AI支出不可能永远增长：P Equity Research｜EP 167

If a hyperscaler tells you that they see demand 10 years ahead, they are just talking nonsense. They cannot predict further than 2 years. Is that true?

I believe that the memory market is cyclical and will remain so, mainly because of the very nature of these costs. People assume that hyperscaler costs are fixed, as if they can grow constantly. This is just a guess.

For this market not to be cyclical, costs would have to increase indefinitely, and I don't see that being possible. I think there will come a time when these costs will slow down.

Chinese open-source developments are maybe 6 months behind, right?

They are not that far behind, and many businesses have already started switching to Chinese models. If you're talking about slowing down model development, you're just letting China catch up to you. They're not going to slow down, right?

I think China already has a lot of the safeguards that are being talked about so much in the US. If you want to slow down, go ahead, but it will likely be risky for your business model. I don't think you can afford it.

1. P Equity Research’s Background

Great, P Equity Research. Thank you very much for joining me. I'm glad to see you here. I really like what you post on Twitter and on Substack. As we chatted a little before the recording started, congratulations on your success and your growing rating on Substack. Your posts are impressive, and I look forward to our conversation about your views on the market.

Yes, thank you. I'm also happy to discuss computing power and AI with you. It's been a pleasure growing on Substack and X, and I'm glad we had this conversation.

You're always active online, 24/7, and I appreciate that.

Yes, I'm online quite often. I have to say that I use the tweet-scheduling button a lot.

To plan your posts in advance. Maybe this is the secret to success.

Yes, this is the real secret: use scheduled tweets.

Perfect. I would like to know a little more about how you started to delve deeper into the topic of computing, because you are very active in your research and, I think, are gaining significant momentum in the investment community, especially in terms of building various data centers. I would appreciate a brief story about yourself, if you're willing to share, and how you came to equity analytics.

I started tweeting around the beginning of May. At that time, I had maybe 30 followers, so I was really small. I had this account for a long time, but I never used it. Then I wanted to start using it, and I began tweeting.

I started publishing a lot on Substack, and initially I wrote many articles about niche companies in Japan, Taiwan, and Korea. It went well, and I started to gain more recognition on X, so I started posting more and more. Around July, I really changed my approach.

I tried to become someone who just shares a lot of information. I try to have an unbiased opinion and let people decide for themselves how to interpret it. I do this, and my success has skyrocketed since then.

I have no technical knowledge at all. I learned everything on my own. I don't have an engineering degree; I have a bachelor's and master's degree in accounting. I worked in the field of public auditing, and that's actually all my experience.

I'm not an engineer, so if someone asks me what the difference is between this and that, I can't answer. I don't know that much.

I think the people I consider exceptional in their fields are always self-taught anyway. Before diving deeper into AI, I spent a lot of time in the crypto market. Before that, I worked at Tesla, and a good friend of mine told me, “You don't have to be who you were yesterday; you can learn new skills.”

If you're constantly learning, then in my opinion, you will always remain at the forefront.

Yes, exactly. I think that's one of the reasons I write so much: just learn everything you can for free, because there's no limit to what you can learn. It's always good to share information.

100%. Perhaps we should start with this, because a lot has happened in the field of data centers for AI. To a large extent, I feel like things started to accelerate with Opus 4.5 sometime around the end of 2025. Since then, we've moved more toward API-based models rather than OAuth and subscription models, at least for regular people, which is really what has driven this revenue explosion.

Maybe it started even earlier, but as a result, we have a lot of these hyperscalers spending trillions on capital investments every year. I know you've been diving very deeply into the material specifications and where those capital-investment dollars are going, so I'd like to start with a general overview of what you're tracking and how you see the state of the market today.

2. Where Hyperscaler Capex Actually Goes

When it comes to hyperscalers, I think there are 4 important areas: logic, memory, power, and networking technologies. A lot of money is now being spent on logic—GPUs and ASICs—and also on memory.

It's expected that next year, memory will account for about 50–60% of hyperscaler capital expenditures. For reference, if expected hyperscaler capital expenditures are $1.1 trillion or $1.2 trillion, you should expect $500–$700 billion to be allocated to memory alone. This amount will probably be higher, but there are many nuances here.

There is HBM, DRAM, regular DRAM, and also NAND. On the computing side, hyperscalers spend a lot of money purchasing ASICs and GPUs. We all know that Amazon and Google are doing everything they can to increase their own silicon production to wean themselves off Nvidia.

I think that brings us to the question of whether computing power remains scarce. A lot of people talk about computing: We have enough computing power, but we don't have enough energy. Some say we have quite enough energy.

I haven't heard anyone say we have enough energy, but people say we have enough of other things. We don't have enough computing power. We are limited by computing power. This is what the labs say.

There are 2 thoughts I want to discuss. There was one person I spoke to from AMD during an expert call. Before I mention that, there are 2 ideas here.

First, there is a group of people who say that there is enough computing power, but we lack the energy to connect and use these chips. This comes from a discussion Satya Nadella had on a podcast. I don't remember exactly who it was with; it seems it was SemiAnalysis, but maybe someone else.

He essentially said that we don't have ready-made pads to connect these chips. The second opinion is that labs and vendors claim there is huge demand for computing and that we are limited in capacity.

