I originally planned to write about Micron’s history.

But if the goal is to forecast the future from the past, it’s more accurate to dissect the memory industry’s chicken-game mechanics than to trace a single company’s timeline.

This piece isn’t about who wins, but what it takes to win.

As my final post of the year, I want to start with the most important question.

Will there be another chicken game in the future?

I can’t say “no” with absolute certainty, but I want to argue that the probability is extremely low.

1. In the classic sense of a memory chicken game—one that ends with a major incumbent exiting—there are hardly any targets left.

Even if Samsung wanted to wage another memory “chicken game,” the field has narrowed to just three players, so shrinking the player set any further is close to impossible.

In particular, the moment there are any signs of an attempt to start a chicken game, regulators around the world will be watching closely due to antitrust concerns.

2. Semiconductors as strategic assets

Semiconductors have become a strategic industry watched closely by governments worldwide, so this is no longer a market where companies can invest and compete entirely on their own.

In particular, Samsung driving SK hynix into bankruptcy would not be tolerated by the Korean government—and it’s unclear whether it would even be possible.

And if Samsung tried to go after Micron, it would almost certainly invite retaliation from the U.S. government.

3. The remaining possibility—however slim

Still, there is a fingernail-sized chance that a chicken game could return—this time against Chinese memory semiconductor companies.

That said, it’s hard to argue this is likely. Chinese memory makers are the product of massive state backing, with tens of billions of dollars in subsidies and the full weight of the Chinese government behind them. Could Samsung win a power struggle against the Chinese government? I don’t think so.

Alright—before we dive into the history in earnest, let me first explain what it takes to survive a memory chicken game.

Survival Rule #1: Get bigger

For example:

If you spend $100 million to build a fab and produce 100 million chips, the cost per chip is $1.

If you produce 200 million chips? It becomes $0.50.

In other words, the moment you cut production, the cost per chip spikes.

In the end, to survive in an industry like this, you have to push utilization to the absolute limit and drive costs down as much as possible.

And this is exactly when Samsung makes an extremely bold move.

In the late 1990s, when DRAM prices collapsed, most companies hunkered down and argued that they should cut output—but Samsung Electronics moved in the opposite direction. As a chaebol, Samsung could pull cash from other affiliates and held massive liquidity, so during the downturn it aggressively increased capex and even raised its production targets—an unusually decisive stance.

Samsung, which had already become the world’s No. 1 memory maker in 1993, understood that if it expanded output further, it would gain an overwhelming advantage over its competitors.

Survival Rule #2: Spend like crazy on R&D

In the DRAM industry, shrinking the process node—reducing the nanometer width—to survive leads to an increase in supply, as it allows for a higher net die per wafer. (This is largely a narrative of the past, as the logic has become less applicable in today’s 10nm-class era.)

This, in turn, triggers a vicious cycle that exerts downward pressure on prices. Conversely, companies that fail to scale their processes produce lower volumes, which translates to a higher burden of fixed cost absorption.

Survival Rule #3: Hold an enormous cash war chest

People often describe the memory semiconductor business as one where you make a ton of money during upcycles—and then give a ton of it back during downcycles.

The problem is that if you don’t have enough cash to survive the downturn, the company goes bankrupt. On the other hand, if you have an enormous cash war chest, a downturn becomes the moment you’re holding the knife that can cut down your competitors.

Samsung, as a chaebol, could pull massive funding from its financial affiliates. I think a big reason Samsung has been able to defend its memory throne to this day is precisely because it’s a chaebol.

The introduction ran a bit long. Now, let’s get into the main body.

From 2007, Taiwanese companies began aggressively expanding supply, which triggered the first “chicken game.” Starting in 2008, this overlapped with the severe demand slump caused by the U.S. financial crisis, and DRAM prices collapsed from $6.8 three years earlier to $0.5.

At the time, there were around 10 companies in the market. Market share was roughly Samsung Electronics 30%, Hynix 19%, Japan’s Elpida 15%, U.S. Micron 11%, Germany’s Qimonda 10%, Taiwan’s Powerchip 5%, and Taiwan’s Nanya 4%.

The chicken game that began in 2007 was, surprisingly, initiated first by Taiwanese companies.

It may feel quite surprising that smaller Taiwanese players struck first, but in fact, they had their own rational outlook.

It was the alliance.

Back then, Taiwanese firms lacked independent technological capabilities. Instead, they pursued a coalition strategy: [Japanese/German technology + Taiwanese manufacturing capacity].

Structure:

Technology providers: Elpida (Japan) and Qimonda (Germany) provide leading-edge process technology.

Manufacturing: Taiwanese companies such as Nanya, Inotera, Powerchip, and ProMOS keep the fabs running.

This is how they calculated it: “Samsung is alone, but we are united. If we combine Japan and Germany’s technological edge with our low-cost manufacturing, we can encircle Samsung with volume and bleed it dry.”

But reality was ruthless. When the crisis hit, the “alliance” crumbled like sand. As the technology providers (Elpida and Qimonda) began to falter, the Taiwanese fabs that depended on them saw process upgrades come to a halt. Meanwhile, Samsung—an IDM (integrated device manufacturer)—was able to move fast, with rapid decision-making.

Beyond this, Taiwan made two major misjudgments.

First, the Taiwanese government’s “Two Trillion Twin Star” policy.

I think this may have been the biggest reason. At the time, the Taiwanese government pushed the “Two Trillion Twin Star” project, aiming to grow semiconductors (DRAM) and displays (LCD) as the nation’s two core pillars.

The government’s money line: the Taiwanese government poured massive subsidies, tax benefits, and low-interest loans into DRAM makers.

Taiwanese companies thought like this: “Even if we run losses, the government has our back, so we won’t die. If we flood the market and drive prices down, firms in other countries without government support will die first.”

But, unfortunately, as we all know, industries that receive government subsidies usually do not end well. Instead of building self-sustaining competitiveness, they only grew in size, and when the financial crisis hit, losses piled up to a level that even the Taiwanese government could not bear. In the end, Taiwan’s DRAM industry was left as nothing but an empty shell.

Also, they failed badly in their market forecasting.

Microsoft’s “Windows Vista,” released in 2007, was expected to be the savior of the PC industry.

Memory makers imagined it like this: “Vista is going to consume an enormous amount of memory. If PCs around the world upgrade to Vista, DRAM demand will explode. We need to expand our fabs now so we can rake in money.”

But as always, MSFT is terrible at operating systems. Vista turned out to be the worst OS. It was heavy and full of bugs, so consumers turned away. Demand didn’t rise, yet DRAM kept pouring out of the Taiwanese fabs they had expanded. Oversupply + the financial crisis = a massive price collapse—the worst-case scenario brought to life.

Even Samsung Electronics posted massive losses at the worst point of the cycle (4Q08–1Q09), with an operating margin around -30%. For the other players, survival was uncertain without government support—let alone the ability to keep investing in technology.

3Q08: Samsung Electronics was the only company to remain profitable with $200 million in operating profit, while Hynix -$400 million (OPM -28%), Micron -$450 million (OPM -35%), Powerchip -$530 million (OPM -100%), and Nanya -$300 million (OPM -69%) all posted losses.

4Q08: Samsung Electronics recorded -$900 million (OPM -14%), Elpida -$930 million (OPM ??), and Nanya OPM -105%.

Ultimately, Qimonda—then the No. 5 player with roughly 10% market share—filed for bankruptcy in January 2009. Its cumulative losses from 3Q07 to 4Q08 exceeded $3 billion, and despite receiving $500 million in support from the German government, it still ended up going bankrupt.

Japan’s Elpida also came close to collapse due to massive losses, but managed to barely survive after receiving $240 million in Japanese government support and $90 million in creditor support in February 2009.

Smaller Taiwanese players suffered even more severe losses relative to their size, but they managed to hang on thanks to KRW 2.8 trillion in Taiwanese government support and backing from their parent companies.

Hynix held up relatively well at the time because it ranked #2 in both market share and technological capabilities. However, the scale of its losses was so large that there were talks about selling the company to another conglomerate or even to Micron. Its stock price collapsed from around KRW 40,000 to the KRW 5,000 range—an eightfold drop.

With a massive operating loss (OPM -53%), its BPS was cut in half, and its PBR fell from about 2x to 0.5x. But after surviving the first chicken game and the global financial crisis, Hynix fully recovered its share price during the upturn that followed.

After the first chicken game, as the economy recovered, the DRAM industry briefly entered a boom period.

However, in 2010, aggressive investment ramp-ups in Japan triggered a second chicken game. DRAM prices quickly fell from $3 to $1 (based on DDR3 1Gb).

A notable point is that this chicken game started with Japanese companies first.

As with the first chicken game, it wasn’t the #1 or #2 players that kicked it off—what’s worth paying attention to is that a relatively smaller player made the first move.

Of course, the Japanese firms had their own reasons for doing so.

There are three main reasons.

1. They tried to break through the “curse of a strong yen (Endaka)” with volume.

The most fundamental cause was the exchange rate. At the time, the yen surged to an all-time high (USD/JPY in the 70–80 range).

Back then, Korean players (Samsung, Hynix) could still make a profit even selling a single chip thanks to a weak won, but Elpida was structurally in the red even if it did nothing, because of the strong yen.

So Elpida began to think: “Because of FX, the loss per chip is huge. Then we need to increase production volume (Q) to an extreme degree and sharply lower the fixed cost per chip (depreciation). That’s the only way we can match Samsung’s pricing.”

Following that logic, they forced their fabs to run at full capacity and flooded the market with supply—but this ended up triggering global oversupply and ultimately backfired on them.

The second reason is, once again, similar to the case of the Taiwanese companies.

Namely, the massive amount of public funds poured into Elpida.

Right after the 2009 global financial crisis, the Japanese government injected 30 billion yen in public funds and 100 billion yen in loans into Elpida in order to save its pride—the semiconductor industry.

Once it had taken government money, Elpida’s management had to quickly show “visible results (market share expansion, a technological leap)”. There was political and managerial pressure that left them no choice but to invest aggressively rather than stay defensive.

The critical misjudgment was that instead of concentrating this funding on technology R&D to strengthen the fundamentals, they poured it into expanding capacity (Capa, production capability) to protect their near-term market share. In the end, it was as if they simply used Japanese taxpayers’ money to fatten the wallets of semiconductor equipment companies.

The third reason was the shift toward mobile DRAM.

Seeing the smartphone market opening up at the time, Elpida tried to restructure itself from PC DRAM toward mobile DRAM.

The strategy was: “PC DRAM is a red ocean. Let’s increase the share of mobile DRAM (supplying iPhones, etc.), which we do well.” The intention was good. Even looking back now, it was truly forward-looking. The problem, however, was yield.

As the process transition to mobile DRAM progressed more slowly than expected, a flood of defective or off-grade products came out, and by dumping them into the PC spot market at rock-bottom prices, they further fueled the plunge in DRAM prices.

In the end, Elpida judged that if it stayed still, it would inevitably die—crushed by the strong yen and a lack of cost competitiveness—so it borrowed government money and placed a gamble to win through “economies of scale.”

But Samsung Electronics had already stabilized its 40nm-class advanced process by this time and pushed its costs even lower, and it comfortably absorbed Elpida’s volume offensive (a counter-fire strategy), ultimately driving Elpida into bankruptcy (2012).

Ironically, Micron ended up reaping the greatest rewards from the memory “chicken game.”

When you interview Micron engineers, there’s a line you hear again and again:

“The whole company is obsessed with efficiency.”

Let me explain in more detail.

Inside Micron, an extreme cost-cutting culture—born out of crisis—has taken deep root across the organization. For example, they simplified manufacturing processes as much as possible, minimized die size to pack more chips onto each silicon wafer, and pushed every other step of the flow to run faster and process higher volume with maximum efficiency.

And they had one more hidden card: Micron’s headquarters wasn’t in Silicon Valley—it was in Idaho. That meant lower employee churn, and dramatically cheaper land and electricity.

All of these factors combined allowed Micron, especially in its early days, to compete head-on on price with Japanese memory semiconductor companies.

In the late 1990s, when an extremely severe downturn hit the entire memory market, Samsung waged a chicken game, and the market ended up flooded with DRAM. Prices collapsed due to massive oversupply, and because of Samsung, the oversupply period dragged on far longer. On top of that, every company was in a situation where even if they sold chips, they were barely breaking even—or taking losses. Because of the chicken game Samsung Electronics ignited back then, countless semiconductor companies fell into financial distress.

But interestingly, Micron’s playbook was completely different from Samsung’s. Instead of fighting an all-out war like Samsung, Micron chose a strategy of buying companies that had fallen into financial trouble. In other words, they decided to use the moment to increase their weight class. First, in 1998, when Texas Instruments gave up its memory business, Micron acquired all of TI’s memory fabs and instantly grew into a core player in the memory industry. Then in 2002, Micron even attempted to acquire Hynix in Korea. That deal fell through because Hynix rejected it, but if the acquisition had succeeded, it was almost certain Micron would have surpassed Samsung in scale and taken the No. 1 market share position at the time.

Micron continued to push the same playbook afterward. As the Taiwan-driven memory chicken game pushed the industry into crisis, Germany’s Qimonda fell into a liquidity crunch in 2008—and Micron moved quickly, snapping up Qimonda’s 35.5% stake in Taiwan’s Inotera. Likewise, in 2012, Micron acquired Japan’s Elpida after it went bankrupt. You could call this a true masterstroke. The total acquisition price was 200 billion yen, but the actual cash Micron paid was only 60 billion yen. The remaining 140 billion yen was on an interest-free installment basis—so it was close to a conditional, almost-free acquisition. And at the time, Micron was strong in PC DRAM, but it lacked mobile DRAM capabilities. But who was Elpida? A major supplier of mobile DRAM to Apple. So that acquisition meant Micron caught two huge rabbits at once: it entered the mobile DRAM market overnight, and at the same time it moved into Apple’s supply chain.

Micron didn’t stop there. In 2016, it acquired the remaining stake in Inotera and turned it into a fully owned 100% subsidiary. So what happened then? First, by fully integrating Taiwan’s process operations into Micron’s global R&D network, it could maximize manufacturing efficiency. It also secured collaboration routes into Taiwan’s semiconductor ecosystem. In other words, rather than Samsung’s push-forward expansion, Micron pursued a fairly conservative expansion strategy—placing precise bets when opportunities appeared. And as this strategy worked, Micron’s weight class kept growing, and it ultimately became part of the memory Big 3 alongside Samsung and Hynix.

Micron’s conservative DNA also showed up very clearly when ASML’s EUV lithography tools emerged. In the late 2010s, semiconductor scaling became far more complex than ever before, and ASML’s tools were essentially monopolizing the world’s fists. The problem was that EUV tools were unimaginably expensive, and because it was a new technology, early yields were low and it took quite a long time to stabilize the process. At that time, Samsung Electronics and SK Hynix boldly adopted EUV, but Micron instead delayed adoption. They chose to push existing technologies to the absolute limit in order to lower the cost structure. This directly connects to Micron’s inherent character. From its earliest days, this company has been almost obsessive about cost reduction and extreme efficiency, and that conservative investment discipline has continued all the way to today.

