The underlying logic of storage investment has changed.

A few days ago, Nomura Securities issued a report on the storage industry.

Nomura is the largest securities firm in Japan and has been conducting semiconductor research for many years. Its coverage of Samsung and SK Hynix is regarded as a benchmark in the institutional circle. In short, Nomura's words carry weight.

I searched around. There are many people translating research reports, but hardly anyone really breaks them down and explains clearly what they are talking about. It's a pity because this report is worth a good read.

01

Nomura's core message can be summed up in one sentence: The way to make money in the storage business may be changing.

In the past 40 years, the storage industry has basically followed a single script: price increase, production expansion, oversupply, price drop, loss, production cut, and then price increase again. It goes round and round, no different from pig farming.

The most recent example was in 2018. That year, Samsung's semiconductor division set a record profit, and so did SK Hynix. Everyone was shouting that "the super cycle of storage has arrived". Then what?

The price of DRAM dropped by 40% in less than a year, and Samsung's profit was directly halved.

I looked up the data:

From 2016 to now, DRAM has experienced at least three complete cycles of price increases and decreases. The script of each cycle is almost the same: when the price is rising, people think "this time is different", but when the price drops, they remember that "the same thing was said last time".

This cyclical characteristic determines how the market values storage companies.

The logic is simple: You may earn 10 billion this year, but only 3 billion next year, and you may even incur losses the year after. With such unstable profits, the market is reluctant to value you based on your profits.

So what should be done? Look at the net worth.

How much your factory and equipment are worth, that's how the valuation is calculated. That's why storage companies have long been valued using the PB ratio, that is, the price - to - book ratio.

If your assets are worth 100 billion, and the market gives you a 1.5 times PB, then you are worth 150 billion. As for how much you earn this year, the market doesn't really care because it thinks you probably won't earn that much next year.

SK Hynix's forward PE for the next 12 months is about 6 times. What does 6 times mean? The PE of any bank in the A - share market is higher than this. One of the world's most profitable storage companies, the market is only willing to value it at 6 times the price - to - earnings ratio.

The market thinks that the money it earns is "not sustainable". This has been the consensus for 40 years. Every cycle, someone tries to break it, but every time they hit a wall.

So, when Nomura shouts again that "this time is really different", hardly anyone takes it seriously. After all, this sentence has appeared too many times in the history of Wall Street. Every time it is said, it is basically a signal that the cycle has reached its peak.

02

Let's first look at the data of this round of price increases. I'll highlight a few key points:

Citi expects the average price of DRAM to increase by 88% throughout 2026. What does that mean? It means the price of chips is almost going to double.

SK Hynix's operating profit margin in Q1 was 72%, breaking the record since the company was founded. The profits of Samsung and Micron also soared collectively. The contract price in Q1 alone increased by more than 90%.

The numbers are really astonishing.

If you only look at these, it's really no different from 2018. It's still price increases, record profits, and everyone is still shouting that "the super cycle has arrived". The script is exactly the same.

What's different this time is another thing. Cloud providers like Google, Microsoft, and Meta have started signing 3 - to 5 - year volume - lock contracts with storage original manufacturers and are also paying upfront deposits. These are written commercial agreements.

The industry calls it LTA, the long - term supply agreement.

I looked through the investor day materials of Kioxia. It has already signed three such long - term agreements, locking in a minimum income of over $42 billion. SK Hynix has also signed 2 - to 3 - year LTAs with NVIDIA and AMD, and the production capacity of HBM is basically locked by the contracts.

Then you may ask, how were storage products sold before?

The price was negotiated every quarter, and they were sold at the market price. It was similar to selling vegetables. If cabbages cost $2 today and $1.5 tomorrow, you had no bargaining power and it all depended on supply and demand.

Now? The cloud providers come and say: I'll sign a 3 - year contract with you. The volume is locked, and there's a price floating range. I'll give you a deposit first.

The change from "bargaining every quarter" to "signing a 3 - to 5 - year contract" has never happened in the 40 - year history of the storage industry.

Why are cloud providers suddenly willing to sign such contracts? Because they're afraid they won't be able to buy enough. The DRAM demand of an AI server is 8 to 10 times that of an ordinary server, and the NAND demand is 3 times.

