Author Profile:
Takashi Uesugi is the Director of the Microdevice Research Institute in Japan. He has long been engaged in semiconductor R&D at the forefront of Japan's manufacturing industry. He obtained his Ph.D. in Engineering from Kyoto University in 2000 and has since been involved in teaching, research, consulting, and journalism related to the semiconductor industry. He has authored books such as "The Failure of Japan's 'Semiconductors'", "Lessons from the Defeat of 'Electronics and Semiconductors'", and "Lost Manufacturing: The Defeat of Japanese Manufacturing".
From 2025 to 2026, anyone who has been following the semiconductor industry for a long time will see a series of numbers that make them doubt their own eyes. The global semiconductor market size has surpassed $1 trillion for the first time in history and is heading straight for $2 trillion. Among them, the two major memory markets, DRAM and NAND, are exhibiting a rapid expansion that can only be described as "abnormal".
Spot prices for DRAM and NAND have surged to about 10 times their previous levels in just over a year and a half. Voices from all over the world are exclaiming that memory for digital consumer electronics such as personal computers, smartphones, and game consoles is "completely insufficient".
Based on public data, this article will sequentially analyze the following two questions: (1) Why is the memory market experiencing such abnormal growth; and (2) How long will this growth last.
01. The Rapid Expansion of the Semiconductor Market
Figure 1 is a chart showing the changes in the global semiconductor market size from 1990 to 2027, based on data from the World Semiconductor Trade Statistics (WSTS) and the spring 2026 forecast. The market, which was about $50 billion in 1990, has grown steadily driven by waves such as the popularization of personal computers, the expansion of the internet, and the emergence of smartphones, reaching over $400 billion around 2020. Up to this point, it can be described as "normal growth".
Figure 1 The global semiconductor market rapidly expanding due to the AI boom (2026-2027 are forecast values); Source: Compiled by the author based on WSTS data and WSTS's Spring 2026 semiconductor forecast.
The fundamental shift in the situation occurred on November 30, 2022, when OpenAI released ChatGPT. The market size shown in Figure 1 surged to $630.5 billion in 2024 and to $795.6 billion in 2025. It is projected to reach $1.5112 trillion in 2026 and $1.9137 trillion in 2027, representing a virtually vertical skyrocket.
In other words, within just 2 to 3 years, the market size will inflate to more than double its original size. In the over 30-year history of the semiconductor industry, there has never been such a steep growth. The release of ChatGPT on November 30, 2022, is evidently the catalyst that caused the drastic change in market trends, which is clear at a glance.
02. The Expansion of the Memory Market Can Be Described as "Abnormal"
It is a fact that the overall semiconductor market is expanding rapidly, but if its composition is broken down, it becomes clear who the main driver of this expansion is.
Figure 2 is a chart that categorizes semiconductors into four product types and shows market changes. These four categories are: MOS Micro (including processors and microcontrollers); MOS Memory, which mainly includes DRAM and NAND; Logic, which includes GPUs from NVIDIA and smartphone processors; and Analog. The most conspicuous element here is the abnormally steep rise of the curve representing MOS Memory.
Figure 2 Memory including DRAM and NAND, and logic chips including GPUs are growing rapidly (2026-2027 are forecast values); Source: Compiled by the author based on WSTS data and WSTS's Spring 2026 semiconductor forecast.
Although the MOS Memory market once fell to $92.3 billion in 2023, it is expected to reach $1.0621 trillion by 2027, growing more than 10 times in just over three years. The combined market size of DRAM and NAND surpassing the $1 trillion mark is unprecedented in the history of the semiconductor industry.
In the same Figure 2, the logic chip market, including NVIDIA's GPUs, will also reach $522.8 billion by 2027, breaking through the $500 billion mark. Notably, the growth momentum of the memory market even exceeds this rate. When mentioning the AI boom, people are often first drawn to NVIDIA's GPUs, the protagonists of AI semiconductors. But in terms of monetary value, the abnormal inflation of memory is actually the most significant characteristic of the current situation.