The Survivor’s Peace: Why the War is Likely Over

In the end, what was the result of the blood-soaked memory chicken game that lasted for decades? The Warring States era—when dozens of players ran rampant—is over, and the world has now been divided into a “Big 3”: Samsung, SK hynix, and Micron.

What’s interesting is that each of these three survivors endured in a completely different way.

Samsung: Survived by crushing opponents with overwhelming capital strength and preemptive capacity investment (The Conqueror)

SK hynix: Survived by holding on to the very end through bone-grinding technological development, even while facing a life-or-death crisis under creditor control (The Fighter)

Micron: Survived by growing its size through ruthless efficiency and an M&A strategy of absorbing fallen enemies (The Hunter)

The market that these three companies—each with different DNA—have created by surviving is completely different from the past. When there were ten players, even if I cut output, someone else could simply increase theirs. But now that only three remain, everyone knows all too well that if anyone recklessly expands supply, everyone goes down together.

In other words, the market paradigm has shifted from a “market share (M/S) war” to “profit maximization.” This is the strongest structural basis for my argument that “the probability of another large-scale chicken game happening again in the future is extremely low.”

And there’s a particular anecdote I want to share.

1978, the year Micron was born, was the worst possible timing to start a semiconductor business.

This was the era when Japanese companies were flooding the market with cheaper and better DRAM, pulling ahead of American firms across the board. It was a time when most U.S. semiconductor companies were bleeding heavily year after year.

Back then, Micron’s founders, Joe Parkinson and Ward Parkinson, didn’t go looking for Silicon Valley billionaires—they went to Jack Simplot, the richest man in Idaho, nicknamed the “Potato King.”

Jack Simplot didn’t know much about technology, but he grasped one fundamental truth that cuts through the memory industry.

“Memory is a commodity.”

He saw through the situation: with Japanese companies pouring DRAM into the market, prices would crash to the floor. Having spent his entire life dealing with potatoes, he knew exactly what that meant—this was the bottom of the market, the moment when prices would fall as far as they could and competitors would collapse, and therefore the right time to enter.

So he boldly invested millions of dollars into the Parkinson brothers—and succeeded in multiplying it by tens, even hundreds of times.

However, I want to argue that this “memory is a commodity” truth is gradually breaking down.

The birth of HBM is a perfect example.

HBM, for now, is still a product you can manufacture and pile up as inventory.

But we’re entering an era where customization is added to the base die, foundry nodes are used, and the specs demanded by each customer diverge.

From that point on, HBM from DRAM manufacturers starts to resemble a foundry business.

The expansion of cHBM (custom HBM) is inevitable, and this will ultimately spread the effect of signing LTAs across the DRAM industry.

Right now, DRAM prices are exploding, and HBM prices are being raised accordingly. Later on, even if commodity DRAM rides a cycle and prices rise, it will become difficult to raise HBM prices in the same way—but in exchange, you gain guaranteed sales.

Both an escape from the commoditization of memory, and the “foundry-ization” of memory, will happen. Within the next two years.

Some people also point out that even an LTA can be terminated mid-stream.

If that happens, then we simply cut cHBM output. Just like TSMC reduces wafer starts when iPhones sell less.

Of course, I’m not claiming that the cycle of the entire DRAM market will disappear. I just want to argue that going forward, it will inevitably be less affected by the cycle.

Shouldn’t we stop valuing memory companies by P/B now?

Summary:

1. In 2025 Q4, FII’s GB200 and GB300 rack shipments each account for ~50%, and monthly capacity for both is 1,700–1,800 units. Capacity is expected to peak in May–June 2026, reaching 8,000–10,000 units/month. Next year’s GB300 demand is ~100,000 racks.

2. Incremental capacity mainly comes from North America and Mexico plants, and is expected to be released gradually around mid-next year.

3. The Rubin series is planned for small-batch production in 2026 Q3, with potential delays; next year market demand will still be dominated by Blackwell.

Read more

Before we begin,

* This content is not investment advice.

* Some parts of this information have been cross-checked.

* This expert call transcript is excerpted from AceCamp (you can think of AceCamp as the Chinese version of AlphaSense).

Q: Did the ODM utilization rate for GB300 recently reach the expected 80% level? Is there an update on delivery expectations?

A: Currently, the utilization rate for GB300 is basically above 80%, with some manufacturers even reaching 85% to nearly 90%. This high utilization is mainly due to the smooth transition from GB200 to GB300 production lines, as many product structures and production experiences could be directly reused. Regarding shipment volume, since NVIDIA’s supply has not yet been fully released, GB300 deliveries in October were less than 600 racks. By November, the supply exceeded 1,000 racks. It is expected to expand to over 1,200 racks in December. NVIDIA originally projected that at full load production, it could reach 5,000 racks per month. Ideally, it is expected to approach this level by late 2Q26. According to manufacturer statistics, the total shipment volume of GB200 and GB300 in 2026 is expected to be over 60,000 racks. Note: This figure is inferred based on production capacity and does not directly come from customer demand. Customer demand is currently still quite vague, with many variables involved.

Q: What are the considerations for some customers preferring to switch their original GB300 orders to B300?

A: Recently, many foreign cloud vendors found that the overall delivery timeline for new data centers deployed for GB300 has been delayed. Additionally, the base price of GB300 is high, and customers have not yet completed the tuning and optimization of large-scale clusters. B300 allows for rapid deployment for inference business needs and significantly lowers the requirements for data center construction. Many manufacturers place more importance on B300, but B300 has just started delivery, so its progress is still slower than GB300.

Q: Update on the number of GB series racks received by major customers by November?

A: Microsoft has currently received about 100 GB300 racks, with a total rack deployment of about 5,000, making it the largest customer by volume currently. Most racks have been allocated to OpenAI, and OpenAI is primarily using these resources. Microsoft’s own data centers and European partner data centers are also deploying GB200 and GB300, but progress in European data centers is slow. The original plan was to deploy 1,000 racks by the end of the year, but it now seems unlikely to be completed on time; however, orders have been placed, and ODM manufacturers have started production. Microsoft’s overall demand this year is about 6,500–7,000 racks. Amazon originally planned to deploy over 2,000 racks but has currently only completed 1,300; the remainder is not yet online, and they are accelerating the delivery progress. Google has a small data center with over 1,000 racks planned for deployment this year. Meta has had about 800–900 racks delivered. Oracle just received over 1,000 racks at its Texas data center. CoreWeave also has close to 1,000 racks. Other customers are generally at the scale of a few hundred racks.

Q: Is the investment of approximately $30 billion per GW for GB300 clusters the level for a brand-new site?

A: Yes, especially in North America. This includes purchasing or leasing land, building a brand-new data center, equipping it according to liquid cooling standards, rewiring electricity, and deploying transformers and power supply lines. If retrofitting an existing site, the cost is expected to be controlled around $25 billion.

Q: Has the recent increase in ODM utilization for the GB series affected the demand and pace of third-party TPU deployment by OpenAI, Meta, etc.?

A: Third-party user deployment of TPUs has not yet unfolded on a large scale, so there is no significant impact in the short term. It is expected that the delivery times for both will likely overlap in the second half of 2026. Initially, GB300 may have an advantage in terms of user experience and consistency risks, but later, as customers gradually complete testing on TPUs and more business models begin to be deployed on TPUs, the cost-performance advantage for inference will gradually emerge.

Q: Meta is discussing a multi-billion dollar TPU order, expected to be deployed in its own data centers in 2027? What are the corresponding TPU quantity and price?

A: It is expected to be used mainly for Meta’s future inference business needs and possibly for the training of some Llama models. The price for Meta’s TPU procurement is preliminarily estimated to be between $14,000 and $16,000. The official quote is higher, but there may be discounts for large-volume purchases. Meta currently has no plans to deploy Google TPUs in its own data centers because its data centers are not expected to be completed until 2026, and resources are limited; existing data centers mainly use NVIDIA products.

The section behind the paywall includes NeoCloud’s TPU leasing status, the Rubin timeline, the progress of HBM4, and the CSPs’ in-house ARM CPU developments.

Read more

Welcome, readers.

It’s been a while since I’ve returned with a subscribers-only post.

Yesterday, Fubon published the following note, and here’s what it said:

“INCREASING COWOS NUMBERS. Sherman writes in 2026, based on our CoWoS model, we calculate and estimate total TPU production to be 3.1-3.2mn. TSMC’s current AP8 is already fully loaded, and phase 1 of its new AP7 is for Apple’s WMCM while phase 2 will not be ready until late 2026. In our earlier note, we mentioned that TSMC will start to outsource some mid-to-low end CoWoS products to ASE, but the outsourcing is limited to CPU and networking chips, and all the accelerators will remain at TSMC’s AP fabs. However, **from our recent checks with TSMC’s CoWoS suppliers, we think TSMC is now becoming aggressive in expanding its CoWoS capacity in 2027**. Based on our latest checks, TSMC’s in-house CoWoS capacity will reach 120k by the end of 2026, which is for 1Q27 production, and 140k by the end of 2027, compared to our earlier forecasts of 110k and 130k, respectively. With a more aggressive expansion plan in 2027, we think TSMC will also provide more capacity support for Broadcom and MediaTek, Google’s TPU partners. Although it is too early to talk about the CoWoS capacity allocation in 2027, **our checks suggest that total TPU volume may double to 5-6mn units in 2027**, including 5mn from Broadcom and potentially 1mn from MediaTek.”

However, the expert call my Chinese partner conducted yesterday suggests that Fubon’s projections may actually be on the conservative side.

No need for a long preface — let’s get straight to the point.

Read more

Before I begin the main text, I will first explain what “copper foil” is.

Copper foil is a thin copper sheet made by rolling, and it can be broadly divided into two categories: battery copper foil and circuit copper foil for IT substrates. Battery copper foil is a thin copper sheet attached to the battery anode material,

and circuit copper foil serves as the copper circuit layer that forms PCB circuits. Circuit copper foil can be classified by substrate type: general-purpose rigid, mobile HDI,

FPCB, MLB substrates, and semiconductor packaging substrates, in that order of increasing added value.

Battery copper foil is an industry with over 100% excess supply. This is because Chinese companies have all rushed to expand capacity.

In contrast, circuit copper foil for substrates is a market where supply is limited to around five major manufacturers outside of China. In particular, for circuit copper foil used in semiconductor packaging substrates and server/network equipment substrates, Japan’s Mitsui is in effect a monopolistic supplier. It is not easy to convert production lines between battery copper foil and circuit copper foil. This is not a market where Chinese players can freely switch lines and ramp up supply just because it is profitable. While circuit copper foil is currently suffering from an acute shortage, if substrate demand increases, the shortage of circuit copper foil could become prolonged. Following glass fiber used in substrates, key materials such as gold plating and prepreg are also rising in price. For substrate/material categories, supply–demand is gradually tightening in tandem with rising costs.

1. The MLB (multilayer board) circuit copper foil market is expected to grow rapidly from 15,000 tons in 2025 to 31,000 tons in 2028 and 54,000 tons in 2030. For circuit copper foil makers, orders for 2026 have already started to exceed their installed capacity. Moreover, the ASP of MLB (AI server and network equipment) circuit copper foil is more than three times higher than that of battery copper foil and general-purpose mainboard substrate copper foil, which is positive.

2. Furthermore, orders for circuit copper foil used in semiconductor packaging substrates are also increasing. Some substrate companies have already begun to face difficulties in securing circuit copper foil from the fourth quarter. The ASP of packaging substrate circuit copper foil is a premium market that is more than 10 times higher than that of conventional general-purpose copper foil. (Note: The above ASPs are based on T-Value excluding copper cost.)

   ---

Outlook for Circuit Foil: Reducing Dependence on Chinese and Japanese Materials

The steep improvement in circuit foil supply–demand dynamics stems from the demand within the substrate value chain to “de-China” its sourcing. Recently, Solus Advanced Materials decided to sell its European circuit foil plant to Chinese capital. This has accelerated the trend of decoupling copper foil from China. In addition, in high-end circuit foil, Japan’s Mitsui commands a market share exceeding 90%. Semiconductor and substrate manufacturers are understood to be reviewing supply chain diversification measures to gradually lower their dependence on Japan’s Mitsui.

Korean circuit foil companies are currently working with Panasonic, which leads Japanese CCL, as well as Doosan, a Korean CCL player. They have also begun sample testing with two Taiwanese CCL companies and are reportedly in discussions to supply circuit foil to all of the global top-four CCL players. Looking ahead, Korean-made circuit foil is highly likely to be supplied to the next-generation AI accelerator models of North American AI semiconductor companies. Signs of change are also emerging in packaging substrates, and Korean semiconductor substrate makers have begun to consider diversifying their circuit foil procurement channels. It is particularly encouraging that development of circuit foil for ultra-fine pattern substrates used in high-layer-count FCBGA is underway domestically.

IT Investment Strategy: Circuit Foil as a Beneficiary of the AI Shortage Theme

On top of AI momentum, it is necessary to select components/materials that are expected to face shortages. For the copper foil industry, AI substrates are a more important growth driver than ESS. Copper foil for ESS batteries is not significantly different from that for EVs; production lines can be shared. This implies that copper foil for ESS is also a market with oversupply. By contrast, circuit foil is a separate market, where a shortage is anticipated. We expect a Korean substrate value chain to form, spanning Korean-made circuit foil > Doosan Electronics > Isu Petasys. If a post-shortage price (T-Value) increase occurs, the pace of earnings improvement for the Korean copper foil industry will be even steeper. The related company here is Lotte Energy Materials.

Lotte Energy Materials’ stock has shown very strong performance over the past month.

This copper foil shortage is strikingly similar to the glass fiber shortage.

1. There is a dominant, absolute leader in the market (Nittobo, Mitsui).

2. The dominant players are conservative about capacity expansion.

3. Therefore, they are suffering from a severe supply shortage during this AI supercycle.

What is even more noteworthy is that the dominant players are shifting their production lines toward the highest value-added products. (In Nittobo’s case, glass fiber for AI semiconductor packages; in Mitsui’s case, copper foil for AI semiconductor packages.)

This spreads the supply shortage to non-AI semiconductor package markets, which have lower value-add than AI semiconductor packages but are still sufficiently high in value-add to attract new entrants, and enables new players that were not previously regarded as suppliers in these markets to enter. (Taiwan Glass, Lotte Energy Materials)

However, if we look at the difference between Taiwan Glass and Lotte Energy Materials, we can say that Lotte has a narrower technology gap with the leading players than Taiwan Glass does and has ample room to ramp up capacity in a short period of time.

Hello, readers.

I’m back with the first planned series, an “Investment Idea Roundup.”

To explain the background a bit, one reader pointed out that I seemed to be too heavily focused on memory, so I decided to come up with a new series.

This past week… was a nightmare for Korean tech stocks.

As Kioxia’s earnings were crushed, most Korean memory names were not spared from declines.

As I was the first in the industry to point out, I see the fall in Kioxia’s results as stemming from a misjudgment by management: Kioxia signed an LTA with Apple before this supercycle, so even though the earnings themselves were not that bad, they fell short of the market’s overly elevated expectations.

I remain bullish on AI hardware and maintain a positive view on the memory shortage.

Accordingly, I added more Korean memory stocks during this pullback.