According to data from JPMorgan Chase, the storage consumption of AI inference is 3 times that of training. Inference runs continuously 24/7, not just once. The demand is not only growing but also devouring.

The cloud providers did some calculations:

If they don't lock in the production capacity in advance, they may not be able to buy any when they really need it. So, they'd rather pay a deposit in advance and accept the price floating range to secure the volume for the next few years.

The most vivid example of this happened within Samsung.

At the end of last year, Samsung's semiconductor division rejected the request of its mobile phone division to sign an LPDDR supply agreement for more than a year and only offered a quarterly contract. Even when the executives of the mobile phone division went to negotiate in person, they only got a quarterly volume.

The reason is simple. The semiconductor division gave priority to the production capacity to AI customers like NVIDIA and cloud providers who had signed long - term contracts. Why? Because they offered higher prices.

The price of 12GB LPDDR5x has risen from $33 at the beginning of the year to $70, more than doubling.

The mobile phone division couldn't get a long - term contract, and it couldn't control the cost. The Galaxy S26 was forced to increase in price, and the Galaxy Z TriFold was discontinued three months after its launch. Samsung's mobile division may even face its first annual loss ever.

The world's largest memory manufacturer not selling memory to its own division. Abstract, right?

There have been years with sharp price increases before, but the highlight of this event is that "even Samsung's own mobile phone division can't compete with AI customers who have signed long - term contracts". The right to allocate production capacity has shifted from market price to contract terms.

In the past, price increases were determined by supply and demand. Once production capacity caught up, the price would come back. But this time, a part of the production capacity is locked in advance by 3 - to 5 - year contracts.

Even if the supply catches up later, this part of the production capacity won't be released to the open market. This is a structural change, completely different from the cyclical one.

03

In short, LTA doesn't just change how storage companies sell their products. It changes how the market should value storage companies.

Let me give you an example:

Previously, storage companies were like farmers whose harvests depended on the weather. If the weather was good this year, the harvest would be good, but if there was a drought next year, they might have no harvest at all.

If you were to value this farmer, you wouldn't base it on this year's harvest because it could be zero next year. You could only look at how many acres of land and how many cows he had and value him based on his assets.

This is the logic of PB valuation. Now the situation has changed.

Someone has signed a 3 - year purchase contract with this farmer in advance, locking in the volume and price and paying a deposit. The farmer's harvest next year may not be better than this year's, but at least a part of his income is guaranteed.

If the predictability of income has changed, it's unreasonable to value him based on his "assets". It should be based on "how much he can earn stably every year", that is, the PE, the price - to - earnings ratio.

This is not my wild speculation. This is the core logic of Nomura's report.

I checked. Nomura made a comparison. TSMC's revenue growth rate in the next 3 to 5 years is about 30%, and the market gives it a 20 times forward PE. Samsung and Hynix have a similar growth rate, but the market gives them a valuation that is less than one - third of TSMC's.

In the past, this discount was reasonable because TSMC's revenue is locked by OEM contracts, while the revenue of storage companies follows the market trend. One is stable, and the other is unstable, so the pricing is naturally different.

Now that LTAs are being widely adopted, a part of the revenue of storage companies is also locked by contracts. Is this discount still reasonable?

Nomura thinks it's not. It has raised SK Hynix's target price from 880,000 won all the way to 4 million won, more than quadrupling it in half a year. It has also changed the valuation method from PB to PE.

Interestingly, Nomura is not the only one thinking this way.

I looked through the reports from the same period. SK Securities used the PE framework to value storage companies for the first time on May 7th. Macquarie significantly raised the target price on May 15th. JPMorgan Chase shouted about a "multi - year super up cycle" as early as the end of last year.

The analysts from several investment banks may not have communicated with each other, but they almost simultaneously reached the same conclusion.

SK Hynix's financial report is the strongest evidence for this logic. Its revenue in Q1 exceeded 50 trillion won, the operating profit was 37.6 trillion won, and the profit margin was 72%.

The proportion of HBM in DRAM revenue jumped from 20% in the previous quarter to 35%. The price of a single HBM chip exceeded $10,000, and the gross profit margin exceeded 90%. At the end of the quarter, it had 54.3 trillion won in cash, and the net cash on hand was 35 trillion won.