3. Is Compute Still Tight?

If you listened to Google's and Amazon's reports, they mentioned that they are limited in computing resources and that their cloud revenue would be a certain percentage higher if it weren't for that. Additionally, CLSA Research came out today stating that the gap between ASIC and GPU demand and supply is 73%, and they expect the shortage to continue until 2030.

If you look at user comments, people say they feel a lack of capacity, and rental prices confirm this, because 4–5-year-old chips are more expensive than before. These are chips that were expected to be fully depreciated or unusable.

A year ago, there was a major debate about depreciation, started by Michael Berry. Now you see that these chips last for more than 5 years. They're starting to reach the 10-year mark, aren't they?

It seems Nebulous recently released data—and correct me if I'm wrong or you have better data—that they are using chips up to 8–9 years old, extending their lifespan similarly to the H100. I think H100s are still going for around $25,000 used, which is pretty amazing.

Yes, and I believe Nvidia's Volta architecture is also still in use, even though it's almost 10 years old. These are chips that are being used beyond their natural or expected lifespan.

It's funny that last month or the month before, there was a comment about a small startup that was trying to get a small batch of chips but couldn't. When they succeeded, someone from the team brought a cake and started celebrating, because that's how important it is to get computing power now.

We've even heard this from several of our portfolio companies. They've tried to do a lot of interesting things, but without access to compute power, you're stuck.

To your point, we have to literally knock out these capacity constraints. It seems like this is what everyone has to do now. Even Elon, while building the terrafactory, says there isn't enough computing power for both the robots and, I think, the cars they want to build—not to mention future space data centers.

It seems like we're generally limited by computational resources, but as you noted, is it more to do with energy or with actual physical capacity?

Energy is a big problem. I would say that the main difficulty now is the numerous delays in the construction of data centers. We don't know exactly how many data centers are delayed. Some analysts say it's not a big problem, but from all the news I'm reading, it seems like there are quite a few delays.

There is also significant public and political resistance to building data centers. I think energy is a serious problem. Energy is probably the worst bottleneck right now.

Expanding on the opinion of an AMD expert I spoke with, he believes that from the standpoint of the chips themselves and computing technology, we have enough of them. The limitations lie in other areas: data-center power, advanced packaging, and memory. Memory is the main problem.

I think we'll see a lot of debate about whether we have enough GPUs or not. However, rental prices indicate that computing power is currently in short supply.

Yes, I think I saw this morning that you posted it. It was something like B200 or B300; the price per hour of GPU work went from $5 to $7 or $8. Yes. Madness. The price continues to rise.

4. Memory’s Share of the AI Bill

Maybe it's worth delving into the topic of memory, because I've spent a lot of time on it, and we've discussed it with Bubble Boy. Especially now, when data centers are largely switching from model training to inference, which is more memory-dependent. The weight of the models is getting bigger, and the contexts are also getting longer in general.

As you rightly noted, it seems that various kinds of neoclouds, or even just the demand for memory, are growing significantly. Do you have a more detailed breakdown of this 60% memory share? That's a pretty crazy statistic—trillions of dollars literally being buried in the ground every year.

Yes, it's actually very difficult to estimate how much capital expenditure goes into memory, so let me explain the situation. First, everyone has different numbers depending on whom you ask. I don't think anyone knows exactly how much is spent on memory. For example, CLSA Research stated a few months ago that the share of memory in capital expenditures will be 48% in 2027. SemiAnalysis predicted 36%, and Citrine predicted 60% by 2028.

I don't know what data they have for 2027. JPMorgan called for 49% in 2027 and 60% in 2028. UBS says 63% in 2027, and later they pointed to 73%. So, do you understand what I mean? We can't be sure who's right, and I don't think we'll ever get an exact number because the mix consists of a lot of HBM, NAND, and commodity DRAM. We have a pretty wide range depending on whom you ask.

In general, no one will argue that memory is expensive, right? This will be a very large expense. UBS estimates that $900 billion will be spent on memory alone next year. To put that into perspective, this is actually more than all the capital expenditures made by all hyperscalers this year.

This is madness. That's a lot of money, right?

Yes. This is truly madness.

That's a huge amount of money when it comes to memory, and I think right now, in terms of supply shortages, HBM is probably the scarcest. Or maybe it's commodity DRAM—it depends on the situation.

Do you have any distinction in your research, or in the research of others that you follow, regarding the structure of memory allocation? It seems that so far—and Bubble Boy and I have discussed this on a general level—the industry has not focused on maximizing throughput, which makes sense given how they are paid, for example, for tokens. This is very reminiscent of Groq and Cerebras, as well as technologies like high-bandwidth memory, where weights and contexts are ultimately stored.

But now more and more tasks are being shifted to NAND as agent systems with longer runtimes emerge, performing increasingly complex tasks. What have you noticed from research that would indicate other types of memory allocation, if you found anything at all?

I haven't actually seen any information regarding this. I know that memory vendors are trying to make HBM a larger share of their revenue. From an economic perspective, they want to make HBM an important part of their revenue because they believe in its long-term stability.