That said, I’m growing a bit more worried about Kioxia itself. Recently, the Taiwan supply chain has been buzzing with rumors that SanDisk will end its JV with Kioxia and outsource NAND production to Taiwanese foundries to increase supply.

Considering that Kioxia has been able to compete thanks to its partnership with SanDisk… this is quite concerning from my perspective.

In this series, I will present investment ideas that can capture both structural changes and asymmetric opportunities looking ahead to 2026.

1. ASML

2. Innolight (Hong Kong)

3. AIXTRON

4. Google

5. Korean memory stocks

6. NAND photoresist

7. Glass fiber (Nittobo)

1. ASML(ADR)

ASML CEO Christophe Fouquet said,

“There is no scenario in which 2026 revenue will be below 2025.”

Just three months ago, he was cautious, saying he “could not guarantee growth,” but has now shifted to a stance of clear optimism.

I want to say that this optimism from ASML’s CEO is not at all without basis.

Let me explain why I am bullish on ASML.

1. TSMC, additional capacity expansion

   

   According to Morgan Stanley’s analysis, after Nvidia CEO Jensen Huang met CC Wei in Taiwan, he asked TSMC to increase Nvidia’s 3nm capacity, and TSMC responded by agreeing to increase its 3nm capacity by up to 20,000 wafers.

   Originally, TSMC’s 3nm capacity for 2026 was projected at 140,000–150,000 wafers per month, but this is now expected to rise to 160,000–170,000 wafers per month.
   

   What’s interesting is that this capacity increase does not directly translate into new EUV orders for ASML.
   

   Local supply-chain interviews conclude that TSMC can handle this by driving the output and utilization of its existing EUV tools to extreme levels.

   That said, here is what I want to emphasize:

   “TSMC, which is unusually conservative in executing CAPEX, chose to expand capacity.”
   

   This implies that the strength of AI demand is far higher than previously expected,

   and in turn, this is directly linked to an acceleration of the transition to High-NA EUV at the A16 node and beyond.
   

   In addition, the current approach of coping with more advanced nodes through multi-patterning on existing EUV tools is also expected to soon hit its limits, which allows us to foresee an expansion in High-NA EUV sales.
   

   In other words, this round of capacity expansion should be interpreted not as “TSMC is placing incremental orders with ASML right now,” but rather as “an eventual explosion of EUV demand at the high-end nodes is unavoidable.”

2. The delay of 3D DRAM and the inevitability of EUV adoption

   

   This is, in my view, the core pillar of the ASML bull thesis.
   

   Within the memory industry, the prevailing view used to be that 3D DRAM would be adopted at the 1c node, or at the very least at 1d.
   

   The goal was to stack 24–36 layers.
   

   However, as 6F square has continued to evolve, it now appears that 3D DRAM will not be adopted until after 2033.
   

   By that time, the layer count will need to be on the order of 90–100, whereas the 3D DRAM that each company has developed so far tops out at around 16 or 24 layers.
   

   As a result, the current consensus is to first cover 3–4 generations with 4F square in the interim, and then move on to 3D DRAM afterward.
   

   6F square and 4F square refer to the minimum cell area in DRAM. In a 6F square structure, a single cell is a rectangle of width 3F and height 2F (area = 6F square), whereas in 4F square it is a square of 2F by 2F (area = 4F square). By transitioning to 4F square, you can theoretically fit 30–50% more cells within the same chip die size. This is advantageous in terms of density, speed, and power, but it also creates the challenge of having to place the gate, channel, and capacitor on a much smaller cell area.
   

   3D DRAM, on the other hand, inevitably “relieves” EUV demand within the DRAM industry.
   

   The reason is that, just as in today’s NAND flash, it adopts a structure that stacks upward, which reduces the pressure for extreme node scaling and lowers EUV dependency per bit.
   

   Instead, the center of gravity in equipment demand shifts sharply toward etch, deposition, CMP, plasma processes, and inspection/metrology.
   

   However, because of the current technological hurdles, 3D DRAM has been pushed back to beyond 2033, and as the industry progresses in the order of 6F square → 4F square → 3D DRAM, EUV demand in the DRAM industry is expected to surge at least up through the 4F square phase.
   

   I will go so far as to make what is probably the boldest call in the industry: I see a possibility that EUV demand from the three DRAM makers will surpass that of the three foundries (TSMC, Intel, and Samsung Foundry).

   ---

   2. Innolight (HongKong)

   

   A specialist in ultra-high-speed optical transceivers (Optical Transceiver) headquartered in Shenzhen, China, and currently one of the top-tier suppliers that delivers the largest volume to global hyperscalers (especially the U.S. Big 3).

   • Main products: 400G, 800G, 1.6T optical modules

   • Main customers: AWS, Meta, Microsoft, Google

   • Market position: More than 70% of revenue comes from North American CSPs

   • Unlike Chinese local module vendors, it makes “ultra-high-speed products for global CSPs.”

   In other words, Innolight can be seen as a company whose revenue structurally grows together with the expansion of AI clusters.

Originally, the U.S. government wanted to exclude Chinese companies from the optical module space.

However, the reality is completely different.

Because: if you exclude China, the build-out of AI infrastructure comes to a halt.

• There are only a handful of companies in the world that can mass-produce 800G/1.6T optical modules quickly and cheaply

• U.S./Israeli/Taiwanese vendors lag in production capacity (scale) and cost competitiveness

• Players like Coherent and Lumentum have the technology but lack large-scale manufacturing capability

In the end, AWS, Meta, Microsoft, and Google all

cannot build out racks unless they buy high-speed optical modules from Chinese companies.

That’s why Innolight has become one of the few Chinese hardware companies that have been able to sidestep sanctions, and in return its dependence on U.S. CSPs has only increased.

I would argue that, over the medium to long term, this is not a “policy risk” but, on the contrary, a situation that becomes a moat.

Especially with Innolight set to list in Hong Kong next year, I’m very happy that I’ll finally be able to buy the king of Chinese optical modules.

I fully expect it to enjoy an H-share premium, and I want to classify Innolight as an irreplaceable Chinese company, like CATL.

3. AIXTRON (ETR: AIXA)

Power consumption of AI chips is increasing very rapidly. NVIDIA’s current Blackwell-generation GPU chips have already reached around 1.2 kW, and the power consumption of a GB200 rack-level server is about 120 kW.

The Rubin Ultra, scheduled for release in 2027, is expected to reach even three times the level of Blackwell. Rack-level server power consumption is projected to rise to around 600 kW.

By the time we reach the Feynman architecture, power is expected to increase by an additional 1.5x, exceeding 8x the current Blackwell level, and there is a high likelihood that rack-level server power consumption will reach around 1 MW (megawatt).

To solve this problem, NVIDIA has announced a new power system called “Kyber,” which adopts an 800V architecture. This dramatically increases the value of power semiconductors per system, and all the major power semiconductor companies are now competing to get into this supply chain.

At present, the value of the power semiconductors going into a single rack is about $12,000–$15,000. Taking the reference design and the increase in power demand into account, Barclays estimates that power semiconductor BoM content will increase more than sixfold by 2027, and then grow a further 50%+ by 2029, reaching roughly $140,000 per rack in today’s price terms.

However, the unfortunate reality is that there are simply too many players in power semiconductors today. Looking only at NVIDIA’s 800V partners, countless companies such as Navitas and Innoscience are already competing.

That is why I would recommend investing not in the miners working to dig up the gold, but in the company selling the pickaxes. In other words, the ASML of power semiconductors: AIXTRON.

According to Kerrisdale Capital, Aixtron’s MOCVD tools hold an overall market share of around 70%, and in several of the most attractive growth segments that figure exceeds 90%.

Notably, at a time when companies such as Infineon and TI are announcing the in-house production of 300mm GaN, AIXTRON is one of the very few vendors that offer 300mm GaN MOCVD tools.

Some worry about what happens if China localizes this equipment, but just as Kerrisdale’s supply-chain work suggests, my own interviews with industry contacts indicate that the technology gap between Chinese suppliers and AIXTRON remains substantial.

Kerrisdale Capital’s original report can be found here:

https://www.kerrisdalecap.com/investments/aixtron-aixa-gr/

4. Google

Apple’s strength lies in devices, Meta’s in advertising, Amazon’s in cloud, and Microsoft’s in Copilot and OpenAI — each has a clearly defined core competency.

By contrast, Google is the only company that simultaneously owns all of these pillars — literally the “Big Tech among Big Tech.”

And I believe that in the AI era, the company with the highest likelihood of building the most profitable AI business is also Google.

Let’s think through why.

First, Google has already secured an overwhelmingly dominant position in the market for search ad targeting.

In practice, there are very few platforms that can substitute for Google; the vast majority of global users, advertisers, and traffic sit on top of Google’s ecosystem.

Given this backdrop, Google has the highest probability of successfully integrating AI in a highly sophisticated way into its existing ad model.

The search, YouTube, Android, and Chrome ecosystems it has already built out are the largest distribution platforms in the world, and Google has the power to naturally layer new AI features and ad formats on top of them.

From the advertiser’s perspective, it’s actually very hard to refuse the new AI-based business tools Google offers.

Because, ultimately, “if you want your product to sell inside Google’s ecosystem, you have no choice but to follow Google’s algorithms.”

In other words, through a combination of its massive user base + advertiser dependence + proprietary AI infrastructure (TPU) + distribution platforms, Google is positioned to become the player with the highest margins and the strongest operating leverage in the AI era.

I also think many people are underestimating YouTube as a platform.

The more AI advances, the more AI-based creator tools will shine (subtitles, auto-cut editing, multilingual translation).

And as AI drives the cost of production in creative work toward zero, YouTube upload volume will increase in a near-linear explosion.

This, in turn, will expand ad inventory, make ads more sophisticated, and structurally grow revenue.

Lastly, there is Google Cloud.

A lot of people still treat Azure and AWS as the only truly meaningful players in the market, but I would argue that Google Cloud has a tremendous opportunity ahead.

Azure does not have its own large language model and is overly dependent on OpenAI, while OpenAI is trying to break away from Microsoft.

Meanwhile, AWS’s Trainium and Inferentia performance is frankly very poor.

Google Cloud, with Vertex AI + TPU + Gemini API, owns a vertically integrated SaaS/infra stack —

which effectively makes Google Cloud the only hyperscaler that is independent of the GPU stack.

Why do you think so many large customers — xAI, Anthropic, Meta, and others — want to keep using Google Cloud?

To sum it up in one sentence, in the AI era, the only company that combines the highest potential, the lowest cost structure, and the broadest user base distribution is Google.

The following section is behind a paywall.

Beyond the paywall, you’ll find:

• My view on the current memory supercycle

• My overall outlook on Korean memory stocks

• Some interesting insights on photoresist used in memory that I picked up from recent supply-chain checks

• Updates related to fiberglass

Read more

1. Outlook for the Display Industry
   
   

2. Outlook for Humanoid Robotics
   
   

3. Current Status and Outlook of the Memory Semiconductor Market
   
   

4. Foundry Market Outlook, including a Perspective on Mature Node Foundries
   
   

Among these, the display section and humanoid section will be summarized briefly. Instead, the memory semiconductor and foundry market sections will include detailed explanations along with attached images.

• MacBook Pro OLED is scheduled for release in Q2 2026.
  
  

• MacBook Air OLED has been delayed to 2029.
  
  

• MacBook Pro will adopt tandem OLED.
  
  

• Mass production of MacBook Pro OLED panels is planned for Q2 2026.
  
  

• Due to aggressive capacity expansion by Chinese display manufacturers, there is a risk of oversupply in OLED notebook panels.
  
  

There was also mention of Meta’s OLEDoS-based XR device, Meta Project Puffin:

• 110g
  
  

• 0.9-inch OLEDoS
  
  

• FOV < 90° (TBD)
  
  

Regarding humanoids, much of the content was quite abstract and speculative.

One particularly interesting point was that the Tesla Optimus 3rd generation has been delayed to early 2026. The reason is that the robot hand of the 3rd generation Optimus currently has a lifespan of only about six weeks, which is far from meeting commercial requirements.

And now, we will move on to the memory semiconductor section.

The growth of the DRAM market in 2026 is essentially driven by AI servers and general-purpose servers. In particular, the recovery in general-purpose servers is expected to make a meaningful contribution to DRAM demand. Also, while PC shipments are relatively modest, the notable increase in DRAM content per device deserves attention (driven by AI PCs).

Server DRAM capacity per box has effectively entered a phase of rapid acceleration.

We are now seeing a 15% increase in memory content per server.

The direction of investment is clearly shifting toward servers and GPUs.

CapEx intensity (CapEx as a percentage of revenue) is projected to rise to 22% in 2025 and 27% in 2026.

The shipment volume of general-purpose servers is expected to grow by only around 1%, which essentially indicates a state of stagnation.

I share a similar view.

When AWS began placing large orders for general-purpose server DRAM earlier this year, many interpreted it as the beginning of a long-awaited replacement cycle for general-purpose servers. However, my view has been that CSPs simply do not have the bandwidth to refresh their general-purpose fleets right now. And even if they wanted to, the recent surge in memory prices may make such upgrades unfeasible in practice.

(It’s worth noting that in general-purpose servers, DRAM accounts for nearly 10 times the share of total BoM cost compared to AI servers.)

AI server share is increasing from 13% → 17% → approaching 20%, indicating that the proportion of AI servers within the overall server market is steadily expanding each year.

As much as the market is benefiting from AI server demand, it’s also fair to say that dependence on AI servers is increasing as well.

The DRAM CapEx increase of +81% in 2025 and +14% in 2026 may appear lower than one might expect.

However, this is essentially unavoidable.

Even if equipment is ordered now, most suppliers (except Samsung) currently lack available cleanroom space to install it. In other words, they must wait for new fab construction or expansion to be completed before any meaningful ramp-up can occur.

This means that for the two DRAM suppliers other than Samsung, it will take at least another 10 months for CapEx spending to translate into actual capacity expansion.

TrendForce interprets this as Samsung not aggressively increasing its HBM allocation.

In other words, Samsung is signaling that it may continue to allocate part of its capacity to general-purpose DRAM rather than pouring capacity into HBM.

Instead of expanding new fabs for HBM, Samsung appears to be prioritizing maintaining existing production levels and focusing its CAPEX on process efficiency improvements.

Also, SK hynix’s massive capacity expansion stands out (an increase of about 60K wafers per month).

The additional 30K increase in TSV capacity suggests that SK hynix is entering the HBM4 phase with a considerable level of confidence.

Micron’s capacity increase is extremely limited.

It is likely to remain minimal until the Singapore and New York fabs come fully online.

I will also comment on CXMT.

While the company produces a large volume of DRAM, it remains significantly behind in HBM R&D, indicating that its ability to compete in the HBM market is still weak.

The growth in graphics DRAM is noticeable.

It seems a bit early for this to be driven by Rubin CPX…

Perhaps this is being influenced by the RTX Pro 6000 refresh cycle?

The slowdown in HBM demand also stands out.

However, I don’t think there is much need for concern here.

The main reason is that Blackwell Ultra and Rubin have similar HBM capacity configurations, which temporarily dampens incremental HBM demand.

This should not be interpreted as weakening demand for HBM itself.