03. Abnormal Expansion in Both DRAM and NAND Markets
So, what changes are DRAM and NAND exhibiting respectively in the memory sector? Figure 3, based on data from TrendForce, shows the global market trends for DRAM and NAND on a quarterly basis from 2005 to the second quarter (Q2) of 2026.
Figure 3 Global market for DRAM and NAND (Q2 2026 is a forecast value); Source: Compiled by the author based on TrendForce "Data Track" data.
It can be seen from this chart that until around 2022, both had been cyclically experiencing booms and busts along the so-called "silicon cycle". The peak around 2018, known as the memory bubble, the peak around 2021, known as the pandemic-induced special demand, and the subsequent troughs—these fluctuations were all within the scope of past economic cycles.
However, the upswing from 2025 to 2026 is not on the same level as any past peak. From Q1 2023 to Q2 2026, the DRAM market surged from $9.7 billion to $145 billion on a quarterly basis, while the NAND market skyrocketed from $8.7 billion to $81 billion, an increase of about 9.3 times. Both are exhibiting breakthrough and rapid growth.
After the release of ChatGPT, the two markets, which had once bottomed out, leaped upward. This situation can no longer be explained by the conventional concept of the silicon cycle. The "simultaneous" abnormal expansion of DRAM and NAND is an important characteristic of the current situation.
04. The Reason Lies in the Skyrocketing Spot Prices of DRAM and NAND
Since the market is expanding sharply in monetary terms, there must be a sharp increase in prices behind it. Figure 4 shows the changes in spot prices for DRAM (DDR5 16Gb 2Gx8) and NAND (1Tb TLC wafer). From 2023 to the first half of 2025, DRAM prices remained relatively stable in the $4 to $6 range, while NAND prices stayed in the $2 to $3 range. However, entering the second half of 2025, the situation changed completely.
Figure 4 The skyrocketing spot prices of DRAM (DDR5 16Gb 2Gx8) and NAND (1Tb TLC wafer); Source: Compiled by the author based on TrendForce data.
The spot price of DRAM rose from $4.7 in early 2023 to $46.0, about 9.8 times the original; the wafer price of NAND also increased from $2.4 to $25.0, about 10.4 times the original. In just about a year and a half, both have increased by nearly 10 times, an astonishing surge.
Here, since the market size (monetary value) of various types of memory is determined by "price × quantity", it is precisely this round of price skyrocketing that directly pushed up the abnormal expansion of the memory market seen in Figures 2 and 3 from a monetary perspective. It can be said to be the "chief culprit".
A natural question arises here. Why have the prices of DRAM and NAND skyrocketed to such an extent? The following will gradually reveal the reasons from both the demand and supply sides.
05. Whether It's Google Search or Generative AI Inference, Human Behavioral Patterns Are the Same
To understand the fundamental reasons for the price skyrocket, let's first look at human behavior. Figure 5 compares searching on Google with asking questions to generative AI represented by ChatGPT. In other words, on one side are "people using Google to search", and on the other side are "people asking questions to ChatGPT".
Figure 5 One search on Google vs. one question to ChatGPT
It is reported that over 5 billion people worldwide use Google search, and over 1 billion use generative AI. As the arrows indicate, people's behavior is gradually shifting from searching to generative AI.
The key here is that, from a human perspective, the operations for both have hardly changed. Facing a smartphone or personal computer, inputting what one wants to know, and then getting an answer—from the user experience perspective, searching and asking generative AI are almost the same.
In other words, users are completely unaware of the enormous load generated behind the scenes after they shift from "searching" to "asking AI". This gap, where "it looks the same on the surface, but the actual content differs by several orders of magnitude", is exactly the starting point for the explosion in memory demand to be discussed next.
06. The Computational Load of Generative AI is 10,000 to 1,000,000 Times Higher
Even though it looks the same on the surface, what happens inside data center servers is on a completely different level. Figure 6 compares the computational content of one Google search and one GPT-5 inference by item (estimated as of 2024). Looking at this table, the gap between the two is shocking.
Figure 6 Comparison of computational costs for one Google search and one question to ChatGPT-5 (as of 2024); Source: Author's prediction.