Commodity DRAM is a commodity. The majority of their DRAM sales still come from non-HBM, and they probably believe that in the long run it is more price-sensitive. HBM is a more specialized product. It can be considered something that is not an ordinary commodity. So, in that sense, they believe they can make HBM margins higher than DRAM margins in the long run.

I think the trend now is that memory manufacturers want to depend less on conventional DRAM. They want to sell more HBM.

So, it's a transition from conventional DRAM for home computers and RAM to data centers that consume more high-bandwidth memory and are less sensitive to price.

Yes, yes, exactly. SK hynix is expected to have about 35% of its DRAM revenue tied to HBM. This is almost twice as much as it was this year, and this is expected by 2030. It may not seem like much, but I think every dollar counts for these memory manufacturers because they know they are operating in a cyclical industry with ups and downs.

5. Why the Memory Cycle Isn’t Dead

What is your personal opinion? I mean, they're trading at relatively low P/E ratios right now because they were more cyclical in the past. The famous last words are, “This time it's different,” but do you think this industry will remain cyclical, in your opinion or based on your research?

Or do you think that with the development of AI, the growth of gigawatts of power, and the greater need for memory for these models, the cyclicality will become less pronounced?

No, I still think it's a boom-bust cycle, but the downturn phase can be smoothed out a bit with long-term agreements, or LTAs, so it's a very good question indeed. I've been asked this before, and I've stated publicly that I believe the memory market is cyclical and will remain so, mainly due to the nature of the costs.

People believe that hyperscaler costs are fixed and can continue to increase. This is an assumption. For it to stop being cyclical, I believe costs must increase infinitely, which I don't expect. I think there will come a time when spending will slow down.

6. Inside a Long-Term Agreement

But speaking of the boom-and-bust cycle, I would say there are 4 points when it comes to long-term deals that memory manufacturers make. The CEO of SanDisk explained it quite simply. They do what they call new business models, or NBMs, although they should just call them LTAs—long-term agreements. I'm not sure why they call it NBM, but it has to sound fancy.

Yes, I think that would sound elegant. They want to be unique.

But there are 4 parts here. These are time commitments, volume commitments, pricing, and financial guarantees. I think most people are familiar with the pricing part. But when it comes to term commitments, it's usually 4 to 5 years. I think LTAs are usually concluded for 5 years now.

I don't think many people know this, but 2 days ago, Dyson Securities reported that Samsung had started receiving inquiries for 10-year contracts. So, yes, this is effectively a doubling of the original LTA term. And they have already started to extend contracts with some of their clients. They mentioned this before.

They are essentially saying that if the customer needs larger volumes, they can extend the LTA. They can change the conditions. So that maybe turns 5 years into 7 or something. But the length of the contract ultimately depends on the provider, and it wouldn't be surprising if it were 10 years.

I think even the CEO of SanDisk, at the September 9 event, said there might be a point where they get 10-year visibility into demand, but they're not there yet. Emphasis on the word “yet,” because that's exactly what he said. But I don't think that's possible because a 10-year contract seems negative for both sides.

You don't want to tie a client down for 10 years if there's a high probability of a cyclical downturn, right? No one has visibility 10 years ahead. If a hyperscaler tells you they have 10 years of demand visibility, they're just fooling you. They can't make predictions beyond 2 years. Isn't that right? So, there is a lot of instability in a contract that lasts 10 years.

But there is a downside to signing a 10-year contract: you lose the opportunity to take advantage of higher prices. So, this goes to the time commitment aspect.

In terms of capital expenditure, perhaps with hyperscalers, initially they funded a lot of it from their balance sheets and just from the excess cash they had from building a profitable business. I think right now, if you look at the free cash flow of all of them, they're either going to zero or they're borrowing money, and I think a lot of them have to continue to finance that.

So, is that your base-case scenario—that because revenues may be a little behind their upfront capital expenditures, spending will slow down?

I think every major technology has a period where it gets a lot of investment, and then it starts to slow down, because the argument for endless spending would be that the AI industry would never mature, right? I believe that every industry matures at some point. Whether it's PCs or smartphones, you have costs, but then you reach a point where the industry becomes mature, so you don't need to spend as much.

You just need to maintain a certain amount of capital expenditure, whether that's by cutting costs or simply keeping them at a certain level for the next few years. So, I believe that's what will happen in the future. I don't think this is something where we're going to see capital spending constantly increase.

Of course, this would change if all hyperscalers were to generate huge amounts of free cash flow in the coming years. It is expected that this could happen in 2028, but we will see.

So, it seems you're not a fan of the AI supercycle, where we get recursive self-improvement and all that. AI agents are going to take over everything.

I believe in AI, but I think we have to see what level of ROI they can achieve.

Yes, that's a big question because right now they're just draining free cash flow. So, until the free cash flow situation changes, people will continue to ask questions.

But speaking of memory, I talked about volumes, prices, and financial guarantees. Regarding volume commitments, there is a certain amount of product reserved for hyperscalers that they are obligated to purchase under long-term agreements.

So, I believe that SK hynix has 50% to 70% of its supplies locked in long-term contracts. And then, as for the price, they all essentially have a certain price minimum attached. If you look at SanDisk, they recently talked about 80% margins by the end of 2030.