In this table, the following can be read:

• Samsung achieved almost no growth from HBM in 2025.
  
  

• However, in 2026 it will grow in HBM by about 9–11%.
  
  

• SK hynix achieved tremendous growth in HBM in 2025.
  
  

• However, this HBM growth is expected to partially slow in 2026.
  
  

• Instead, to the extent of that slowdown and more, it will grow in general-purpose DRAM.
  
  

• Micron grew in HBM in 2025 more than SK hynix.
  
  

• Micron will grow the most steeply among the three.
  
  

• CXMT is aggressively increasing its growth rate to achieve economies of scale.
  
  

Additional comment: Existing DRAM = process miniaturization (Technology): achieving higher output per wafer (bit density) by shrinking processes in existing fabs—e.g., to 1-alpha (SK hynix) and 1-gamma (Micron)—without building new fabs.

TrendForce is harshly critical of Samsung, lol.

It seems right to say that SK hynix captured most of the gains from HBM growth in 2025.

One interesting point is that TrendForce appears to think SK hynix will also take the lead in HBM4.

As evidence, they point to SK hynix being the first of the three memory makers to provide samples.

However, I can’t help thinking they may be underestimating Samsung.

I believe Samsung will take roughly 30% of the HBM4 market next year.

I’m also hearing from inside sources that Samsung internally is very confident about HBM4’s performance.

This table contains quite a few important points.

First, TrendForce expects SK hynix to mass-produce 16-high HBM3E.

This is also the first time I’m hearing this. As far as I know, 16-high HBM3E has been at the sample level for technology demonstration.

It may be that someone on the ASIC side (perhaps Google?) requested 16-high HBM3E.

Second, TrendForce sees no delay for Micron’s HBM4.

In other words, they regard the GF report I shared on X a few days ago as not factual.

The view that HBM3E will be the best-selling HBM product even in 2026 has been around for some time.

However, in the material I just reviewed, the projected share of HBM4 within total HBM sales in 2026 was more aggressive than any sell-side estimate.

In other words, isn’t TrendForce implicitly suggesting that Rubin is being pulled forward?

DRAM is expected to sustain a higher revenue growth rate in 2026 compared to 2025.

What stands out is the projection that the market will continue to suffer from chronic supply shortages in 2026 as well.

This is because HBM wafer demand continues to eat into general-purpose DRAM wafer allocation, keeping supply persistently tight.

Really interesting material.

In the fourth quarter of 2025 alone, a price increase of roughly 30% for DRAM is expected…

This means I’m very much looking forward to the Q4 earnings of SK Hynix, Samsung, and Micron.

Another noteworthy point is that prices are projected to continue rising through the fourth quarter of 2026.

This is the longest upcycle in the history of the DRAM industry.

“DDR5 prices have already recovered to their initial launch peak, and are expected to surpass it next year.”

Normally, when a product reaches EOL, demand collapses and prices plunge as it gets replaced by newer products like DDR5.

But look at the chart. Even DDR4 (the green line), which has already been declared EOL, begins to surge sharply again starting in 2025.

Isn’t this the strongest evidence of HBM cannibalization?

Manufacturers are halting DDR4 production to allocate their limited DRAM production lines to HBM and DDR5. As a result, despite being an EOL product, supply to support legacy server demand disappears, creating the bizarre situation where even “older products” are experiencing price spikes.

Samsung is expected to join NVIDIA’s HBM3E supply chain next year, which means HBM3E prices are projected to decline by around 30%.

HBM4 will command a 50–60% price premium compared to HBM3E.

However, since the majority of HBM sold in 2026 will still be HBM3E, a decline in HBM3E ASP is unavoidable.

Jukan’s view:

A somewhat disappointing estimate.

Some sell-side houses had published reports reflecting the expectation that HBM3E prices would fall less, due to the ongoing supply shortage (there were rumors that Samsung would divert some of its general-purpose DRAM capacity to HBM3E production).

However, TrendForce seems to see it differently.

According to sources familiar with the matter, Samsung will indeed join NVIDIA’s HBM3E supply chain, but the volume will not be large. And even that comes with the condition that Samsung must buy back the Blackwell systems that use its HBM3E…

It feels like TrendForce’s projection may be overly conservative.

Jukan’s view:

HBM3e (orange line): Prices are expected to peak in 1Q25 and then begin to decline, falling nearly 40–50% by the end of 2026 compared to early 2025.

Server DDR5 (blue line): Supported by solid server demand, prices are expected to either slightly increase or remain stable.

Result: The price gap between HBM3e and DDR5 narrows sharply from 4.6x in 2Q25 to 2.0x in 4Q26.

Here’s the interesting part: HBM requires about 3–4× more wafers than DDR5.

This implies that the COGS of HBM is 3–4× higher than that of DDR5.

Combining these two points:

The price premium begins to fall below the cost premium.

In other words, starting in 2026, selling DDR5 becomes more profitable than selling HBM from a margin perspective.

The recent rumors that Micron may shift capacity from HBM back to server DRAM are likely related to this.

The share of HBM output as a percentage of total DRAM remains at 8% in 2026, the same as in 2025.

However, its revenue contribution is projected to decline to 26%.

This once again reminds us of the ASP decline for HBM3E mentioned earlier.

Alternatively, it could be interpreted as general-purpose DRAM prices rising more sharply, leading to faster revenue growth on that side.

It can also be seen as a signal that the era in which SK Hynix effectively monopolized HBM is ending, and that competition—at least to some extent—has begun.

The memory section ends here.

Let’s move on to the foundry section next.

The memory session is provided for free, while the foundry session is behind a paywall.

Read more

Today in Korea, SK Securities made headlines by setting a target price of 1 million won for SK Hynix.

Since this aligns with what I’ve been asserting for some time, I plan to process and share that report with all of you.

Before we begin, let me briefly note that SK Securities is no longer owned by the SK Group.

Therefore, I believe—and trust—that SK Securities’ decision to assign the highest target price for SK Hynix among both domestic and international brokerages was made without any external pressure.

Why was memory valuation based on PBR: high earnings volatility

The valuation methodology for memory stocks was PBR. Because downturns and upturns were stark, confidence in the long-term trajectory of earnings was low. It was a cyclical industry linked to the macroeconomy. TSMC, which shares the same industry and end markets, was different. Despite being in the same capital-intensive manufacturing industry, it continued stable growth. Confidence in medium- to long-term earnings justified the application of PER.

Why is earnings volatility so high: build-first, order-later structure

In memory, a “build first, take orders later” structure is typical. Suppliers expand capacity by forecasting future demand on their own. Since the actual volume of demand is determined by the macroeconomy, suppliers ultimately have to match the macro. That is impossible. Large-scale memory capacity additions meant impending oversupply. TSMC follows an “order first, expand later” structure. Within specialties and technological differentiation, it expands on an order basis and is less sensitive to macro slowdowns. Large-scale expansions at TSMC signaled upward revisions to forecasts.

Signal of value expansion: the PBR ceiling has begun to be breached

In the AI cycle, PBR has started to lose its explanatory power for memory share prices. SK Hynix, which has been consistently achieving differentiated profits based on HBM, has seen its PBR begin to break above its historical ceiling. EPS and PER are also rising together. Considering that PER typically fell during upswings when EPS was being revised up, this is judged to be a signal of value expansion.

From now on, PER: structural expansion of earnings stability and growth

Memory will also transform into an order-first, expand-later structure. The start of the AI scale-out cycle is expanding the memory benefit beyond HBM to memory overall, including server DRAM and SSDs. It is an even stronger cycle than the past two years. By contrast, given the lack of room for capacity additions, a flexible increase in supply will be difficult. Without stable procurement of memory, achieving AI roadmaps is impossible. The share of long-term supply contracts of about 2–3 years will now expand rapidly. In the AI cycle, memory is PER.

New paradigm, New multiple: a new path

If the industry has changed, the method of valuing companies should also change. Applying PER, the target prices presented by SK Securities are KRW 170,000 for Samsung Electronics and KRW 1,000,000 for SK Hynix.

1. Why was the memory valuation tool PBR

Cyclical: high earnings volatility

Historically, the valuation tool for memory stocks was PBR.

As the memory industry went through the early-2010s smartphone cycle, followed by the server and COVID cycles, the contrast between downturns and upturns was stark. Because a large share of costs in the memory industry are fixed, the high profit elasticity to memory price fluctuations was a double-edged sword. A Valuation based on the more stable Book value, rather than highly volatile Earnings, was effective.

It was also a typical Cyclical industry linked to the macroeconomy.

From 2010 until the pre-AI cycle, no cycle deviated from the direction of the macroeconomy. The same held for earnings and share prices. In that process, PBR explained share prices well.

But TSMC was different.

Despite the same cycles, the same end markets, and the same capital-intensive manufacturing industry, TSMC continued stable growth. High confidence in stable profit growth provided the basis for applying PER Valuation. TSMC’s share price is well explained by PER.

The same holds for the directions of PER and EPS within a cycle.

In an upcycle, while the direction of EPS increases is the same, the memory sector typically derates to a low PER even as EPS enters a full-fledged upward trend. The reason share prices still rise in this phase is that the slope of the EPS increase exceeds the slope of the PER decline. This implies that once the slope of EPS growth in the memory sector begins to slow, stock returns deteriorate.

The reason PER was low during booms has been confirmed by the subsequent sharp drops in earnings. PER failed to explain stock performance in the memory sector; if anything, buying high-PER and selling low-PER proved more effective. The cause is high earnings volatility and low stability.

The background is that memory follows a build-first, order-later structure.

In the build-first process, suppliers forecast future demand on their own and increase future supply through capacity expansion. Given that the actual volume of demand has been determined by the macroeconomy, suppliers effectively have to match the macroeconomy. This is virtually impossible.

Moreover, in periods when memory is standardized and the technological gap among vendors narrows, what remains is a battle for market share. Large-scale capacity additions mean imminent oversupply, which further amplifies industry volatility.

By contrast, TSMC follows an order-first, expand-later structure. Based on specialties and clear technological superiority, it expands capacity to respond to orders, so earnings volatility tied to the macroeconomy is low. This is generally why TSMC’s earnings and share price are insensitive to economic slowdowns, and even with large-scale expansions, the market does not worry about oversupply.

2. Signal of value expansion

The traditional upper bound of the PBR band has already begun to be pierced

PBR, which has long explained the share prices of the memory sector, is beginning to lose its function.

This is because SK Hynix’s PBR, having delivered differentiated profits based on strong competitiveness in HBM since the start of the AI cycle,

has begun to break above the historical upper band.

Considering that it had not lost that function even once since 2010,

this shows that, in the AI cycle, a clearly different expansion of value from the past is underway.

Recently, upward revisions to EPS and a rise in PER have been occurring together in the memory sector.

In the past, cases where upward EPS revisions and PER increases occurred together in the memory sector were the phase when the cycle

was beginning to turn from downturn to upturn. In this period, expectations of a resilient rebound in conditions

are reflected in share prices, so the elasticity of share-price gains exceeds the magnitude of the EPS revisions. Except for this case,

it is generally low PER in booms and high PER in downturns.

The memory sector is not currently in a phase turning from downturn to boom. Since the start of the AI

cycle in 2023, the industry began to turn around, and the boom has continued up to recently. We judge that we are at the outset where

the assessment of the memory sector within the AI cycle is beginning to change rapidly.

It is time to consider whether maintaining PBR valuation for the memory sector is appropriate.

If the basis for PBR was high earnings volatility, then as earnings stability strengthens within the AI cycle and structural

growth becomes prolonged, PBR will gradually begin to lose its rationale.

Change in valuation methodology: now PER

Change in the character of the memory industry: lower volatility, a cycle that beats the macroeconomy

Within the AI cycle, the character of the memory industry is changing fundamentally.

Since the start of the AI cycle in 2023, SK Hynix, which preempted the HBM market, has seen its earnings trend upward regardless of the macroeconomy. In 2023, this was interpreted as a base effect, as it was a phase of passing through the worst conditions represented by production cuts. However, this trend has continued for two years.

SK Securities forecasts that the structural profit-increase cycle for the memory sector within the AI cycle will continue at least until 2027. This is because the core of the AI cycle is expanding from the existing scale-up to scale-out, lifting demand, while the room for supply to rise is limited.

With CapEx discipline and HBM strength, and a lack of space available for capacity additions, the start of the scale-out cycle means a prolongation of supplier advantage. In October’s NDR, NVIDIA projected that global data-center CapEx could grow at up to a 40% annual average rate from 2025 to 2030. The AI cycle will expand from scale-up to scale-out and scale-across.

In the process, AI memory will begin to diversify not only through increased HBM adoption but also into server DRAM, GD7, SOCAMM2, and Custom HBM, and the intensity of AI benefits across the overall memory sector is expected to rise sharply.

This implies that the memory industry will shift to an order-first, expand-later structure. Achieving AI roadmaps is impossible without a stable supply of memory. Securing AI-specialized memory such as SOCAMM2 and Custom HBM is also extremely important. Amid limited capacity for supply increases, the share of long-term supply contracts of about 2–3 years to secure various AI memories stably will rise rapidly. The likelihood that memory capacity additions, as in the past, will become the trigger for a peak-out in conditions is low.

SK Securities changes the valuation methodology for the memory sector within the AI cycle to PER. This is because declining earnings volatility, enhanced long-term growth, and profit trends that beat the macroeconomy will begin to raise confidence in earnings. This also implies a structural rise in EBITDA, which is expected to lead to strengthened shareholder-return policies in the memory sector.

This will, in turn, re-justify the higher corporate value that accompanies the shift to PER.

3. New paradigm, New multiple

At the threshold of a new path

SK Securities changes the valuation methodology for Samsung Electronics and SK Hynix to PER.

Accordingly, it raises the target prices to KRW 170,000 for Samsung Electronics and KRW 1,000,000 for SK Hynix.

Applying target PERs of 15X and 11X, respectively, to the two companies’ 2026 expected EPS,

the target prices are revised up by 55% and 108%.

Since applying PER to Korean memory names is the first phase, a period of doubt about the rationality

of the target PER levels is inevitable.

However, as the AI cycle expands from scale-up to scale-out and scale-across, it is a phase in which

memory demand is being driven explosively. By contrast, the suppliers’ limited ability to increase supply

means a prolongation of a price-up-centered cycle. We judge that this is the time when the earnings

growth rate of the memory sector will begin to stand out most conspicuously among global AI-related stocks.

In the AI scale-out cycle, Samsung Electronics will generate upside in general DRAM based on a production-capacity advantage,

and SK Hynix will actively leverage synergy with general DRAM amid the continuation of its competitive edge in HBM.

Both companies can pursue profit maximization.

Considering this, the PERs of memory suppliers such as Samsung Electronics, SK Hynix, and Micron are inexpensive

relative to the PERs of TSMC and NVIDIA, the key beneficiaries in global AI.

Also, Samsung Electronics’ target PER of 15X is the upper end of its historical 12M Fwd. P/E band,

and SK Hynix’s target PER of 11X is below Micron’s 2026 expected PER of 13X, so they are not excessive.

In the AI cycle, the character of the memory industry is changing structurally.

It has been about two years since the earnings trend began to change, and the cycle is strengthening.