Among these, the most decisive factor is Floating-point Operations Per Second (FLOPS). The number of floating-point operations required for one Google search is 1 billion to 10 billion, while one GPT-5 inference reaches 100 trillion to 1,000 trillion. The gap between the two actually reaches 10,000 to 1,000,000 times.
In addition, in the search process, processing that originally only required index matching via a CPU now requires 150 to 400 GPUs to perform ultra-large-scale matrix operations in generative AI. The data showing a 30 to 100 times increase in energy consumption and a 500 to 3,000 times increase in cost per query also corroborates this explosive growth in computational load.
For users, "searching" and "asking AI" are the same operations, but the computational load behind generative AI is 10,000 to 1,000,000 times greater. If the behavior of billions of people worldwide shifts in this direction, the required computing resources and memory will surge astronomically.
07. The Reason for the Abnormal Inflation in Capital Expenditure by Hyperscale Cloud Service Providers
The computational load per operation is 10,000 to 1,000,000 times higher, and over 1 billion people worldwide repeatedly perform such operations every day. Preparing for this unimaginable computational demand are the ultra-large IT companies represented by GAFAM, namely the so-called hyperscale cloud service providers. Figure 7 shows the changes in capital expenditure for the top four hyperscalers: Amazon, Google, Microsoft, and Meta.
Figure 7 Massive capital expenditure of the top four hyperscale cloud service providers; Source: Compiled by the author based on financial reports of each company and data from TrendForce, etc.
The total capital expenditure of these four companies expanded from $21 billion in 2015 to $91 billion in 2020, and reached $217 billion in 2024 after the release of ChatGPT, and $355 billion in 2025. It is expected to finally reach $755 billion in 2026. Calculated at an exchange rate of 1 USD to 160 JPY, this investment is about 120 trillion yen, a scale exceeding Japan's national budget (general accounting basis).
In particular, the surge in the most recent year is extraordinary and truly deserves to be described as "abnormal". What needs to be noted here is that this is ultimately only data for the top four companies. If we add the AI investments from OpenAI, other cloud service enterprises, and various national governments, the total investment in AI infrastructure will become even more massive. They are pouring these huge sums into the construction of AI data centers and the procurement of semiconductors to fill these data centers.
08. The "Black Hole" of AI Data Centers is Devouring Memory
The aforementioned development trend directly leads to the mechanism of price skyrocketing. Figure 8 illustrates the structure of the price surge in a single diagram. The AI data centers being built one after another by hyperscalers are like "black holes", sucking in GPUs as AI semiconductors, High Bandwidth Memory (HBM) stacked with 8 to 12 layers of DRAM, and SSDs equipped with NAND as storage media, one after another.
Figure 8 Why DRAM and NAND prices are skyrocketing
On the other hand, DRAM and NAND used for digital consumer electronics such as personal computers, smartphones, and game consoles have fallen into a state of "severe shortage". The production capacity of memory manufacturers is limited. As a result of the majority of the limited supply being prioritized for allocation to AI data centers, the DRAM and NAND that can be allocated to consumer devices have become extremely insufficient.
If demand significantly exceeds supply, an increase in memory prices is an inevitable market result. This is the truth behind the skyrocketing prices of DRAM and NAND (Figure 4). And once prices rise, the market size (monetary value) will inflate—thus forming the abnormal expansion of the DRAM and NAND markets (Figures 2 and 3).
09. Will This Abnormal Growth Continue?
Up to this point, we have seen the mechanism of the abnormal expansion of the memory market starting with generative AI. The next question we must face is how long this growth will last. The author believes that, from the perspectives of historical context and physical limitations, this frenzy will reach a turning point in the not-too-distant future.
[Part 1] The History of Memory—Year-over-Year Positive Growth Has Never Lasted More Than 5 Years
Figure 9 shows the year-over-year growth rate of the memory market from 1991 to 2028. In this chart, the author has annotated past events that drove changes in the memory market, such as the release of Windows 95, the dot-com bubble, the Lehman crisis, the memory bubble, the pandemic-induced special demand, and the current AI boom.
Figure 9 Year-over-year growth rate of the memory market; Source: Compiled by the author based on WSTS data and WSTS's Spring 2026 semiconductor forecast.