This is the price floor mechanism they have, but there is also the caveat that they can cancel the price ceiling. One of the Micron employees I spoke with actually said that the company could raise its DRAM margins to 94–96% if it wanted to. Which is madness. Yes, that’s right. This is crazy, isn’t it? You’ve never heard of a business with such high margins. This is effectively selling their DRAM for next to nothing.

This is unlikely to happen, in his opinion and in mine, but the nuance is that you have a price floor and there may not be a price ceiling. So memory manufacturers can just keep raising the price if they want. In general, there are long-term agreements that provide for a fixed capacity of new supply that is put into operation, and this varies depending on the company.

On the other hand, you have some free volume that is not contracted, and it potentially has an even higher margin simply because you have not signed long-term agreements. It works on a first-come, first-served basis—or you pay more—so they can potentially charge even higher markups than under long-term agreements. Yes, they could set higher margins, but obviously, the reason they want to do more long-term deals is because they know it’s a cyclical business.

They just want to make sure their demand is stable. They make a fixed amount of money over the next few years, regardless of what happens with the hyperscaler capex, and I’ll get to that in a bit. The last part is financial guarantees, right? Micron says they have all these strategic customer agreements that are also long-term contracts, and they mentioned receiving $22 billion in upfront payments.

So there is some kind of penalty, whether it’s a prepayment or a cancellation penalty. I think that’s where the cycle looks different, and that’s why I say the business cycle will soften. It won’t be like before, when margins became catastrophic during recessions and memory manufacturers had to close.

The amount of spending on large language models is unprecedented, isn’t it? On top of that, you can make the argument about what I call capital maintenance costs. Let me explain this. Imagine that tomorrow hyperscalers announce that they have overestimated their capex capacity for the coming quarters and are going to reduce it.

Personally, I don’t believe AI spending will reach a point where hyperscalers will cut capex by 50%—maybe 30%, right? In fiscal year 2023, Oracle, Google, Microsoft, Meta, and Amazon together spent about $150 billion. If you cut next fiscal year’s capital spending, estimated at $1 trillion, by 50%, it would still be $500 billion. That would be 3–4 times more than the capital expenditure in 2023.

Do you foresee a situation where they cut costs by 50%? But even if they do, it’ll still be a lot of money, right?

I think we’re at a point where even if hyperscalers reduce costs, memory manufacturers will still have good margins, and a new average will be established. Yes. Personally, I find this unlikely.

When I look at OpenRouter and see the number of tokens generated, whether it’s open source or closed source, it seems to be only increasing. Of course, token usage is different from revenue, but the consumption of tokens itself, even with things like Agent, Muse, Hermes, and Grokbot, seems to me to be the next frontier.

That’s especially true with the use of computers by regular people who may not be as deep in computer science or software engineering as we are. We started creating agents for accounting or operations, handling basic tasks. It seems like it’s still in its early stages, and that’s what excites me.

Yeah, I mean, the demand for memory is going to be huge, right? The only thing is to make sure the hyperscalers don’t get angry at you.

7. What Happens If Customers Cancel?

I think you’ll start to see Micron—I think Micron reports next week, on the 30th—start to comment that margins will be stable from that point forward. The margins won’t be higher because they’ll have more long-term deals, so it’s similar to the comments about SanDisk. But yes, I think the memory market is still cyclical, although the situation will not be as bad as before.

I would like to add that the same specialist I spoke with who worked at AMD also worked at Samsung. This guy is a kind of all-rounder. He worked at Samsung, at AMD, and at Renesas, and he still works at Renesas.

He worked in their memory division and actually said that long-term deals are overrated because he believes that contracts can simply be quietly canceled by both parties behind closed doors with no penalties. He says, “You know, during the last recession, we had long-term deals with some big customers.”

“When demand dropped, we had to continue to supply guaranteed volumes. But if you force them to do it, companies start stockpiling your products for years. So hyperscalers can just take your DRAM and ship it to a warehouse.”

“After the contract expires, they no longer need your memory because they have accumulated it in warehouses. Accordingly, your income falls even more rapidly.” His argument is this: “If I force a customer to take 50 billion products that they no longer need, then my income will simply drop to zero over time.”

Do you understand? Not only does the share price split in half, but demand simply evaporates. Besides, you are ruining the relationship with the client. This is his view on the situation: if we reach a point where AI spending is unaffordable, there is a chance that such deals will simply be silently canceled.

That makes sense. I hope this doesn’t happen. The show continues. People continue to pay for AI services, or even through advertising, which I don’t think we’ve fully utilized yet. Rather, through a freemium model.

I’ve seen the statistics on X, and I haven’t verified them, so be skeptical about them. It turns out that only 2–3% of the world’s population actually pays for AI, which seems to be a good guideline. Given that few people have actually worked with these tools—or, if they have, only with the simplest things in the past—it’s really driven by a lot of exceptional users with extreme requests.

Yes. Yes. This is certainly the case. You also spend a lot of time, judging from how I’ve been following you, on in-depth analysis of a wide range of materials, which I find quite interesting. Given the trillions of dollars in capital investment, it’s important to understand where exactly those resources are going.

You obviously mentioned the computing side and memory. What else has caught your attention or is growing significantly in terms of overall value and costs?

