies, the Chinese companies. But in the case of the Korean companies in particular, a lot of their manufacturing capacity is actually in China. SK Heinex has its Whooshi facility, which I don't know if this is still true, but at one point, 40% of the entire memory making capacity of the entire company was located in China. So when the export controls hit and said China's not allowed to buy all these kinds of advanced machines, et cetera, those facilities were at-risk assets. Because again, memory, if you can't produce at the frontier, you might be lower performing and more expensive. So not being able to upgrade your facilities is terrifying if you're a memory maker. Most memory facilities swap in and out machines every year or two. Whereas a logic facility, sometimes they just buy the machines they buy, which maybe are state of the art when they buy them. And then that's the same set of machines they're using 20 years later, right? Because they just go from being a state of the art facility to being a Stone Age facility and they just keep making the same quality of chips the whole time. Memory, you kind of can't do that. You need to refresh the equipment to stay competitive, to stay at the leading edge. And so when the export controls threatened those facilities, that was terrifying to the memory companies. And they ultimately got a permit from the American government to at least maintain those facilities at their current capacity, but not to expand the facilities in any way, basically saying all future growth for your company needs to take place outside of China, which, by the way, what a gift to the Korean people the American government gave. Basically said, sorry, these companies, your national champions, are required to create jobs in your country, not in China going forward. Remember how I said SK is like tripling or whatever their manufacturing capacity? That tripling is going to take place in Korea. It's not going to take place in China, which, in the absence of these export controls, absolutely could have been the case. But companies didn't like that. It's not like it was fun to be threatened to be cut off from chipmaking equipment. And so, at least according to the reporting that we have, and this at this stage is people talking off the record, it's unconfirmed. But the allegations are that they are now testing some Chinese semiconductor manufacturing equipment. So there's a lot of different types of semiconductor manufactured equipment. There's lithography, there's etching, there's deposition, there's inspection metrology, there's water cleaning systems and spraying systems. Some of it is, oh my God, I can't believe humans did this complicated. And some of this is like, yeah, it sprays water and it spins the plate. Okay. Some of it's crazy impressive. Some of it's, you know, just sort of pretty simple automation and robotics, but in a high purity type of context. So the Chinese, obviously, have been trying to increase their share in the semiconductor manufactured equipment market for ages. For a long time, they really only had progress to show in the packaging segment of the market, which is sort of the least technologically sophisticated. But slowly but surely they've been working their way up in sophistication to like that water spraying disc plate, you know, spinning thing that I was talking about. And now they actually have some stuff that's halfway decent in, for example, like the etching market. So when you want to create a chip, when you look at it, it looks like a square or a rectangle. But at the microscopic stage, it actually looks more like a city, like a forest of skyscrapers. And those layers are printed one layer at a time. And so it's all about just like depositing material according to the photolithography plan and then etching away material. And that staging of lithodepo etch, lithho depoetch, on and on and on happens a bajillion times. And that's how you turn this sort of flat silicon plate into having these microscopic cities on them that are stacks of many, many different kinds of layers. And so AMEC in particular is probably the Chinese semiconductor equipment company that is climbing the technological sophistication ladder the fastest. And the Koreans, if this story is true, one interpretation of this story is they're kind of threatening the American government. They're saying, hey, if you cut our facilities off again, we might have to buy equipment from the Chinese. Now, that could be sincere. That could be a negotiation tactic. That could be a way to curry favor with the Chinese government. There's a lot of different interpretations of that behavior, which again has yet to be confirmed. But if you're an export controls guy, it's an exceedingly concerning development.

SPEAKER_01 28:00

So, Greg, both companies involved have said that this is about maintaining existing lines and doing the existing business, but not expanding beyond that. Why is servicing the thing that scares both of them?

SPEAKER_00 28:13

Yeah. So I guess I should say, you know, specifically like what we know and what we don't know from the story. So Reuters did go to Samsung and SK Heinex, the big Korean chipmakers, and say, is this true? You know, or can you confirm our deny our reporting here? And Samsung has said it's not testing AMEC equipment for use in its China factory. SK Heinex says it has not tested AMEC tools for its use in its China factory, but neither company is denying testing. So we really don't know the full extent of the story here, but there's definitely something, right? Like they're not denying testing the tools. What they, I think, are worried about is that the United States is going to make it impossible to service the equipment that they have in China, or that the United States is going to make it impossible for them to keep getting new equipment as stuff wears out, like spare parts, materials, et cetera. At this stage, I don't think that's incredibly likely. I mean, I think something that is like a real risk for these companies is China nationalizing those facilities at some point. And that's another reason why you might want to play nice with the Chinese equipment companies, right? To lower your perceived nationalization risk. If you're like, oh, you know, we play nice with the Chinese companies as well. The Koreans are in a tough spot here, but I don't think we should let them use these Chinese tools. I just think the overall incentive structure for the United States and for its allies is very clear. Every step on the staircase of progress that Chinese semiconductor equipment companies make is bad for the allies, full stop. And so getting some takeup from a world-class chipmaker like an SK or a Samsung, they're going to learn a lot, right? Samsung and SK, they want the tools that they spend their money on to work. So they're going to troubleshoot a lot of stuff and they're going to educate the companies a lot. If it was anybody but China, they would love additional competition in the semiconductor equipment market. Just the problem is inevitably China is going to point those machines at CXMT, at YMTC. So the better they get, the more competitive those competitors of Samsung and SK Heinex can get.

SPEAKER_01 30:35

Yeah. And Greg, it's probably worth just having a baseline for this part of the conversation. How good are the Chinese edge tools now? And how much cheaper are they just to operate and for the customers?

SPEAKER_00 30:47

Yeah. So there's there's a lot of diversity in types of equipment and the levels of technological sophistication differ from one piece of equipment to another. We talked a bit about lithography not that long ago and reported progress in Chinese lithography, depot etch inspection of metrology, like their level of technical sophistication differs. I mean, Dan Hutcherson, vice chair of research at Tech Insights, which is like a semiconductor industry consulting firm, he made this sort of ballpark statement that broadly Chinese etch tools are 20 to 30% worse compared to their foreign competitors. Worse can mean a lot of different things, but I think that's probably true. These pieces of equipment are nowhere near as good as the global state of the art. You need the global state of the art for a lot of stuff, but they're also not useless.

SPEAKER_01 31:41

So uh we know that Chris McGuire at the Council on Foreign Relations has argued that Fabs should have to choose US tools or Chinese tools. What happens if Washington forces that decision?

SPEAKER_00 31:55

Well, I mean, I think it's a very plausible outcome here. I do think that's a credible step that the U.S. government can take because South Korea in this scenario would be saying something to the effect of, hey, you threatened to cut off our tools. And so we need to explore other options to reduce our dependency on you. So we're going to dip our toe in the water of Chinese manufactured semiconductor equipment. And the Americans would respond, if you don't have our tools, you're doomed immediately. And if you dip your toe in that water, we will crush you. Right. And that would be the nature of the threat. So we'll see how that conversation goes.

SPEAKER_01 32:42

Well, Greg, it seems as though we revisit the conversation about memory and kind of the memory shortage, you know, every couple of weeks. And I expect us to have this conversation again in a few weeks. But for now, that wraps up our discussion about China's entry into the global memory chip and equipment supply chains. Thanks as always to Greg and to our audience for listening. We'll be back next week with a new episode. With the pace at which things have been developing, I'm sure we'll have plenty to talk about.

SPEAKER_00 33:08

Hey, thank you, Adam. This was great.