As confidence in a foundation of stable earnings growth rises, the methodology for valuing corporate value must also change.

SK Securities maintains its previously presented view of a memory supercycle.

We interpret the recent steep share-price gains not as a temporary phenomenon, but as the beginning of the reflection

of structural changes in the industry. Supercycle, Buy all.

In today’s premium article, we’ll cover updates on Samsung’s HBM status, the current state of glass substrate supply, and key issues related to the Samsung Galaxy S26.

Read more

1. Introduction — “Korea is a Wafer Powerhouse, but a Packaging Laggard”

Semiconductors consist of four major stages: design, fabrication, packaging, and testing.

Among them, the final stage—packaging (back-end process)—has long been regarded as simple assembly work.

Yet today, as Moore’s Law has effectively reached its limit, the importance of packaging is growing explosively.

At Computex 2025, Jensen Huang, founder and CEO of NVIDIA, stated:

“In the age of AI, advanced packaging is the most important factor.

And right now, the only company that truly excels at it is TSMC.”

His words carry more weight than mere praise.

Advanced packaging now stands at the forefront of semiconductor competitiveness—

and it is precisely in this domain that Korea has fallen far behind.

However, Korea was not behind in packaging from the very beginning.

At one time, Amkor Technology, now known as a global semiconductor packaging giant, was originally a Korean company.

Likewise, STATS ChipPAC Korea, which still maintains a notable presence today,

originated from the semiconductor assembly division of Hyundai Electronics (now SK Hynix).

(Although it has since been acquired by China’s JCET and become a foreign-owned company.)

Then why did Korea’s semiconductor packaging industry lose its own roots?

In this article, I will dissect the current state of Korea’s packaging sector

and trace the causes of its decline through history.

Read more

Today, Europe accounts for less than 10% of the global semiconductor market.

Considering that it once produced nearly half of the world’s semiconductors, it’s no exaggeration to say that Europe has fallen from grace in this industry.

Today, instead of writing another piece focused on technology or market analysis, I’d like to take a step back — to look into the past in order to understand the future.

This article examines “Why Europe Lost Semiconductors” from three perspectives:

① The absence of market mechanisms

② The limits of bureaucratic industrial policy

③ The structural gap between Europe and the U.S./Asia in ecosystem dynamics.

“Europe was once the cradle of semiconductors.”

You might not know this, but Europe was once not just a leader — it was a creator in the semiconductor industry.

Right after Bell Labs invented the transistor in 1948, research institutes in France and Germany independently developed their own versions around the same time.

In 1948, just a few months after Bell Labs invented the first transistor in 1947, two scientists in Paris — Herbert Mataré and Heinrich Welker — independently developed a similar device called the “transistron.”

Their invention was well received by the French press, and in 1949 the French government held a major press conference to announce it to the public.

By the end of 1949, Westinghouse’s French subsidiary had even built a factory capable of mass-producing the transistron.

The First Miss: However, after witnessing the atomic bombing of Hiroshima, the French government decided to focus its resources on developing nuclear technology, cutting off state funding for the communications industry and leaving it to the private sector.

• Without government support, research on the transistron gradually faded away, and the inventors eventually departed — Welker joined Siemens, while Mataré emigrated to the United States.

Time then moved into the 1950s.

The 1950s: The Loss of Technological Leadership

France struggled with the unpredictability of early semiconductor manufacturing, which led to a more conservative approach — favoring abstract, theoretical research over practical development.

The “unpredictability” here refers to the fact that semiconductor manufacturing itself was inherently difficult to control, chaotic, and required a complex mix of interdisciplinary knowledge.

At the time, early semiconductor production was extremely inefficient — the precise techniques and know-how needed to mass-produce devices like transistors reliably had not yet been established.

While Bell Labs in the United States addressed this by openly sharing its production knowledge, dispersing expertise, and leveraging collective intelligence, France approached the problem in a closed, fragmented manner within separate academic circles — and failed.

As a result, France lost its early-mover advantage and, by the late 1950s, had fallen behind even its peers in Japan and across continental Europe.

And now, as the needle of time moves into the 1960s, the integrated circuit emerges.

In the 1960s, the U.S. semiconductor industry was driven primarily by military demand—notably from the Minuteman II ICBM and the Apollo Program—with the Department of Defense purchasing nearly half of all chips produced.

Military applications demanded smaller, more power-efficient, and highly reliable devices, effectively forcing American companies to refine their technologies at an accelerated pace.

In contrast, Europe’s semiconductor demand came mainly from the consumer sector, not the military or government. In 1963, consumer demand accounted for 56% in West Germany and 35% in France, while in the U.S., it was only 5%.

Consumer markets prioritized price over performance.

European industrial giants like Philips and Siemens viewed semiconductors merely as components within finished products. They focused on incremental improvements in low-cost germanium transistor technology, believing that silicon was suitable only for high-temperature applications.

Meanwhile, American startups rapidly mass-produced silicon planar transistors, embracing disruptive innovation and widening the technological gap.

By this time, Europe was no longer just trying to catch up with the United States—it now had to defend itself from Japan, an emerging competitor.

Japan, through its consumer electronics industry, swiftly adopted transistor technology and caught up with Europe by implementing import restrictions and signing over 250 technology licensing agreements, which helped build its domestic manufacturing capacity.

France, for its part, launched the “Plan Calcul” computer development program in 1966, prioritizing the rapid creation of a national computer system — even at the expense of its domestic semiconductor industry.

In short, France—and much of Europe—was indifferent to nurturing its own semiconductor ecosystem.

In the end, European chip manufacturers failed to achieve the necessary production scale to keep up with the aggressive cost-reduction pace of their foreign competitors.

Looking back now, the 1960s can be seen as Europe’s last real opportunity to take leadership in the semiconductor industry.

However, that opportunity was lost — and the gap would only grow, permanently.

1970s: The Rise of LSI Complexity and the Widening Gap

With the introduction of Intel’s 4004 microprocessor in 1971, the era of Large-Scale Integration (LSI) began — a period marked by a dramatic increase in circuit complexity.

In this new age, enormous capital investment was required for any new entrant to succeed.

In other words, what was once a craft-like, workshop-scale industry was now transforming into one defined by advanced equipment, sophisticated materials, and highly specialized talent — an industry with a genuine economic moat.

• In the United States, computers became the primary demand driver for high-performance LSI devices, distinct from missiles or televisions.

• Europe, however, lacked not only the production capacity to manufacture such advanced LSI chips but also the strong demand base necessary to justify their development.

European computer firms were much smaller than their American counterparts, military and avionics demand was weak, and the telecommunications market was deeply fragmented.

In other words, Europe was not a single unified market, but rather a collection of divided national markets, each with its own standards and regulations.

Also, European consumer electronics companies still wanted a different kind of chip — not the cutting-edge, high-performance ones.

During this period, the United States was dominant — yet there was one semiconductor powerhouse determined to catch up: Japan.

In 1976, Japan launched the VLSI (Very Large Scale Integration) Project to counter IBM, investing hundreds of millions of dollars into a joint effort to develop next-generation technologies.

Although the project ultimately failed to achieve its goal of creating a “one-chip computer,” it played a crucial role in strengthening Japan’s semiconductor manufacturing industry — refining its technical capabilities and leading its companies to adopt CMOS technology, which would later become the global standard.

1980s: The Failure of National R&D Subsidies and Cooperation

In the 1980s, Europe’s information technology (IT) trade deficit surged dramatically — from $4.1 billion in 1980 to nearly $7 billion by 1985.

European companies tended to prefer technological alliances with U.S. firms rather than collaborating within Europe itself.

(For example, Philips partnered with AMD and TI, while Siemens allied with Intel, Toshiba, and Fujitsu.)

Alarmed by the growing IT trade deficit, European governments began launching state-led projects and subsidy programs.

France announced a five-year microelectronics plan, while West Germany and the Netherlands jointly launched the “Mega-Projekt” to support Siemens and Philips — described at the time as “Europe’s last chance to fight back against the Japanese chip invasion.”

Although Siemens successfully demonstrated a 4-megabit chip, it had to license 1-megabit DRAM technology from Toshiba, and by then Japan’s NTT had already unveiled a 16-megabit DRAM, showing just how rapidly technology was advancing.

The Mega-Projekt later evolved in 1989 into JESSI (Joint European Submicron Silicon Initiative), but it suffered from major setbacks — Philips withdrew from the memory business, and Siemens abandoned its 16-megabit memory development plans after partnering with IBM.

As a result, the program’s budget was sharply reduced, marking yet another missed opportunity for Europe’s semiconductor ambitions.

1990s: The Shift of Semiconductor Hegemony to Asia, Europe Reduced to Specialized Niche Markets

In the 1990s, the technological and financial barriers to entering the semiconductor industry became extremely high — a 250nm memory fab cost between $1 billion and $2 billion.

During this period, the companies that still dominate the global semiconductor industry today emerged — TSMC, Samsung, SK Hynix, and others.

As a result, Europe faced intensifying competition from Asia’s electronics supply chains, led by TSMC’s pioneering independent foundry model and conglomerate semiconductor giants like Samsung.

Europe lost the consumer electronics market to Asia, military and government spending declined, and the computing market remained firmly under the control of U.S. companies.

While Europe has maintained a strong presence in its remaining domains — such as industrial, telecommunications, and automotive sectors — these now account for only about 10% of the global market.

Conclusion: Europe Had Reasons to Lose Semiconductors

Europe was once the birthplace of the semiconductor, but it failed to commercialize its invention.

When government support was needed, it withdrew.

And when private-sector demand was finally needed to drive the industry forward, governments stepped in too late — trying to lead from the top down.

In short, policy timing was completely out of sync.

Europe spent decades in this cycle of hesitation and misalignment.

And now, when it declares it will double its semiconductor market share, one can’t help but ask — is that even possible?

Europe lacks both the political will of the United States to pressure Taiwan and the financial capacity of China to pour massive subsidies into its industry.

It remains indecisive, forever chasing ideals instead of execution.

I would argue that Europe’s relative economic decline today ultimately stems from losing semiconductors.

Because Europe lost semiconductors, it missed the era of the personal computer.

Because it missed the PC era, it lost the software revolution that created Microsoft.

Because it lost the software era, it never gave birth to Big Tech.

And because it never gave birth to Big Tech, its economic growth stagnated.

AMD founder Jerry Sanders once said, “Real men have fabs.”

I’d like to give it a slight twist:

“Real countries have fabs.”

OpenAI’s indicated monthly DRAM demand of 900K units is understood to be based on 2029 projections, according to sources.

As of the end of 2025, the global DRAM capacity is estimated to be around 2,000K.

That means the demand presented by OpenAI would account for roughly 45% of the total.

For the high-value DRAMs (including HBM) that OpenAI requires, only the three major DRAM manufacturers are capable of supplying them. Based on the effective capacity of these three companies, the share rises to 63%.

However, this raises one particular concern for me.

It reminds me of the memory supercycle of 2017.

Back then, Amazon (AWS) urged Samsung to build a memory fab dedicated to its needs, implying massive DRAM demand to come.

But when Samsung completed the fab in 2018–2019, Amazon never actually placed those orders, as if nothing had happened.

As a result, Samsung was left with idle capacity, and those unused fabs quickly turned into enormous depreciation burdens and heavy losses.

That’s how we fell into the memory downcycle that lasted until 2020.

My point is this: during booms, everyone insists that demand will keep soaring and pushes semiconductor makers to make huge investments.

But when the cycle turns for whatever reason, those same players simply stop ordering — and those investments translate directly into losses.

The more aggressively you invest during the boom, the deeper the losses become when the downturn hits.

A bit off-topic, but one of the members in my Discord once shared an interesting idea with me.

(If you become a Founding Member, you can join my Discord! It includes several hedge fund professionals, sell-side analysts, and people working in the semiconductor industry.)

He said: “Google is doing so well with its TPU that maybe NVIDIA and AMD won’t be able to sell shovels anymore.”

So this time, when NVIDIA invested as much as $100 billion into OpenAI, and AMD offered OpenAI equity just to establish ties, maybe both of them are desperately trying to bind themselves to OpenAI in order to fend off the threat of ASIC-based competitors.

And then another thought hit me:

Not everyone can be a winner in AI.

There are challengers like xAI, Anthropic, and OpenAI — but the crown can only sit on one head.

Eventually, someone will drop out of the race. The question is — who and when?

One thing seems certain: this AI arms race can’t last forever, and the number of people willing to buy shovels isn’t infinite.

In other words, once one of them falls out of the race, could that be when GPU and HBM prices start to collapse?

Lately, I’ve been feeling anxious about that. Like Buffett said, “Only when the tide goes out do you discover who’s been swimming naked.” Those words keep echoing in my head.

Anyway, let’s get back to the main topic.

What does the easing of the separation between industrial and financial capital mean?

President Lee Jae-myung of South Korea has instructed his administration to “review the easing of the separation between industrial and financial capital.”

He made this remark during a meeting with Sam Altman, CEO of OpenAI, on the 1st.

The separation of industrial and financial capital is a regulation introduced in 1982 to keep industrial conglomerates and financial institutions apart.

President Lee cited “securing investment capital for semiconductors” as the main reason for reviewing the policy.

But what does the separation of industrial and financial capital have to do with semiconductors?

To understand the background of President Lee’s remarks, we need to start with SK Hynix.

SK Hynix, which has risen to become a “semiconductor superstar” in the age of artificial intelligence (AI) thanks to its high-bandwidth memory (HBM), faced a severe crisis just two to three years ago — to the point where its very survival was in question. The company was hit hard by a brutal memory downturn.

In 2022, its operating profit of 6.81 trillion won plunged 45.1% from the previous year’s 12.41 trillion won. In 2023, it even recorded an operating loss of 7.73 trillion won. The situation was so dire that rumors spread throughout the industry that “SK Hynix might not survive.”

It was HBM that saved SK Hynix.

As the era of generative AI began in late 2022, demand for HBM started to surge explosively. From 2023 to 2024, SK Hynix effectively monopolized HBM orders from major customer NVIDIA, barely making it through the harsh memory downturn.

Then, in 2024, the company posted an enormous operating profit.

Yet, despite that success, SK Hynix’s management could not hide their growing sense of unease.

To handle the flood of HBM orders pouring in from customers such as NVIDIA, the company had to rapidly expand its production capacity. However, manufacturing HBM requires an enormous amount of capital — in other words, more investment spending.

Although SK Hynix’s performance began to recover in 2024, the company was still short on cash.

So the idea conceived by SK Hynix’s top leadership was to create a fund dedicated to semiconductor facility investment.

The SK Group would establish special-purpose companies (SPCs) to build semiconductor fabs or purchase equipment, and then set up funds to attract investments from major global firms such as BlackRock and Blackstone.

The profits generated through these ventures would be distributed to investors.

While it meant the company wouldn’t be able to keep all the profits, they judged it to be a better option than being unable to expand production capacity due to a lack of funds — and ultimately having to hand over HBM orders to competitors.

However, this plan was soon blocked by South Korea’s regulations on the separation of industrial and financial capital.

Under Korean law, industrial conglomerates like the SK Group are prohibited from acting as general partners (GPs) to create funds and raise external capital.