Observing it anew, it is clear that the memory market has always experienced positive and negative growth in repeated short cycles. In the second year of high-speed growth such as 43% or 52%, there is often a sharp decline of 30% or 50%.
There is an extremely important empirical rule here. Carefully tracking Figure 9 reveals that over the past 30-plus years, year-over-year positive growth in the memory market has never consecutively exceeded 5 years. No matter how strong the demand or how prosperous the market situation, due to oversupply or cyclical demand decline, it will inevitably turn to negative growth within a few years. This is the destiny of the memory industry—the silicon cycle.
In this round of AI boom, the growth rate in 2024 was 79.3%, and it is expected to reach an unprecedented 249.5% in 2026. If this historical empirical rule is applied directly, it can be considered that around 2027 to 2028, this abnormal growth is highly likely to pause temporarily.
[Part 2] Demand Grows Exponentially, Supply Infrastructure Grows Linearly—The Barrier of "Unable to Produce"
Besides historical empirical rules, another factor that cannot be ignored is physical limitations. Figure 10 shows in a schematic diagram the relationship between chip demand and the infrastructure capacity supporting this demand.
Figure 10 Chip demand grows exponentially, infrastructure capacity grows linearly
The chip demand driven by generative AI, especially memory demand, will grow exponentially due to the superposition of increasing user numbers and expanding model sizes. In the chart, it can be seen that the chip demand curve, indicated in green, climbs upward like a rocket.
On the other hand, the infrastructure capacity required to meet these demands—namely semiconductor fabs, the power supplied to the fabs, the water used for cooling, and the production capacity of various manufacturing equipment—can only expand linearly, as shown by the red straight line.
Building a new fab takes several years, and the expansion of the power grid, water resource assurance, and the production capacity of equipment manufacturers all involve long lead times and significant constraints. Therefore, the exponential curve and the straight line diverging significantly at a certain point is a mathematical certainty.
The expected outcome of this is the "unable to produce even if there is demand" shown in the upper right of Figure 10. No matter how much AI data centers need memory, once they hit the bottlenecks of power, water, and manufacturing equipment, supply cannot continue to increase. Ironically, this is not because demand weakens and growth stops, but a scenario where growth hits a ceiling due to the physical upper limit of supply. In any case, linearly growing supply cannot support exponentially growing demand forever.
10. Conclusion—Understanding the Structure of the Frenzy and Preparing for the Turning Point
Let us summarize the preceding discussion. The global semiconductor market has shown a vertical upward trend, with the emergence of generative AI as the dividing point, and the expansion of the memory market is particularly prominent. Both DRAM and NAND have exhibited abnormal growth, and the direct reason is that spot prices have risen to about 10 times their previous levels.
Tracing back to the root of the price skyrocket, it ultimately comes down to the shift in human behavior from searching to generative AI, and the 10,000 to 1,000,000 times difference in computational load behind it. This enormous computational demand has caused the capital expenditure of hyperscalers to inflate abnormally. The "black hole" of AI data centers sucks in GPUs, HBM, and SSDs one after another, while at the same time, memory for consumer devices faces extreme shortages. This is the complete reason for the abnormal growth of the memory market.
However, this growth will not last forever. First, historically, there has never been a case where positive growth in the memory market lasted for more than 5 years. Furthermore, linearly growing supply infrastructure cannot fully support exponentially growing demand, and the physical barrier of "unable to produce even if there is demand" is waiting ahead.
Based on these two reasons, the author believes that the current abnormal growth will reach a turning point around 2027 to 2028. Will the history of the "silicon cycle" repeat itself, or will the AI boom come to an end due to the bottleneck of infrastructure? Even if demand is structural, the physical upper limit of supply cannot be changed.
In any case, calmly understanding the frenzied structure of the current AI boom is indispensable for preparing for the upcoming turning point. Because it can be considered that we are standing on the steepest uphill part of the memory supercycle.
Source: Content compiled by Xinpianlaosiji (ID: xinpianlaosiji) from "JBpress", Author: Takashi Uesugi