Behind the Memory Chip "Carnival": In the Age of AI, Does the Old Script Still Work?
Hello everyone, I'm your financial journalist. If you've been following the U.S. or Korean stock markets in the past two weeks, you've probably been confused by the drastic fluctuations in the memory chip (DRAM/HBM) sector.
On one hand, the financial reports have been astonishingly positive: companies like SK Hynix, Micron, and Samsung have reported incredible profits, with some even claiming that selling chips is more profitable than selling gold. On the other hand, stock prices have plummeted, with declines of 30-40% from their highs, and the Korean stock market has even seen rare consecutive circuit breakers.
It's like someone who has just won the lottery, but their face is getting paler by the minute. What exactly is the market afraid of? Is the traditional "cycle curse" about to strike again, or is this really different?
Today, we'll break down the logic behind this in five easy-to-understand parts to help you sort out this mess.
---
Part 1: Why Are Profits So High, but Stock Prices Falling? – The Fear of the "Old Script"
First, we need to understand why the market is so panicked. The memory industry has been trapped in a 20-year "vicious cycle," commonly known as the "cycle curse."
Here's how it works:
1. Strong demand: When phones and computers sell well, memory prices rise.
2. Crazy expansion: The three giants (Samsung, Hynix, Micron) see the money and immediately invest in equipment and build new factories.
3. Overcapacity: Two to three years later, new factories start producing, but there's not enough demand for the increased supply.
4. Price collapse: Memory prices drop to rock-bottom levels, sometimes even below cost.
5. Huge losses: Manufacturers lose money on every chip sold and can only survive by reducing production and laying off employees.
The latest painful example was in 2023: Micron's annual revenue was cut in half, and its gross margin turned negative (-9.1%), resulting in a loss of $5.8 billion. The worst part was that major customers like Apple took advantage of the high inventory levels to push down prices. Memory that used to sell for $30 was now being sold for $10. Since memory is a standard product, if one manufacturer doesn't sell, customers will turn to others, leaving the manufacturers with no leverage.
The current market logic is simple: Seeing the high profits of Samsung, Hynix, and Micron, and their record-level capital expenditures (especially Hynix's $1100 trillion Korean won investment plan), the market thinks, "Oh no, another round of overcapacity is coming."
So, despite the good financial reports, investors are worried that this prosperity is just the precursor to another crash. This is why stock prices are falling even when there are positive news—everyone is trying to sell their shares before the next downturn.
---
Part 2: Variable 1: Changing Demand – From "Mobile Phones" to "AI"
If it were just about the cycle, this time might really be different. The reason is that the customers have changed, and so has the nature of the demand.
In the past: Memory was mainly sold to mobile phones, computers, and ordinary servers. These were "stalled markets" where reducing prices wouldn't increase demand significantly.
Now: The demand comes mainly from AI models.
- HBM (high-bandwidth memory) has become essential: AI chips (like NVIDIA GPUs) require HBM.
- HBM is a "wafer devourer": Producing one bit of HBM consumes more wafer area than ordinary DRAM, meaning more HBM production takes up capacity that could otherwise be used for DRAM.
This leads to an counterintuitive phenomenon: Instead of overcapacity, we're now facing shortages. On September 7, Korean brokerage KB Securities warned that Samsung and Hynix' inventory would last less than 10 days. This is not overcapacity; it's a shortage of production capacity.
So, when we see the giants expanding production, AI analysts believe they are not just preparing for the regular market but also aiming to capture the massive AI market. As long as AI continues to grow, this shortage will persist.
---
Part 3: Variable 2: Changing Products – From "Standard Parts" to "Customized Products"
Another key change is that the threshold for memory production has increased.
In the past: Memory was a standard product, and there was little difference between Samsung's and Hynix's 8GB chips, so cheaper options were more popular.
Now (with HBM4): HBM4 includes a custom chip called the Base Die that can be tailored for specific customers (like NVIDIA and Google's accelerators). This means manufacturers need to collaborate with customers years in advance for design and testing.
The consequences?
- Changing suppliers is extremely difficult: If a customer completes the certification process, switching suppliers would be time-consuming and costly.
- Evidence: In 2026, hotels around Samsung and Hynix' factories were booked by U.S. tech companies. AMD CEO Lisa Su visited to discuss cooperation, and Samsung CEO Kim Yong-soo even wrote "Please Make More" on Hynix's booth.
- This gives manufacturers more bargaining power: They are no longer at the mercy of customers, as they become either exclusive or semi-exclusive suppliers.
---
Part 3: Variable 3: Changing Contracts – From "Spot Market Negotiations" to "Long-Term Pricing**
The most significant and often overlooked change is the shift in the business model.
In the past: Manufacturers feared two things: sudden stops in demand and price crashes. They could only watch spot prices fluctuate and suffer from inventory depreciation.
Now: Manufacturers are demanding long-term quantity and price agreements.
- Micron's data: As of June, it had signed 16 strategic customer agreements (SCAs) covering 2026-2030.
- Key terms:
- Take-or-Pay: Customers must pay for the agreed quantity, regardless of actual demand.
- Price Floor: The contract ensures a minimum price, regardless of market fluctuations.
- Deposits: These agreements total about $100 billion, with $22 billion in cash deposits.
This means risk has shifted: In the past, manufacturers bore all the risks of demand fluctuations and price drops. Now, customers also share the risk.
- Benefits for manufacturers: They can plan their production more carefully, signing contracts and receiving deposits before investing in new equipment.
- Cyclical buffer: Even if AI demand slows or spot prices drop, long-term contracts ensure a steady income stream, preventing sharp losses.
Example: Analysts used 2023 data to simulate this. If 50% of Micron's revenue was protected by such contracts (with a 62% gross margin), the 2023 losses (-9.1% and $6.1 billion in free cash flow) would have turned into profits.
Similar simulations in 2009 and 2016 showed similar results: The previous crises would have been less severe with these contracts.
---
Part 4: The Ultimate Question: Can AI Break the Cycle of Overcapacity?
Despite the long-term contracts, we can't be overly optimistic. The memory industry still has cycles, and AI isn't a panacea.
Potential risks:
- Slowing AI investment: If AI companies cut back on spending due to high costs or poor returns.
- Chinese competition: Chinese manufacturers like ChangXin Memory will continue to expand, increasing supply pressure.
- Technological evolution: As HBM standards mature, alternatives may emerge, reducing customization barriers.
But a more fundamental question is: Is supply creating its own demand?
In traditional industries, increased supply usually leads to saturation. However, in AI, there's a "flywheel effect": Lowering computing costs makes more AI applications feasible, leading to increased demand and further infrastructure investment.
If this cycle holds true, then: More memory and computing power will not just follow demand; they will create new demand.
Conclusion:
- The old script: Expansion → Overcapacity → Crash.
- The new script: Expansion → Meeting new AI demands → Creating more applications → Growing demand.
Advice for investors: When evaluating memory stocks, looking at spot prices and inventory levels is no longer enough. You need to consider the long-term contract coverage rate (SCA Coverage). Companies with a higher proportion of long-term contracts are more resilient to cycles. Even if spot prices drop, their profits will be more stable due to contract income.
Finally, a thought-provoking question: When supply itself can create demand, what exactly defines "overcapacity"? If AI demand keeps growing faster than production capacity, the cycle curse might be altered. But if AI demand slows and production capacity is already high, how much protection will long-term contracts provide?
This is a battle between technological belief and **cyclical laws. For now, AI seems to be providing manufacturers with significant protection. However, whether the cycle will still strike and how severe it will be remains to be seen.