But now, with President Lee Jae-myung mentioning the potential easing of these restrictions, there’s a possibility that SK Hynix’s original proposal could finally come to life.

However, I have some concerns about this.

As I mentioned earlier, memory is a cyclical industry — and no one knows how long this boom will last.

Historically, investments made during upcycles have almost always exceeded optimal levels, which in turn led to massive losses during downturns.

Now, SK Hynix is bringing leverage into the picture.

In the past, most memory industry investments were financed directly from companies’ own cash reserves. That allowed them to weather downturns even after periods of overinvestment, since the losses weren’t catastrophic.

But things change once leverage enters the equation. The scale of investment will grow much larger — and so will the risks.

Whether this capital comes from Korean banks or Wall Street remains to be seen.

I just hope that by the time this plan materializes, around 2029, the current cycle will still be intact and everyone in the industry will be thriving.

I sincerely hope my uneasy feeling doesn’t become reality.

In the paywall section, we’ll take a more rational approach — using actual data to forecast where the memory cycle might pause, and how far it could flourish.

Read more

SK hynix has overtaken Samsung to claim the No.1 position in the global memory semiconductor market. Once on the brink of bankruptcy, Hynix has now surpassed Samsung, which had held the top spot for 33 years. Behind this dramatic reversal lies not just technology, but a deeper foundation of “organizational culture” and “trust.”

Today’s discussion is based on the book The Trust Game, written by former presidents and vice presidents of SK hynix.

SK hynix’s performance in the second quarter of 2025 clearly demonstrates its growth. While Samsung Electronics posted a total operating profit of 4.6 trillion KRW, SK hynix, with its focus on the memory business, recorded an operating profit of 9 trillion KRW. In the DRAM market as well, the tables have turned, with Samsung holding a 34% share compared to SK hynix’s 36%.

In particular, in the high-value HBM (High Bandwidth Memory) segment, SK hynix commands over 70% of the market and has exclusively secured the trust of NVIDIA, the world’s largest AI company. Samsung, on the other hand, has yet to pass NVIDIA’s quality certification (qualification).

Just looking at the current results, SK hynix may appear to be a long-standing leader, but only a little over ten years ago, the situation was quite the opposite. Facing a bankruptcy crisis in 2000, a workout program in 2001, and eventually being acquired by the SK Group in 2011, Hynix’s very survival was uncertain. At that time, Samsung was the world’s No.1 semiconductor company, holding a gap of more than 20 percentage points over Hynix.

The book explains that the driving force behind Hynix’s overtaking of Samsung lies in the “three trust games.”

The first is the “cultural trust game.” After acquiring Hynix, SK Group did not replace the existing management team wholesale but chose to trust the organization as it was. Based on the SKMS (SK Management System), it reinforced autonomy and a horizontal culture, instilling in employees the sense that “this organization trusts me.”

The second is the “strategic trust game.” Through the “Shift Left” strategy, Hynix created a structure where quality, production, planning, and all departments moved together from the early stages of development. Instead of the old method of handing projects off after development, a new approach was introduced in which everyone started together and saw the project through to the end. With stronger interdepartmental collaboration, the overall organizational capability rose exponentially.

The third is the “human resources and organizational trust game.” SK hynix moved away from a performance-only personnel system and built a structure focused on nurturing people for the long term. A CEO with a technical background served for an extended period, demonstrating leadership grounded in a deep understanding of the field, while maintaining stability without frequent leadership changes. As trust accumulated between frontline employees and leaders, decision-making also became faster.

Behind this paywall, I cover detailed anecdotes introduced in the book, as well as my selection of notable Korean materials, components, and equipment (SoC/SoP) companies worth paying attention to right now.

Read more

‘Why doesn’t Google Maps work well in Korea?’

The staff smiles and replies,

‘People in Korea rarely use Google Maps. We use Korean apps like Naver Map or Kakao Map instead.’”

This phenomenon—where foreign platforms operate only in a limited way compared to the global standard while local platforms grow overwhelmingly—is often called the “Korea Galápagos.” In the past it was even mocked as a symbol of closedness and exclusivity.

For example, Korea’s public institutions and courts still use the domestic word processor Hangul (Hwp) rather than Microsoft Excel.

Messaging is no different. As of February 2025, KakaoTalk holds a staggering 98.9% market share, an almost unheard-of monopoly globally.

In search, Naver still holds a majority. Although it has fallen from the 70–80% range of the past, it remains a rare case of a local service beating global Big Tech.

Today, however, I don’t see this “Korea Galápagos” merely as closedness. I argue that, in the coming AI era—especially in the race for sovereign AI—it can become a secret weapon only Korea possesses.

High-quality “sovereign” data is essential for true sovereign AI

Why are U.S. Big Tech firms so far ahead in developing LLMs?

Capital surely plays a part, but I’d point to the sheer volume of data in English within the total corpus.

Because American internet platforms and services are used worldwide, distributing U.S.-made LLMs globally is that much easier. Interactions with users then further improve those LLMs, creating a virtuous cycle.

However, there are signs this equation is starting to break.

With the Trump administration now in power, a growing number of U.S. allies and countries around the world believe they must not depend on the U.S. for AI, which will have profound implications for security and the economy.

Even so, I believe only a tiny number of countries can actually achieve sovereign AI.

Why? Because sovereign AI requires that a country’s local companies own that country’s data.

If they don’t—if data sovereignty has been ceded overseas—then no matter how much money is poured into building a domestic LLM, its performance will inevitably lag U.S. models. This is why a so-called data barrier is necessary.

In other words, if you want to cultivate sovereign AI now, you must, like the EU, prevent U.S. companies from taking your national data.

Demand stringent personal-information security. Force data centers to be built domestically so data cannot be taken out.

You must. If you don’t have domestic internet companies like Korea or Russia, such tough measures are necessary.

I also believe one reason China is ahead in AI LLMs is that it erected the Great Firewall early, blocking U.S. companies from accessing its national data.

Even if the original intent was to surveil citizens, in retrospect it was the right decision.

Model cases of data sovereignty from Korea

Korea’s digital ecosystem is so unique that, from the outside, it can look almost aberrant. Local monopoly platforms—not global standards—dominate nearly every aspect of daily life.

First, there’s the messenger KakaoTalk. In Korea, it’s no exaggeration to say that essentially 100% of smartphone users are on KakaoTalk. The 98.9% share as of 2025 isn’t just a number; it means people’s conversations, photos, payments, group activities, and shopping—digital records of everyday life—are all piling up within one platform. In other words, vast data on the language, culture, and behavior patterns of Korean society is aggregated in a single channel called Kakao.

Second, there’s the portal and search engine Naver. Google dominates globally, but in Korea Naver still holds a majority. Users search news, post on blogs, and ask questions on Knowledge iN. This data is not just search logs; it functions as a unique qualitative dataset reflecting the concerns, interests, and daily context of Koreans. It’s an asset on a wholly different plane from English-language data.

Third, there is local office software such as Hangul (Hwp). Public agencies, courts, and schools still use this word processor as a standard, shaping the form of administrative and legal documents and research materials. This too enables a level of data independence you don’t see elsewhere. The administrative and legal language data of Korea is preserved within the domestic ecosystem rather than flowing to overseas platforms—hugely significant.

In short, Korea is an unintended case of securing “data sovereignty.” Global Big Tech failed to put down roots, and instead, the language and life patterns of the people have been systematically accumulated inside local platforms. This data pool is not only the foundation for training Korean-language LLMs but also the core resource needed to build true sovereign AI that reflects the sociocultural context of Koreans.

Let’s now address the export of Korea’s map data to the United States, which I mentioned in passing above.

Since first applying to the Korean government in 2007 to export map data, Google has continuously pushed to take Korean map data to the U.S.

What Google wants is a fixed-scale “1:5,000” high-precision map. At that precision, 1 cm on the map corresponds to 50 m in the real world, allowing detailed representation of alleyways and buildings. It’s five times more precise than the 1:25,000 maps Google currently uses.

Interestingly, Google applied for this on February 18 of this year—less than a month after the start of the Trump administration’s second term.

The U.S. Trade Representative (USTR) has labeled Korea’s restrictions on map exports a “non-tariff barrier” and is exerting pressure.

Back in 2011 and 2016, applications were blocked largely on security grounds. This time, however, the U.S. government itself is applying direct trade pressure.

Why are Google and the U.S. government so eager for Korea’s map data?

Because exporting this data doesn’t just mean making Google Maps more convenient.

This map data can be used indefinitely. In the AI era, it’s akin to digital crude oil.

For example, a 1:5,000 ultra-precise map includes detailed information such as lanes, road boundaries, and signs—data essential for training autonomous-driving AI. In other words, Waymo could use it to gain competitiveness in autonomous driving within Korea.

That would mean Korea serves merely as a testbed for future mobility while core technologies and platforms remain dependent on the U.S.

Even the autonomous-driving technologies developed by Korea’s Hyundai or Naver could lose competitiveness within Korea itself.

Moreover, future services—delivery robots, drones, AR navigation—will all operate on fine-grained spatial data. The company that holds data down to Korea’s narrow back alleys will preempt the related service-robot and AR markets.

Therefore, exporting Korea’s map data overseas effectively hands over the very foundation of future industries to other countries. Korea is trying to stop this.

The future of a Korea-style sovereign AI

I don’t think any country other than the U.S. and China can catch up with ChatGPT/DeepSeek.

The Middle East? There’s plenty of money to buy lots of GPUs, but there are far too few developers and researchers.

Korea? It’s fairly capable in AI hardware, but it is very short on both GPUs and research talent.

That’s why I believe Korea should move toward industry-specialized AI.

Take healthcare, for example.

In reality, medical techniques and treatment methods are about 85% the same across countries.

The differences lie in the remaining 15%.

That 15% diverges by country due to differences in body type, weight, diet, and so on.

So let’s say the global commonality covers that 85%.

The real contest is in the remaining 15%.

You have to train and deliver that 15% differently, country by country.

I’d like to call this the 85% rule.

I believe this will become Korea’s export model for sovereign AI.

Thinking more broadly, sovereign AI competitiveness will be determined by that remaining 15% of local data and context.

As I said above, securing that 15% of data is ultimately what matters most.

But at this point, it’s unrealistic to build local internet platforms to take on U.S. Big Tech.

Therefore, let me again offer the policy approach I proposed above.

Even now, follow the EU’s approach.

Block U.S. Big Tech from acquiring your local data.

Demand strict information security so your data remains within your country.

That is the only way to achieve sovereign AI.

In closing…

Although Korea’s Galápagos-ization has been mocked, I believe it has become a powerful moat in today’s AI era.

Now, if anyone wants to do something AI-related in Korea, they must partner with Korean companies to secure data.

For the same reason, I believe Russia also has the potential to achieve sovereign AI.

Russia, too, has a significant number of local internet platforms and thus meets many prerequisites.

Of course, U.S. sanctions blocking imports of NVIDIA GPUs are a major bottleneck.

NVIDIA’s massive performance gains don’t come for free.

They demand a sacrifice in the form of enormous power consumption and heat generation.

Take a look at the picture above.

Taiwan’s supply chain is speculating that Rubin Ultra’s power consumption could reach as high as 3,000 watts.

To put that in perspective, that’s about three times the power draw of the Blackwell B200, an enormous leap.

And it’s not just about higher electricity bills — with greater power consumption inevitably comes massive heat generation as well.

Because of this rising heat problem, what’s getting attention these days are substrates and interposers. The reason is simple: existing interposers and substrates can no longer withstand such levels of heat and power consumption.

It’s easy to dismiss substrates and interposers as low value-added industries, but in today’s AI era, they are in fact the invisible infrastructure that supports computing performance. The way you lay out the data pathways between chips, or between chips and memory, can determine the performance of the entire system.

This is especially true for high-bandwidth memory (HBM), the heart of AI accelerators. To solve the bottleneck in HBM, interposer and substrate technologies are receiving more attention than ever.

Why is resolving the HBM bottleneck so critical? Because no matter how fast a GPU is, if a memory bottleneck occurs, training and inference speeds plummet. That’s why the industry is betting heavily on packaging technologies that can maximize HBM’s performance.

The problem, however, is that interposers and substrates are now emerging as the weakest link. In environments where tens to hundreds of GB/s of data must flow, what’s needed is not just a connector material, but a high value-added solution with optimized electrical and thermal characteristics.

This is where a new competitive dynamic comes in. For years, the industry has viewed glass substrates as the next-generation answer. With their low dielectric constant, ability to handle large-area processing, and relatively low cost, they were expected to be a perfect fit for the expansion of HBM and HPC chips. In fact, global materials suppliers and substrate makers in Korea and Japan have been pouring large investments into mass production of glass substrates.

But recently, another contender has been rapidly gaining ground: the SiC (silicon carbide) interposer. Originally known as a power semiconductor material, SiC offers high thermal conductivity and strong mechanical stability, making it well-suited for packaging next-generation AI accelerators with extreme power consumption and heat output. Especially in environments that require both 3D-stacked HBM and ultra-high-speed I/O, SiC is increasingly being touted as more reliable than glass substrates.

This leads to a key question: if the adoption of SiC interposers accelerates, can glass substrates still hold their place as the star of next-generation packaging? Or will they be pushed aside from a “future technology” to merely an “alternative” before living up to expectations? This article aims to explore that very tension — the balance of power between SiC and glass substrates, and the ripple effects that will follow.

About glass substrates

There are two types of glass substrates.

One is the glass core substrate, and the other is the glass interposer, which replaces the currently used silicon interposer with glass.

Currently, the field where players in the industry are relatively trying to enter more actively is the glass core substrate.

A glass core substrate means that the core material inside today’s substrates — which is plastic, specifically FR4 — holds the center.

What GCS aims to do is, as shown on the right, replace that central core with glass.

Then what advantages would it bring if we replace this with glass?

It is advantageous for large-area processing, meaning multiple semiconductors can be packaged at once.

It is also beneficial for miniaturization, as it allows the wiring width of the package substrate to be significantly reduced compared to today.

In fact, the reason glass substrates have received so much attention is none other than because Intel is leading the way.

For the past 50 years, Intel has been the emperor of semiconductors, and even though its dominance has been shaken in recent years, it still supplies an enormous volume of semiconductors worldwide when you look at shipment numbers.

So then, what exactly is a glass interposer?

A glass interposer is the medium that connects the chip and the substrate from the top side of the package substrate. In terms of TSMC’s 2.5D packaging technology, CoWoS (Chip-on-Wafer-on-Substrate), the ‘W’ corresponds to the glass interposer. Currently, however, this ‘W’ is made of silicon, not glass.

That said, procuring silicon interposers is a headache in semiconductor manufacturing, especially for AI semiconductors. The more CPUs, GPUs, and HBMs that must be packed into a single package, the larger the silicon interposer required. But because silicon wafers are circular, there is a physical limit to how much the area can be expanded. Furthermore, producing silicon interposers consumes foundry capacity, making it cost-inefficient.

By making the interposer from glass, which allows for large-area rectangular processing, these concerns can be resolved. For this reason, the semiconductor industry is pursuing a two-track approach: developing both glass core substrates and glass interposers.