This company is no longer short of money. Its HBM production capacity is locked by long - term contracts for 2 to 3 years, and a part of its future income is guaranteed by the contracts.

In short, a part of the money for the next few years is already in the bag.

However, I have to pour some cold water here. Nomura's own quantitative strategist issued another report almost in the same week, warning that the leverage mechanism driving this round of AI bull market will turn against the market when the market reverses, and the semiconductor ETF may plummet by 15% in a single day.

The same investment bank is shouting "this time is different" with one hand and hedging against extreme risks with the other. What does this mean? It means that even the most bullish people are not 100% sure.

LTA is indeed a brand - new variable.

But I checked and found that signing long - term agreements in the storage industry is not the first time.

When DRAM was in short supply in 2017, the industry also signed similar forward purchase agreements. Later, when the demand slowed down and the inventory increased, the DRAM price dropped by more than 40% within 2 to 3 quarters. Customers delayed taking delivery, and the contract price was renegotiated back to close to the spot level. That round of agreements failed to lock in the cycle.

This was mentioned in a joint report by Morgan Stanley and JPMorgan Chase, not just a rumor.

This time, the terms are indeed much stricter than in 2017. South Korean storage companies have increased the upfront payment ratio from less than 5% in the past to 10% to 30%. If customers don't take delivery as agreed, the upfront payment will not be refunded as a penalty. The contracts that SK Hynix signed with Microsoft and Google also include a minimum price guarantee.

The contract terms are stricter, and the cost of breach of contract is higher. This is a fact. But can it really withstand a downward cycle? There were also contracts in 2017, but it still couldn't withstand it.

So, it's a fact that LTA has changed the revenue structure of the storage industry. Should the valuation framework be changed accordingly? It's reasonable, and the logic makes sense. But there is still a real downward cycle to test whether it's really right.

04

Since LTA is so good, can all storage companies benefit from it? The answer is simple: They are not in the same world.

There are only a few companies in the world that can sign 3 - to 5 - year LTAs with cloud providers: Samsung, SK Hynix, Micron, and Kioxia.

They have one thing in common: they have their own wafer fabs and manufacture their own chips. Cloud providers need this kind of source - level production capacity guarantee. They won't sign a 5 - year volume contract with an intermediary.

There are also a large number of companies in the storage industry that are engaged in the intermediate links. They buy wafers from original manufacturers, do their own packaging and module production, and sell them to mobile phone and computer manufacturers under their own brands. These companies are also signing LTAs, but in a different direction.

They sign LTAs on the procurement side to lock in "how many wafers I can buy from the upstream". The purpose is to ensure supply, not to lock in income.

On the downstream side, when selling to customers like OPPO, Lenovo, and Xiaomi, the price still follows the market trend. When the price rises, the profit soars, and when the price drops, the profit shrinks.

I used AI to analyze the data:

The profit data of the A - share storage sector in Q1 this year is really astonishing. Some companies' year - on - year growth exceeded 2000%, and some had a 1500% growth. The numbers look as impressive as SK Hynix's, but the sources of profit are completely different.

A part of SK Hynix's profit is locked by long - term contracts. Even if the DRAM price drops by 20% next year, this part of the income will not be affected because it's written in the contract.

The profits of these A - share companies almost entirely come from price increases. When the DRAM price rises, the value of the wafers they buy increases, and the selling price also rises, so the profit explodes. But once the price turns around, this logic will work in reverse.

Although they are all storage companies and both had a sharp increase in Q1 profits, the underlying structures of making money are completely different. One has a contract as a guarantee, and the other depends entirely on the market.

What about Chinese storage original manufacturers?

Take Changxin Technology, which is about to go public. It is the leader in domestic DRAM, and its profit in the first quarter was also very strong.

But looking at its product structure, more than 70% of its income comes from LPDDR, which is sold to domestic mobile phone manufacturers. These customers won't sign a 5 - year volume contract with you. They're not even sure if they can sell their own mobile phones, so how can they lock in the volume for 5 years?

So, even for domestic storage original manufacturers, the income structure is different from that of SK Hynix. In the money earned in the first quarter, the price contributed the most, and the volume and efficiency contributed a part. The price is given by the cycle, and once the cycle turns, this part will be gone