8. The Rising Cost of the AI Buildout

Well, if you look at the list of materials for Vera Rubin, the components that showed triple-digit growth were ABF substrate boards and memory. Therefore, I believe that ABF substrates should be given special attention.

There is not enough capacity to produce them. The supply shortage of ABF substrates is likely to continue beyond 2028. Some people seem to have mentioned 2030. Even if you read the earnings reports of companies like Dell, HP, Nvidia, and Broadcom, which reported a few weeks ago, they mentioned DRAM, NAND, ABF substrates, and wafers in the list of restrictions.

So they all agree that the substrates are a problem. I think there is also reason to believe that indium phosphide is in short supply, because many companies, such as AXT Incorporated, Sumitomo Electric, and JX Advanced Metals, are expanding their supply of indium phosphide, which is needed for lasers and optics.

But yes, I believe that printed circuit boards and ABF substrates are 2 areas where there is clearly a supply crisis, as can be seen from the list of materials.

9. Copper vs Optics

Another thing I’m interested in hearing your thoughts on is photonics, because I think they’re trying to disaggregate different types of racks to do different tasks. If you have high enough bandwidth, you can potentially do some pretty interesting things.

I know that many people follow the stocks of companies in the photonics sector, such as Lumentum and others. How do you view the networking side in general, specifically rack interconnects or the elements that go into solutions like NVLink?

I don’t have many specifics about the technology, but when it comes to optics, there is a debate: copper or optics, right? What will win, and what will eventually disappear? There are fears that copper won’t last long, but companies are still finding ways to extend its lifespan. I think copper will coexist with optics for a long time to come.

This won’t happen as quickly as everyone thinks, because co-packaged optics, or CPO, still has many issues with yield and heat dissipation. So, in 2027, we will probably see the mass use of near-package optics, and already in 2028–2029, the growth of CPO will begin.

This is exactly what companies like Ayar Labs, Lumentum, and others are talking about. However, the level of implementation will be low. There will be growth, but we will see the dominance of CPO perhaps only after 2030.

This is largely due to the fact that it’s a new technology. It still needs to be tested and improved. Besides, CPO is expensive, right? As memory becomes more expensive, hyperscalers are likely to be more deliberate in their approaches to the network architecture of their racks.

I know some hyperscalers still emphasize the importance of using copper for as long as possible. They are in no hurry to switch to optics as soon as it becomes possible. They want to use copper as much as possible because they know how much memory spending is expected next year and in the near future.

Therefore, they need to save money on switching to optics. This is how they extend the life of copper. In my opinion, in the coming years, the environment will be hybrid—a combination of copper and optics.

At the same time, optics remains the ultimate goal. In the end, optics will win over copper. We just don’t know when.

This will likely happen between 2030 and 2040. Speed matters, right? When you try to increase speed over copper cables, overheating issues arise. When you strive for higher speeds and the wires are crowded together, crosstalk begins. This is a big drawback of copper, and you don't have to be an expert to know that nothing travels faster than light, right? So, based on this, optics will ultimately win because speed matters.

Yes, I'm extremely interested in the optical side and the different trade-offs, even with the things you mentioned, like heat or copper. I think with light you usually need repeaters, which potentially consume more energy, and right now there's a lot of pressure in data centers, because of the power issues you mentioned, to reduce overall energy consumption. I think that's another reason why people in general want to stay on copper as long as possible before switching to optics. But, as you rightly point out, optics are much faster; when data goes through glass, it's much faster.

Yes. We've already seen this happen, for example, with dial-up internet, the transition to broadband, and then to fiber optics. Of course, fiber, which is what most households are running today, is much better than dial-up internet.

Oh yes. I have fiber internet from AT&T. It is much faster than the modems I used before. Interesting. I think one of the main questions for me, and you touched on it, is how many new gigawatts will be able to be connected and whether there will be enough capacity to actually launch them.

I think SpaceX has about 2 or 2.5 gigawatts in Memphis, which I believe is the largest single data center in the world. They have a few of those there. But they continue to scale, and it seems like during their last earnings call they mentioned that they were targeting a total capacity of around 8–10 gigawatts, which should be interesting. But how many new gigawatts can be put into operation in total?

Not only in terms of building and creating a data center, but also the connectivity in terms of power supply, because a lot of things depend on that as the main revenue driver for token generation using the equipment.

10. Power and Gas Turbines

Yeah, personally I've written a lot about gigawatts, but I don't even track the pace of construction and the number of gigawatts that are being added. But I know that SemiAnalysis seems to say that there will be another 14 gigawatts added next year from advanced labs alone. I really haven't been tracking it, so I can't give an exact number.

11. US Models and Chinese Open Source

This is also something that is constantly changing, because every day you hear about delays in the construction of data centers or some kind of moratorium, so it is becoming increasingly difficult to keep track of gigawatts. Perhaps this is a digression from the topic, but what are your thoughts on the slowdown in progress at the frontier?

Well, I think that's largely nonsense. I think it's just an excuse to make things sound better, but it's a little annoying that they're sounding the alarm, so to speak. Largely because it is impossible to simply plan these data centers when the capital investment is already in the land.