At present, what the semiconductor industry urgently requires is glass interposers. This is because TSMC has already laid out a roadmap for large-area interposers. Currently, the largest interposer size used by TSMC is about 2,700㎟, which is 3.3 times the reticle size (858㎟²). Considering yield, only about 7 to 8 usable interposers can be cut from a single 300㎜ wafer.

TSMC does not plan to stop there. It has announced a roadmap to expand interposer sizes to six times the reticle (5,148㎟²) next year and to eight times (6,864㎟²) by 2027. This is in response to the trend of integrating CPUs and GPUs into a single package—like NVIDIA’s ‘GB200’—and the increasing number of HBMs required. Even now, there is a shortage of silicon interposers, and with such size increases, manufacturers may be able to yield only one or two good interposers per wafer.

However, the hopeful future of glass substrates has once again been thrown into turmoil.

The reason? SiC.

What is a SiC interposer?

Recently, there have been rumors circulating in the Taiwanese supply chain.

NVIDIA is said to be considering adopting a SiC interposer for the CoWoS packaging of its Rubin GPU.

The SiC interposer itself is similar to the glass interposer I mentioned earlier.

However, since SiC is the second hardest material after diamond, extremely powerful cutting equipment, such as those produced by Japan’s DISCO, is required.

So why are NVIDIA and TSMC trying to develop SiC substrates?

The first reason is thermal dissipation. The heat generated from HBM alone is enormous, and the GPU itself also produces significant heat. The required thermal specs have become so demanding that conventional silicon interposers are reaching their limits.

The second reason is its exceptional strength. Unlike fragile glass substrates, SiC is far more robust.

And this second reason is the key point. A glass substrate is, literally, glass—it is extremely fragile. This leads to high transportation costs and high defect rates.

Of course, hearing this, you might think: isn’t a SiC interposer simply a complete upgrade over glass substrates?

But the most important factor is cost.

Substrates made with SiC interposers are extremely expensive—SiC itself is inherently costly.

Industry insiders believe that NVIDIA is currently the company most actively seeking to adopt this technology. The reason lies in the nature of NVIDIA’s NVLink technology. The closer the distance between the GPU and memory, the faster the transmission speed, the higher the GPU power, and the greater the overall performance. In the future, NVIDIA even plans to stack the GPU and memory together and drive them with extremely high currents. Because silicon carbide has higher thermal conductivity than copper, it can effectively alleviate the heat generated by such large currents, which is why NVIDIA is paying close attention to it.

However, when using silicon carbide as an interposer, the requirements differ from those of conventional SiC. According to industry analyses, since foundries use SiC as an interposer material, the structural defects that the traditional SiC industry worries about do not significantly affect interposer manufacturing. The real key is cutting technology. Because SiC is as hard as diamond, insufficient cutting capability can leave the surface wavy, making it unsuitable for advanced packaging.

Since major cutting equipment manufacturers such as Japan’s DISCO are still developing next-generation laser cutting machines, NVIDIA’s first-generation Rubin GPU is expected to continue using silicon interposers. However, once the new equipment is in place, the manufacturing of SiC interposers will become much smoother. Nevertheless, because NVIDIA’s performance requirements are extremely demanding, if the heat generated inside the chip exceeds existing limits, SiC must be adopted. Hence, there is an outlook that “around the year after next, SiC will be introduced into advanced packaging.”

Conclusion: My thoughts on glass substrates and SiC interposers

I think glass substrates will find applications in Intel’s CPUs, or in mobile APs from Apple and Samsung.

However, I doubt they will be widely adopted in the AI accelerator market, which has been in the spotlight recently.

The reason is that NVIDIA currently dominates the AI accelerator space, and TSMC exclusively manufactures NVIDIA’s chips.

In other words, whatever NVIDIA and TSMC choose will become the industry standard. (The logic being that AMD and the ASIC camp will follow suit once NVIDIA makes its choice.)

That’s why I believe TSMC beginning research on applying SiC interposers to CoWoS at NVIDIA’s request is a signal that cannot be ignored.

Moreover, this is clearly a negative for glass substrates. The emergence of SiC interposers effectively blocks glass substrates from penetrating the AI accelerator segment, which they were once expected to enter.

I recommend GlobalWafers as a company worth paying attention to.

Since TSMC has abandoned SiC production, GlobalWafers is effectively the only company in Taiwan capable of supplying TSMC with SiC interposers.

In particular, GlobalWafers appears to be the only player currently researching 12-inch SiC wafers.

Some people point out that as interposers scale to larger areas, it will be difficult to respond with SiC wafers that are only around 200mm. That’s certainly a valid point. This is why the Taiwanese supply chain seems to be pursuing a two-track development strategy.

One track is the SiC interposer I mentioned above, and the other is the SiC carrier.

A SiC carrier refers to TIM1, a thermal interface material placed between the chip and the heat spreader to enhance heat dissipation.

Blackwell GPUs use graphite TIMs, and Rubin is also likely to stick with graphite TIM.

However, based on my recent supply chain research, there appears to be a possibility that the second (later) wave of Rubin Ultra GPUs could use SiC carriers.

Also, if SiC interposers are commercialized, I believe it would create additional opportunities for CMP equipment makers such as AMAT, since SiC interposers require extra planarization work.

Q: What are the evaluation results for the 690 chip, including its improvements over the previous generation 590, and its advantages and disadvantages compared to Huawei's contemporary Ascend 910 chip?
A: The 690 chip performs exceptionally well in single-card and full-system tests, although the testing scope did not cover specific cluster environments, and no specific progress comparison with the previous generation 590 chip has been mentioned. In terms of specific parameters, the 690 chip achieved a score of 813 in FP16 tests, is equipped with 196GB of HBM3 memory, and its measured interconnect performance exceeds 890 Gbps. From an overall assessment, the server version of the 690 chip demonstrates outstanding single-card performance. Compared to Huawei's contemporary Ascend 910 chip, the 690 holds certain advantages in all aspects. Notably, the HBM memory parameters of the Ascend 910C single card are relatively lower, making the 690 more competitive in memory-dependent computing scenarios.

Read more

A few days ago, Hanmi Semiconductor’s CFO, Kim Jeong-young, told the press:

“Our marketing has focused on TC bonders that can translate into revenue through mass production and shipments, and we already built a hybrid bonder back in 2020.”

“We expect to sell hybrid bonders for high-bandwidth memory (HBM) in 2027 and for systems-on-chip (SoCs) in 2028.”

This is being viewed as a rather surprising statement.

In the industry, Hanmi has generally been regarded as strong in TC bonders but at a significant disadvantage in hybrid bonders.

I suspect the remarks were intended to dispel such concerns.

What I want to highlight is Hanmi’s earlier claim.

On July 15, Chairman Kwak Dong-shin issued a press release stating that the company would develop a hybrid bonder for HBM6 with a target launch at the end of 2027. He also added that memory makers would not use the expensive hybrid bonder for HBM5 production.

However, this diverges considerably from what I’ve learned from sources.

According to a source familiar with the matter, Samsung is aiming to apply hybrid bonding starting from HBM4E 16-high, while SK hynix is targeting HBM4E 20-high.

The source also indicated that narratives claiming hybrid bonding is difficult due to contamination are not accurate.

Given this, I think the CFO’s apparent shift in stance is essentially an attempt to change the market’s view that Hanmi is disadvantaged in hybrid bonding.

Even so, I recommend a Sell position on Hanmi Semiconductor.

I will cover Hanmi Semiconductor in more detail later in this piece.

One thing many people overlook about hybrid bonding is that the hybrid-bonding tool itself is relatively easy to build.

That’s why companies that wouldn’t dare enter the thermo-compression (TC) bonder market are targeting hybrid bonding as a new revenue stream.

In that sense, the hybrid-bonding tool itself is easy to make—especially for HBM applications.

The real bottleneck is chemical mechanical planarization (CMP) equipment.

What is CMP?

In hybrid bonding, the planarity, cleanliness, and uniformity of the bonding interfaces determine yield—in other words, the surface must not be uneven.

The process that creates this flat surface is the CMP process.

If, during simultaneous planarization of copper (Cu) and the dielectric (mainly SiOₓ, with some SiCN), you fail to hit nanometer-level specs for roughness, dishing/recess (Cu recess), and WIW (within-wafer) uniformity, both the oxide–oxide bond and the Cu–Cu electrical contact will break down.

So what exactly does CMP do?

CMP (Chemical Mechanical Planarization) combines chemical reactions with mechanical polishing to produce an ultra-planar surface. In hybrid bonding, the goal is to bring the Cu pads and the surrounding SiOₓ (or SiCN) to the same height while reducing surface roughness to the atomic scale. Once this flatness is achieved, the oxides can bond at low temperature, and then Cu–Cu makes gap-free contact to form a low-resistance connection.

Originally, CMP wasn’t needed for TC bonding.

That’s because TC bonding uses bumps as an intermediary to join semiconductor chips, and the spaces between bumps are filled with a compliant molding material, so flatness requirements aren’t stringent.

Hybrid bonding is different. Because chips are directly joined without separate bumps, surface planarity basically dictates bonding yield.

The problem is that AMAT literally has a 100% monopoly on CMP equipment for hybrid bonding.

AMAT also holds a 60% share of the overall CMP market, and in hybrid-bonding CMP, AMAT is, quite literally, the god.

In April this year, Applied acquired a 9% stake in the Netherlands’ Besi, becoming its largest shareholder. The two companies had already been collaborating in hybrid bonding, and the equity investment made their relationship even tighter.

Even now, in the advanced packaging market for logic, the “Applied (CMP) + Besi (hybrid bonder)” alliance has tremendous influence.

Industry insiders unanimously say that if hybrid bonding technology is introduced into the HBM (high-bandwidth memory) industry, the pair is likewise highly likely to dominate the market.

Just as Japan’s Lasertec mask inspection tools are bundled alongside ASML’s EUV lithography systems, the two companies’ equipment could be supplied in tandem.

There are alternatives in the CMP equipment market—Japan’s Ebara and Korea’s KC Tech—but neither has any track record of supplying CMP for hybrid bonding.

In the paywalled section that follows, I will cover which vendors are currently preparing to enter the hybrid bonder market, their respective strategies, and who I see as the potential winners.

Read more

Recently, the stock price of SanDisk, spun off from WDC, has surged.

Similarly, Japan’s Kioxia Holdings, a Toshiba spin-off, has also posted a sharp rise.

The key question is: what explains this surge?

There is a widely cited saying in economics: ‘The market moves faster than anything else.’

I interpret this as a signal that the market has begun to focus on NAND.

Let’s look back for a moment.

During the worst NAND downturn of 2023–2024, everyone ignored NAND.

Unlike DRAM, there were simply too many players in the market, making supply cuts difficult to implement in a flexible way. On top of that, compared to DRAM, China was able to expand capacity more easily, as the technological hurdles were lower.

Even the AI boom that began in late 2022 didn’t bring much benefit. Data center operators, including CSPs, still preferred the cheaper HDDs when building out their infrastructure.

However, I have now begun to think that demand for NAND has grown so strong that the market may soon tip into a supply shortage.

While HDDs are still cheaper today, their mechanical structure makes them slower, less efficient, and more prone to failure.

SSDs, on the other hand, are far more power-efficient than HDDs, reducing energy costs, cooling requirements, and physical space — all of which are critical factors as data storage demand continues to expand.

Currently, there are two types of SSDs attracting attention in the AI field.

One is high-capacity eSSD, and the other is the still-in-development NL SSD.

Let’s start with eSSD. It refers to large-capacity SSDs, mainly used in inference workloads. However, at present, the price is so high that even CSPs are hesitant to adopt them.

The other is NL SSD (Nearline SSD).

Based on QLC (Quad-Level Cell) flash, Nearline (NL) SSD storage offers higher storage density compared to TLC (Triple-Level Cell) flash, while also being lower cost.

This makes QLC flash a realistic option for data centers to replace HDD capacity.

The target of NL SSD is to replace NL HDD, not to substitute the high-performance QLC eSSD used in AI training and inference.

In the current situation of severe HDD shortages, with lead times exceeding one year, CSPs have expressed a willingness to adopt NL SSD in large volumes — as long as the pricing is reasonable.

These factors could certainly cause a supply shortage in the NAND market, but the real reason I’m bullish on NAND in the mid to long term is High Bandwidth Flash (HBF).

What is HBF?

In February 2025, SanDisk announced a new product called HBF (High-Bandwidth Flash) at its Investor Day.

HBF is a new type of memory built by vertically stacking NAND flash. If HBM (High-Bandwidth Memory) is created by stacking multiple DRAM chips, HBF is made by stacking layers of NAND flash. The biggest bottleneck in today’s AI semiconductors occurs when data moves from memory semiconductors to logic semiconductors.

Conceptually, HBF technology is designed to allow GPUs to quickly access large amounts of NAND capacity to supplement the limited capacity of HBM.

This reduces the time spent fetching data from PCIe-connected SSDs, thereby accelerating AI training and inference workloads.

HBF access speed could be tens to hundreds of times faster than SSDs, and the first generation of HBF is expected to provide 8–16x the capacity of HBM while maintaining a similar bandwidth.

So then, does HBF replace HBM?

I don’t think so. Rather, I expect it will be used in a complementary way alongside HBM.

KAIST Professor Jung-Ho Kim, often referred to as the “father of HBM,” stated:

“Samsung Electronics and SK Hynix’s earnings are currently determined by HBM,” but added, “Ten years from now, HBF will take its place.”

The reason, according to the professor, is that improvements in artificial intelligence (AI) performance depend on both memory bandwidth and capacity. Professor Kim explained, “Today’s AI is based on the transformer deep learning architecture. To process models with as many as one million input tokens, terabyte (TB)-scale data is required.” He added, “When reading and writing such TB-scale data thousands of times per second, if memory bandwidth is insufficient, bottlenecks occur.”

If bottlenecks appear, the response speed of large language model (LLM)-based generative AI services such as OpenAI’s ChatGPT or Google’s Gemini slows down. The bottleneck originates from the current fundamental structure of computers, the von Neumann architecture, in which the CPU or GPU and memory are physically separated, making data transfer speed (bandwidth) between them critical. Professor Kim stressed, “Even if you double the size of the GPU, it is useless without sufficient memory bandwidth. AI performance is memory-bound and ultimately constrained by memory performance.”

What Professor Kim argues is that HBF can take the place of LPDDR on top of existing HBM-based dies. In this way, HBF compensates for HBM’s limited capacity by serving as storage for large-scale AI models directly within the GPU. HBM acts as a cache that temporarily processes data at high speed, while HBF functions as the memory that holds the massive AI models themselves.

In the latter, paywalled section, I will analyze the companies that are expected to benefit from this.

Read more

Recently, there was an article like this.

Meta will unveil new smart glasses called Hyper Nova at its annual Connect event on September 17 and September 18.

The prevailing view was that, unlike the previously released Ray-Ban Meta glasses, the Hyper Nova coming this time would have a display and its price would be high—above $1,400.

However, contrary to expectations, according to a report by Bloomberg’s Mark Gurman, it will launch at $800.

Source

Why is that?

I want to find the reason in competition with Chinese manufacturers.

About a month ago, Alibaba released smart glasses called Quark Vision.

But the price is 1,999 yuan, about $280.