If you were really worried about the consistency or, let's say, the safety of AI, I don't think you would slow down or even stop the construction of data centers. These are multi-year contracts, and the centers themselves usually take years to build. So it's hard for me to imagine that they're really slowing down the introduction of data center capacity.

On the contrary, as you said, we see that they are only increasing these indicators. Whether they will be able to launch them is another question, but their actions do not look like slowing down.

Yes, it doesn't seem logical to me either to talk about a slowdown, because if there were a technological gap of 2 or 3 years between the US and Chinese models, then it would be possible. But Chinese open source is only 6 months behind, right? They are not that far behind, and many companies have already started switching to Chinese models.

So when it comes to slowing down model development, you're just letting China catch up to you. And they're not going to slow down, are they? I think that China probably already has many of the safeguards that the US is talking about. If you want to slow down, then go ahead, but it will likely be risky for your business model.

I don't think you can afford it, because you're just inviting more competitors. You will let China catch up with you, and they will capture a much larger share of the market.

But as Jensen said today, and I think there's a YouTube video about it, he said, "If there's a problem with AI safety, then just stop the model, right? Just don't let it go. You don't need government intervention. You don't need government regulation."

There are many cybersecurity and data-breach laws that can be used to hold these labs accountable. You don't need any government policy. It's no different from when hackers break into a system and infiltrate it. You hold them accountable. You can hold these labs accountable under the same laws. But I think all this talk about humanity going extinct by 2030 is nonsense.

Yes, this is madness. I think I saw Mark Zuckerberg comment on this too, saying they had Muse ready a few months before, but they just didn't release it because they wanted to address a few security issues. It's not like they went around and convinced me that Muse was unsafe. They just fixed the problems on the technical side and then released it when they felt confident.

Yes. Muse is great. I haven't used it yet, but I've heard a lot of good reviews. I think I need to start using it. Are there any agents you are currently using?

12. Agents and NAND

I don't have any agents. I do the same as you, because you always keep up with the posts. You're the second person to ask me if I use agents, but no, I don't use agents. Everything is done by hand. I use some models. Personally, I like using Gemini.

I switch between quite a few of them. I started on the agent side with Hermes, where you could switch the underlying model, but it's quite complicated. I would say this is more for advanced users. If you want more customization, they definitely allow it.

It's good that you can switch when, for example, Anthropic overtakes ChatGPT: you can change your model, or when ChatGPT releases Astra, you can switch back, and vice versa. So that part was good, but I switched to Grok bot when Elon and xAI announced it. It was much simpler, and I could do a lot of different workflows that I already had. Now you're tied to this specific model.

I don't know if Muse said what model they're using under the hood. I assume these are Facebook's internal models. But to me, this is all very interesting, which brings us back to the question of memory. As you ask these models to perform increasingly long-term tasks, where in the memory hierarchy, so to speak, do you store it?

I think it's unlikely that it will be something like SRAM, given the capacity, or perhaps high-bandwidth memory as we integrate more chips. But SanDisk and the broader group of NAND manufacturers look interesting as they move more long-term tasks to flash memory.

Yes, I think NAND is probably the option where you'll see a lot more memory usage due to the large number of agents being used. Many brokerage reports, like KB Securities, say that if you have significantly more agents that are distributed, then you are likely to see higher CPU and GPU loads, and that will lead to higher demand for HBM, DDR5, and NAND flash, right?

So I think overall this has a very positive impact on memory demand. I think everything will be used more regardless. I don't think there will be one thing that will be used more than another, although perhaps flash memory will be used more often than others. But I'm not very familiar with this area, so I can't add anything more.

Yes. It's just interesting because I feel like agents are really becoming a broader frontier of everyday use. I hope so. At least, things are moving in that direction, instead of people writing code. But we'll see.

Of course, it's interesting. You're thinking perhaps less about this year and more about 2027 and 2028. What have you explored in this regard, or what has come to your mind as you look ahead a few years?

I think about the discussion you had regarding HBF, right? HBF will be talked about in a year or two. I think it will become more commercialized and used more often. I found Bubble Boy's comments about HBF interesting, where he says he's heard that Chinese customers want HBF, although I personally don't understand why they need HBF.

I wouldn't doubt Bubble Boy's comments, because he has more experience and connections than I do. But the problem with HBF is that it has low endurance, right? So it doesn't look like a replacement for HBM, and CXMT is still working on developing HBM3. Why do you need HBF?

13. Where HBF Fits

As far as I know, HBF is placed next to the GPUs, which get hot, right? So where does endurance come from? You won't get any endurance there at all. It's just a double whammy. This is the worst-case scenario, especially when we know that Chinese chips, like Huawei chips, have overheating problems. The CEO of DeepSeek also mentioned this.

So I wonder how you're going to solve the overheating problem with HBF when these Chinese chips already have difficulties with it.

I won't go into too much technical detail because I'm not qualified to do so, but there are 2 reliable sources that point to the role of HBF. Starting with one I mentioned in a conversation with Nick Doyle from SemiAnalysis, he believes that HBF is likely to find applications where low output packet size is needed.

So this doesn't apply to hyperscalers, right?

This will likely be for smaller systems—local deployments in private companies with a small number of GPUs. There was another source: the president of Nintendo, or rather the CTO of Nintendo, during an expert call organized by Bank of America. When asked about HBF, he agreed that HBM would likely coexist with HBF because of all the aforementioned issues.

No one knows yet if HBF will be effective enough, as there are questions about overheating, endurance, throughput, and cost structure. He mentioned using some other substrate, the name of which I don't remember. We'll see how this technology develops, but I think HBF is promising, because in the semiconductor industry, you can never be sure of anything. Everything can change overnight.

I started to dig a little deeper into different memory architectures because, as you rightly pointed out, a significant portion of the capital expenditure is going there, especially in the private sector of the market.

And if you are developing something in the field of memory accelerators, please contact our Frictionless Cap Auto. We will be happy to chat. Shameless advertising, but my point about High-Bandwidth Flash is that as models get bigger and bigger, going from, say, 1 trillion parameters—I think the latest version of Grok 4.7 had 2.1 or 2.3 trillion parameters. model 4.6 seems to have had 1.5 trillion parameters, and Elon hinted that over time, the total number of Grok parameters could reach 4, 6, 10, or even 100 trillion.

Where will all that scale be placed? Maybe we can come up with something. Bubble Boy and I have already discussed it on a general level. Maybe we'll find some interesting algorithmic way of compression so that we don't have to store all the model weights, using mixtures of experts or other methods.

But if not, we will have to load all the weights either into high-bandwidth memory or high-speed flash memory. I think this could be an ideal solution because, as you mentioned regarding endurance, it's well suited for reads, not writes, like a random-access KB cache. The model weights can be kept in high-speed flash memory, or you can do some preliminary calculations, as Logan has already discussed.

This is something he does quite consistently in his operations. You spend resources on pre-filling, and it might be worth keeping the result in high-bandwidth flash memory so you don't have to recalculate it every time. I think there are unique opportunities here.

Of course, it's still questionable. It's still early, considering it's not even in production, but I like that people are starting to experiment with different specialized memory accelerators, like we did with training. This is an interesting area to watch because we are largely shifting computing power, or building data centers, from pre-training—which will continue—to inference and decoding, and that changes the hardware.

You know more about this than I do. I would like to add more, but I don't go into these things too much, so I can't. It's cool to dive into this. High-speed flash memory is interesting.

Are there any other things that generally interest or fascinate you? You're spending all your time on this. You're always up to date. You really have your finger on the pulse. What excites you personally?

14. What P Is Most Excited About

I'm looking forward to seeing how the optics situation will play out, because I think optics is still very fascinating. Over the years, there will be such a huge demand for networks. I think networks are the fastest-growing segment of capital expenditure right now. That's a 64% annual growth rate over the next 5 years.

The development of networks will be interesting because ultimately, speed will matter for all of these models. They want to achieve the highest speed, so there will be a lot of competition over who can supply the best lasers. We have Lumentum, Broadcom, and Coherent. I think Lumentum is a great company. They have great lasers.

From the people I've spoken to, Lumentum seems to have the best technology compared with Coherent. Even Rational Analysis says that Coherent has bad lasers and so on. The optics are fascinating.

I think energy is also exciting, but the problem is that it's probably not the area worth investing in right now, mainly because the lead times are very long. For example, if you look at gas turbines, they were interesting to invest in maybe 2 years ago. Now, probably not, because all the companies—Mitsubishi, Siemens, and GE Vernova, especially GE Vernova—have orders that extend beyond 2030.

They already have limited capacity and can't take many more orders. Even if they could, the deadlines are too long. If you order a gas turbine today, you will probably receive it in 2030. That's how long the lead time is.

I saw Elon on a podcast talking about blades and nozzles, as he called them, in relation to gas turbines. A lot is moving toward over-the-counter systems, which is also interesting.

When I was doing the research, I was thinking: How many new gigawatts are being put into operation? Going back to the SemiAnalysis report on how many gigawatts they expect to commission, many of these players are only putting in megawatts. I thought there was a huge gap between what these players are talking about in terms of total power and energy and the scale we need to achieve at the gigawatt level.

Personally, I didn't focus much on the energy sector. With that said, you obviously need energy to power these chips, otherwise none of it makes much sense.

The main obstacle right now is energy. Referring to Michael Dell's comment, it seems that was at one of the events at Goldman Sachs or perhaps Citi. This was a few weeks ago, and he was talking about how he sees new cloud services as a way to measure the real demand for computing power, because they are the ones that get resources like land, energy, and infrastructure.

He believes that whoever has the opportunity to get all of this has the best idea of demand, because through them, they can see how much computing power is needed and how many servers are needed. Energy is probably the biggest bottleneck right now, along with memory. Isn't that right?

In my opinion, nothing compares to gas turbines right now because of the long lead times and high market concentration.

That's interesting. Of course, we need to continue building these data centers. I think I saw—I don't remember exactly if it was CoreWeave or Nebulas—that they were going to add a few more gigawatts.

Hopefully, everyone will do their part: a gigawatt here, a gigawatt there, and we can achieve our goals. Everyone has to contribute. No more delays.

15. CXMT and China’s Memory Industry

I know you mentioned CXMT to me before, right? Memory from China. If you want to talk about it, we can discuss it.