So from Meta’s perspective, with China putting products out that cheaply, they might have thought, “How can we charge more than $1,400?” and decided to set a lower price.

I believe smart glasses will become a truly essential device.

Let me give an example. Unlike smartphones, smart glasses are much more convenient to operate.

You don’t even need to move your fingers—you can operate them through humanity’s most concise action: your voice.

For instance, if you were trying to use AI features like Gemini Live on a smartphone, you would have to open the camera app with your fingers and scan things yourself; with AI glasses, however, the camera on the device is effectively seeing what I’m seeing, so there’s the advantage that I don’t have to issue instructions one by one.

This is the industry’s classification of smart glasses by generation.

So far, Meta’s Ray-Ban Meta falls under the first generation of smart glasses.

Now, Meta’s upcoming Hypernova and Alibaba’s Quark Vision are considered part of the second generation.

Starting from the second generation, displays are built in by default, which means AR functions are also supported.

However, the displays are still expected to be low-resolution — mostly monochrome, with some models possibly featuring full color.

That’s why I believe the second generation is the point where we can truly start calling them smart glasses.

From the second generation onward, displays provide visual information, allowing users to access much richer data through the glasses.

To draw an analogy: up to the first generation, they were more like PDAs, but starting from the second generation, they can be called smartphones.

This is data released by Counterpoint Research.

Please note their projection of 60–80% annual growth starting this year.

With so many manufacturers continuously rolling out new products, the market is bound to keep expanding at the same pace.

Here’s a table summarizing smart glasses released since 2018.

Looking at the types of displays adopted by products that have branded themselves as smart glasses, in the early days of this industry, LCoS was the most common. For example, the well-known Google Glass at the time was also equipped with LCoS. Three such models came out in 2018.

Although LCoS has continued to appear since then, it hasn’t really become the mainstream choice. Today, the main display technology in the market seems to be OLEDOS. As you can see here, there are many more models using it, and new ones keep being released every year.

Interestingly, LEDOS was absent in 2018, 2019, and 2020, but starting in 2021, its presence has been steadily increasing.

For now, it looks like the three technologies are still competing, but the key point to note is that LEDOS has been gradually increasing since 2021.

If you’d like to see this more clearly, please take a look at this chart.

Here, you can see the AR glasses market—in other words, the smart glasses segment within the broader AR market—and the penetration rates of each display technology. We’re now looking at this from the 2025 perspective.

As you can see, LCoS currently holds the largest share. However, since 2022, LEDOS has started to enter the market and is expected to steadily expand its presence. Meanwhile, OLEDOS is projected to gradually decline, and LCoS is expected to stagnate.

To summarize: the rise of MicroLED (LEDOS), the decline of OLED-on-silicon (OLEDOS), and the plateau of LCoS.

The reasons behind this trend are as follows.

Here you can see a comparison of the strengths of LCoS, OLEDOS, and LEDOS.

So, what are the most essential advantages required for displays used in smart glasses? In my view, brightness and power efficiency are the most important.

The reason power efficiency is critical should be clear. Since these displays are embedded in glasses, if power efficiency is poor, the product essentially becomes useless. At best, you could only mount the battery in the frame of the glasses, but without achieving low power consumption, it would be impossible to make them practical.

The next key factor is brightness. For VR headsets—like Apple’s Vision Pro—where the field of view is fully enclosed, brightness doesn’t need to be extremely high. Because outside light is completely blocked, the display isn’t affected by external lighting.

But AR is different. If you remember the first Google Glass, it looked just like regular eyeglasses, meaning it is directly exposed to external light. Outdoors, you’re under sunlight; indoors, you’re under fluorescent or other artificial lighting—all of which interfere with visibility. If the display’s brightness is low, then external light overwhelms it, making it hard to see the image clearly.

That’s why I believe these two factors—power efficiency and brightness—are the most crucial. In this respect, LEDOS has advantages, as you can see here. However, it hasn’t been widely adopted so far, likely due to the lower level of technological maturity.

Meanwhile, LCoS has been used extensively because its maturity level is the highest. Being based on LCD, it allows for higher pixel density, decent color purity, and reasonable brightness and power efficiency. While not perfect, it has been the most practical and accessible technology, which is why it has been adopted the most.

Looking ahead, though, LCoS may stagnate, and its place is expected to be taken by LEDOS. That’s the way I would interpret the trend.

Then, if we sum things up to this point, the conclusion is quite simple.

So, you might think: “Well then, since LEDoS is going to rise, we just need to study it hard, develop it, and sell it.”

But in reality, it’s not that straightforward.

JBD (Jade Bird Display, 极米光电) is a Chinese company that leads the world in the field of ultra-small MicroLED displays. It is virtually the only firm capable of mass-producing the microdisplays essential for smart glasses and AR devices, creating a structure in which global tech giants like Apple and Meta have no choice but to rely on JBD’s supply chain.

Right now, in this market—in this industry—if someone says, “Please supply us with LEDoS, even just samples would be fine,” the only company that can actually deliver is JBD, Jade Bird Display.

If you look at the news, there are plenty of companies saying, “We’re working on LEDoS, we’re doing MicroLED displays,” but most of them are still at the stage of just setting up lines or considering small-scale pilot runs, rather than true mass production.

JBD, on the other hand, is virtually the only company that can actually supply if someone like us says, “We’re going to launch smart glasses, so please provide us with Ledos.”

So right now, as you can see here, most of these companies are allied with JBD, receiving their displays and LEDs from them. Even later, when it comes to mass production, the likelihood is very high that they will all end up working with JBD.

I believe that Apple’s smart glasses currently in development are also highly likely to adopt JBD’s display.

What I am truly concerned about is that it is obvious smart glasses will become the key device in the post-smartphone era, yet Korean companies are not really joining the supply chain.

Korean conglomerates like Samsung and LG, up until the 2010s, were running large businesses in LEDs and filing many patents, but as of now they have effectively withdrawn from the business.

If these two companies had continued operating in this business, they would have responded to MicroLED as well, but the two giant Korean conglomerates today are focusing more on OLEDoS, which can be manufactured by leveraging their existing infrastructure.

Even though LEDoS is technically superior.

I am beginning to think that perhaps in the process of transitioning to LEDoS, Korean display makers might end up losing display hegemony to Taiwan or China.

The only fortunate thing is that there is a Korean company that has a close relationship with JBD.

It is called Sapien Semiconductor, and this company supplies the backplane for the LEDoS provided by JBD.

As confirmed so far, it has been verified that they are supplying smart glass backplanes in alliance with Meta, Sony, and Alibaba.

First issue: Samsung’s 1bnm problem

A few days ago, DRAMeXchange, a subsidiary of TrendForce, released a report stating that there are technical issues with Samsung’s 1bnm.

The summary of the report is as follows:

[DRAMeXchange August Server DRAM Contract Price Announcement]

• Negotiations over server DRAM contract prices remained firm throughout August. For DDR4, CSPs including MS are actively proposing additional orders, and lead times have been extended through 2027. Most of the demand stems from SK hynix raising the production plan at its Wuxi fab. As a result, supply to PC OEMs, module houses, and consumer products has been reduced. DDR4 prices in August rose 0–5%, which is in line with the expected QoQ increase of 28–33%.

• For DDR5, MS/AWS have presented additional orders for 2H25, continuing the demand momentum seen last month from Google and Oracle. On the supply side, technical issues on Samsung Electronics’ 1bnm process products have restricted product availability, further worsening market conditions. DDR5 prices in August rose 0–5%, in line with expectations for a 3–8% QoQ increase. The premium on 6400 MT/s high-bandwidth modules is gradually narrowing as Intel’s Granite Rapids platform moves into mass production (previously about a 5% premium).

• Considering the economics of Ice Lake CPUs, CSPs wish to extend purchases and are actively negotiating with Intel. This has generated aggressive incremental demand for DDR4, eSSD, and HDD. As of August, SK hynix is negotiating long-term orders with major CSPs in the U.S. and China, with final shipment timing extended into 2H26–2H27. Micron is also in talks with major U.S.-based CSPs. In contrast, Samsung Electronics has not delayed the final shipment schedule to date, nor has it increased supply.

• Total server DDR4 production in 2026 is expected to be 7 billion Gb, accounting for about 5% of total server DRAM. The supply shares of SKH, MU, and SEC are estimated at 60%, 25–30%, and 5–10%, respectively. SKH’s aggressive supply expansion stems from clean-room expansion, maintaining the 1znm process line at the Wuxi fab, long-term partnership with MS, and strengthened position in the Chinese market. SEC continues to raise the share of supply based on advanced nodes, and flexibility in DDR4 supply is expected to be a key market variable going forward. With orders continually increasing, DDR4 ASP in 4Q25 could rise an additional 15–20%. ASPs are expected to gradually ease beginning in 1Q26, after the final purchase window.

• It is becoming clear that 2026 Capex for U.S.-based CSPs will increase YoY. As completion of new data centers approaches, server DRAM purchasing is being pulled forward into 2H25. Accordingly, total server shipments in 2026 could be revised up by 2.3%. For DDR5, suppliers’ inventories continue to decline to meet incremental demand in 3Q. However, due to technical problems, order visibility is expected to extend into 4Q25, and the aforementioned demand is expected to be fulfillable from 4Q. As a result, DRAM bit demand in 4Q will likely remain at 3Q levels. Therefore, contract prices are expected to continue rising through year-end, with a high likelihood of turning downward from 1Q26.

(Source: DRAMeXchange)

What stands out here is that technical issues have occurred on Samsung’s 1bnm process products.

When I asked Samsung about this report, the answer was that it is not true.

However, after directly checking with my sources, I was able to learn the following facts.

On July 24, TM Roh, head of Samsung’s MX Division, met with Micron CEO Sanjay Mehrotra.

A source familiar with the matter said that the performance and yield of LPDDR5X mass-produced on Samsung’s 1bnm fall short of Micron’s, and that the meeting was for Samsung’s MX Division to increase the volume of LPDDR5X sourced from Micron for next year’s S26.

Another source stated that there are internal concerns within Samsung about the competitiveness of 1bnm.

To quote directly: “We have to ride out next year on 1bnm, but we’re not confident.”

Samsung’s 1bnm technical issues—this is the first issue I believe will determine next year’s memory market.

The existing bullish thesis for next year’s memory market was this.

HBM will, of course, face supply shortages and see ASP increases,

and the reason many industry analysts, myself included, expected Samsung to benefit to some extent even if it lagged rivals in HBM was the logic that more wafers going into HBM would tighten supply of commodity DRAM as well, thereby benefiting Samsung, which has the largest commodity capacity.

However, if issues with Samsung’s 1bnm prevent it from sufficiently supplying LPDDR5X and DDR5, the story changes.

This would create additional revenue for SK hynix and Micron that hadn’t been assumed, while for Samsung it could lead to falling short of the benchmarks the market is currently using.

Second issue: The HBM commoditization debate

I think HBM is bound to become commoditized in the long run. However, I think it’s too early to worry about that now.

First, this entire HBM commoditization debate starts with NVIDIA.

NVIDIA will use its own logic base die for HBM4E / Rubin Ultra.

This base die will be fabricated on TSMC’s N3 node.

Originally, HBM vendors could differentiate performance versus competitors via the logic base die; if NVIDIA supplies this directly, it can accelerate a certain degree of commoditization of HBM. In other words, NVIDIA would make the core of HBM itself, and memory makers would be left making only the HBM “shell.”

I believe the reason memory makers’ stocks are currently undervalued is precisely this commoditization of HBM.

Normally, most stocks discount some portion of future earnings, but this is particularly pronounced for memory makers. Because they are viewed as traditional cyclical stocks, the cycle gets excessively reflected in their share prices.

However, I look at it a bit differently.

By my estimates, NVIDIA will introduce custom HBM in 2027, AMD in 2028, and Google or AWS in 2028–2029.

This means that even including NVIDIA, adoption wouldn’t come until the second half of 2027, and I think it is far too early to be pricing in a 2H27 issue already.

Also, not every ASIC maker can design custom HBM on its own.

Companies building relatively smaller-scale ASICs could co-develop exclusive HBM4E with a memory maker that has design capabilities—like Samsung—and adopt it one-stop.

Alternatively, Micron and SK hynix could collaborate with design houses to build and supply custom HBM.

That said, I do favor Samsung—a maker that owns its own foundry and has the design capability to provide the HBM4E base die from design through production in a one-stop manner.

In any case, I think it’s too early to judge the custom HBM4E issue now. If anything, while they may sell to players like NVIDIA as a commodity, they could also custom-tailor HBM4E for other companies at higher prices, thereby preserving ASPs and profitability.

Third issue: Revocation of VEU status for Samsung and SK hynix

Before getting into the details, I am extremely disappointed in the U.S. government.

Even with the sanctions, the NAND flash fab that SK hynix acquired from Intel should have been granted an exemption.

Sanctioning even the NAND flash fab that SK hynix acquired from Intel strikes me as a policy that is profoundly out of line with basic decency.

Getting to the main point, this is, rather, positive news in the long run for Chinese memory makers and Chinese semiconductor equipment makers.

This measure will instead accelerate the aging of Samsung and SK hynix’s fabs in China, bring forward depreciation, cause short-term supply-chain disruption, and, in the long run, make it easier for Samsung and SK hynix to withdraw from China.

When SJ decided to close Micron’s China fab, a large number of engineers who had worked there moved to local Chinese DRAM makers such as CXMT. As a result, CXMT—which had been subsisting on Qimonda’s rather mediocre technology—went on to make remarkable progress after Micron’s engineers joined.

Samsung and SK hynix will be no different. If things continue on this path, I think Samsung and SK hynix will wind down their China fabs around 2030–2032. This would be an enormous boon for YMTC and CXMT, because they would be able to acquire the advanced IP and skilled personnel they have long coveted with relative ease.

Another scenario is that Samsung’s and SK hynix’s fabs could be operated using Chinese-made equipment.

Chinese equipment is already used in the cleaning domain at some of SK hynix’s China fabs. By making it harder to introduce U.S.-made tools, the U.S. government’s revocation of VEU status could instead prompt Samsung and SK hynix to turn to Chinese equipment.

That would mean Samsung and SK hynix’s China fabs would effectively serve as validation sites for Chinese equipment, potentially opening export channels for those Chinese tools—an extremely serious development.

I want to criticize the U.S. government’s revocation of VEU status as a profoundly short-sighted decision.

The revocation of VEU status can actually be seen as a positive for Micron.

Micron operates no fabs in China, so it bears none of the risks associated with this VEU revocation. Moreover, if Samsung and SK hynix’s fabs are not upgraded, that will create shortages in advanced DRAM.

Even so, it’s hard to say that Samsung and SK hynix will suffer massive losses from this alone.

Recently, DDR4 prices have surged, and SK hynix has decided to ramp additional DDR4 production at its China fabs—through 2027.

Samsung, likewise, can import equipment without limit until January of next year, so it will try to bring in as much equipment as possible to upgrade its fabs.

This is one of the reasons I expect semiconductor equipment makers to post strong results next quarter.

Before the U.S. government’s sanctions take effect, it’s a foregone conclusion that Samsung and SK hynix will place large equipment orders. Additional revenue can be expected from pull-in demand.