A lot of people are talking about the memory market being flooded with Chinese products because they're just going to produce cheap memory and all that. I don't think that will happen anytime soon. People see China as a safety valve for every new technology when it comes to semiconductor equipment, chips, or whatever.

As for CXMT memory, according to one estimate, its share of the global DRAM market will reach about 15% by 2028, and its share of the domestic market will reach about 20%. But the real problem right now is that even if CXMT wanted to flood the market, it can't, because it has to satisfy its own domestic demand.

This is the same problem faced by players from the United States and Korea. If you look at the DRAM sector, according to Bernstein, CXMT will be 3–4 years behind in bit density. In addition, CXMT is trying to release HBM3, and it is 1–2 years behind there because of the low yield of usable crystals.

This is something the company is trying to fully launch into production, but the yield is about 25%, which is quite low. In addition, its monthly wafer-launch volume is only 200,000–300,000. This is unlikely to change anything given the existing high demand.

Therefore, if CXMT wants to saturate the market, it will probably have to increase capacity significantly. It simply can't do that now. Its capabilities are quite limited.

Last night, I was at a local event talking about the development of AI, and someone suggested that I pay attention to CXMT. I shared similar thoughts: I think demand is generally so high that any new capacity will be absorbed because we are in a deficit. I don't think they will massively reduce margins when there is an opportunity to make money in the market. So yes, I agree that it's exaggerated.

In the long term, maybe a few years from now, I wouldn't underestimate China, because it is great at organizing mass production. Jensen also said at the All-In Summit that China is really good at producing products in large volumes.

The situation may change in 2 or 3 years. It all depends on the circumstances. In the long term, I think China can get 20% to 25% of the global memory market share.

Where are they now?

I can't remember the exact number, but they're not even close to 20%–25%. If I checked now—I think Counterpoint Research published this—they have 10% at the moment.

But this is in monetary terms.

A little more than twice.

Yes, that's right. More than twice.

The peculiarity of the memory market is that it is very volatile. If you look back in history, in 1975, about 95% of the market was controlled by the United States.

And 75% of that belonged to Intel alone. Then Japan came along, and they increased their share to 85%, while the US share fell to 2% in 1990. So, all of their market share was lost.

Then, sometime in the mid-1980s, South Korea came along, and they forced Japan out of the market completely. Now they have 62% of the market share. So, if the story is true, CXMT will naturally take a significant stake from SK, Samsung, and Micron, but we just don't know when or how much.

One of the ways they're trying to slow down CXMT's growth to the greatest extent possible is to simply stop exporting the best semiconductor equipment to them, right? So, no EUV scanners from ASML, no Applied Materials equipment, and no Lam Research equipment. I think there are both good and bad reasons for this, which I don't want to get into, but it leads to the Chinese government supporting so many of its domestic companies to accelerate localization.

So let me give you some statistics. The market share of local suppliers in etching was 30% in 2020; now it is about 40%. Deposition was 20%; now it is 40%. Implantation was about 5%; now it is 15%. Therefore, they have significantly improved the localization of equipment and their market share in this area.

They are simply trying to get rid of their dependence on American manufacturers like Applied Materials and Lam Research, and also on Japanese and Korean companies like Kokusai Electric, right?

The same can be seen with NVIDIA. It was, in my opinion, poorly managed export-control policies on both sides, because there was a point when Latnick was talking about how China was dependent on NVIDIA chips. Then China imposed an additional ban on their chips. So they effectively pushed NVIDIA out of the market, where it held a 90% share, and now Chinese chips will occupy 80% of the market in 2028.

This is a huge turnaround. You have effectively blocked the largest company in the world from accessing your market, and you're going to rely on your own chips. Is it because China doesn't want these chips? Probably not at all, right? There have been many reports of smuggling through Taiwan.

There have also been reports that companies like ByteDance want more chips, but there is a quota on how many chips the government allows them to buy. I think it's around 100,000 H100s or something like that now.

That seems like a lot.

Yes, that seems like a lot, but 100,000 units of the H100 is about $6 billion in sales, by my calculations. So it's not really that many chips. They're trying to get them, but they can't because the Chinese government is essentially forcing them to build everything themselves.

I think the plan for memory is the same, because China remains the second-largest country after India, and they consume about 30% of the world's memory just for PCs and smartphones. So I believe that 20–25% is quite achievable from a self-sufficiency perspective.

I think there will come a point when CXMT and YMTC become big enough to not depend so much on third-party players. That doesn't require advanced technology, right? If you're going to install memory in PCs and smartphones, you don't need HBM. You can use regular RAM. It is quite enough.

So I believe this is where CXMT could become a real threat over the years: they can capture a large market share solely through domestic demand.

Interesting. Very interesting. Wonderful.

16. Closing Thoughts

I appreciate you coming to the podcast and sharing all your knowledge and research results. I've been following you and trying to figure out the topic, and I'm amazed at how quickly you figure everything out and how you keep your finger on the pulse.

Thank you very much for coming to the podcast, sharing your observations on the market, and discussing what you expect in the future. Thank you again.

Yes, thank you for inviting me. It was a good discussion.