This round of extraordinary growth will reach a critical turning point in 2027–2028.
Anyone following the semiconductor industry trends from 2025 to 2026 would undoubtedly be shocked by the successive data releases. The global semiconductor market size has broken the $1 trillion mark for the first time and is sprinting towards $2 trillion; among them, the two major memory storage markets, DRAM and NAND, are experiencing an "abnormal" explosive growth.
Over the past year and a half, spot prices for DRAM and NAND have surged nearly 10 times. Alerts are sounding worldwide: memory storage for consumer digital products such as personal computers, smartphones, and gaming consoles is in severe short supply. Relying on public data, this article will progressively clarify two core issues: first, why the memory storage market is experiencing such extreme high-speed growth; and second, how long this growth cycle can last.
The Semiconductor Market Experiences Rapid Expansion
Figure 1, based on historical data from the World Semiconductor Trade Statistics (WSTS) and spring 2026 forecasts, illustrates the trend of the global semiconductor market size from 1990 to 2027. In 1990, the market size was approximately $50 billion. It grew steadily alongside waves such as the popularization of personal computers, the development of the internet, and the advent of smartphones, reaching the $400 billion level around 2020. This phase represents normal growth.
Figure 1 The global semiconductor market expanding sharply driven by the AI boom (2026–2027 are forecast values). Source: Compiled and drawn by the author based on WSTS data and WSTS Spring 2026 semiconductor forecast data.
The node where the market situation underwent a complete transformation was after OpenAI launched ChatGPT on November 30, 2022. The market size reached $630.5 billion in 2024 and climbed to $795.6 billion in 2025; institutions forecast it will reach $1.5112 trillion in 2026 and further surge to $1.9137 trillion in 2027, with the curve pointing almost vertically upwards.
In just two to three years, the market size has more than doubled. Looking back at the development history of the semiconductor industry over the past three decades, such a steep growth curve is unprecedented. It is evident that the debut of ChatGPT on November 30, 2022, became the critical turning point that reversed the slope of market growth.
The Expansion Trend of the Memory Storage Market Is Particularly "Abnormal"
Although the overall semiconductor market size is growing rapidly, breaking down the structure clearly reveals the core driving force behind the growth.
Figure 2 divides semiconductors into four major categories for trend statistics: MOS Micro, which includes processors and microcontrollers (MCUs), MOS memory dominated by DRAM and NAND, logic chips including NVIDIA GPUs and mobile phone processors, and analog chips. The most prominent feature in the chart is the unusually steep upward curve of MOS memory.
Figure 2 DRAM and NAND memory, along with logic chips including GPUs, experience high-speed growth (2026–2027 are forecast values). Source: Compiled and drawn by the author based on WSTS data and WSTS Spring 2026 semiconductor forecast data.
The MOS memory market once shrank to $92.3 billion in 2023, and institutions forecast it will surge to $1.0621 trillion by 2027, expanding more than tenfold in just over three years. The combined market size of DRAM and NAND breaking the $1 trillion mark is unprecedented in the history of the semiconductor industry.
Looking at Figure 2 again, the logic chip market, including NVIDIA GPUs, is expected to reach $522.8 billion in 2027, breaking the $500 billion threshold; however, it is worth noting that the growth momentum of memory storage far exceeds that of logic chips. When discussing the AI boom, public attention often focuses first on AI core chips like NVIDIA GPUs, but in terms of market size, the explosive expansion of memory storage is the most distinct feature of this cycle.
DRAM and NAND Markets Simultaneously Experience Extraordinary Expansion
Within the memory storage sector, what are the respective trends of DRAM and NAND? Figure 3, based on TrendForce data, presents the changes in the global DRAM and NAND market size on a quarterly basis from 2005 to the second quarter of 2026.
Figure 3 Global DRAM and NAND market size (Q2 2026 is a forecast value). Source: Compiled and drawn by the author based on TrendForce DataTrack data.
As seen from the chart, prior to 2022, the two types of memory storage always followed the traditional "silicon cycle," alternating between boom and bust. The memory market peak around 2018, the boom peak driven by pandemic-induced demand in 2021, and the subsequent trough all had fluctuation ranges within the historical boom-bust cycle intervals.
However, the upswing from 2025 to 2026 is on a completely different level from any previous peak. From Q1 2023 to Q2 2026, the quarterly market size of DRAM surged from $9.7 billion to $145 billion, an increase of nearly 14 times; the quarterly size of NAND grew from $8.7 billion to $81 billion, an increase of about 9.3 times, both showing breakthrough surges.
Since the launch of ChatGPT, the two memory markets bottomed out simultaneously and then skyrocketed in a straight line. This phenomenon can no longer be explained by traditional silicon cycle theory. A key feature of the current stage is the simultaneous abnormal expansion of the DRAM and NAND markets.
Behind the Surge in Market Size: Sharp Skyrocketing of DRAM and NAND Spot Prices
The rapid expansion of market size measured by revenue inevitably stems from soaring prices. Figure 4 shows the spot price trends of DRAM (DDR5 16Gb 2G×8) and NAND (1Tb TLC wafer). From 2023 to the first half of 2025, DRAM prices remained in the $4–$6 range, and NAND wafer prices stabilized in the $2–$3 range. However, entering the second half of 2025, the market situation completely reversed.
Figure 4 Significant increase in spot prices of DRAM (DDR5 16Gb 2Gx8) and NAND (1Tb TLC wafer). Source: Compiled and drawn by the author based on TrendForce data.
DRAM spot prices rose from $4.7 in early 2023 to $46.0, an increase of about 9.8 times; NAND wafer prices climbed from $2.4 to $25.0, an increase of about 10.4 times. In just a year and a half, the prices of both products have surged nearly tenfold, which is quite astonishing.
Memory storage market revenue = price × shipment volume. It can be seen from this that the soaring prices are the core factor directly driving the abnormal revenue expansion of the memory market in Figures 2 and 3. This naturally raises a question: why have DRAM and NAND prices experienced such drastic increases? The following text will dissect the underlying reasons from two dimensions: the demand side and the supply side.
Google Search and Generative AI Inference: User Behavior Patterns Converge
To understand the fundamental logic behind the price increases, we must first observe the changes in human user behavior. Figure 5 compares Google search behavior with generative AI questioning behavior represented by ChatGPT, with Google search users on one side and ChatGPT users on the other.
Figure 5 Single Google Search VS Single Query to ChatGPT
The global number of Google search users exceeds 5 billion, and generative AI users have surpassed 1 billion. As indicated by the arrows in the figure, user behavior is gradually shifting from traditional search to generative AI. The key point is: from the perspective of ordinary users, there is almost no difference between the two behaviors. People input the information they want to query on their mobile phones and computers, and wait to get answers—from a user experience perspective, using a search engine and asking generative AI are basically the same.
In other words, when users switch from "search engine retrieval" to "asking AI," they are completely unaware that an extremely massive computing load will be generated in the backend. This gap between similar surface experiences and underlying consumption differing by several orders of magnitude is precisely the starting point for the explosion in memory demand to be discussed next.
Generative AI Computing Volume Is Up to 10,000 to 1,000,000 Times That of Traditional Search
Even though the front-end operational feel is consistent, the tasks executed inside data center servers are worlds apart. Figure 6 compares the computing content of a single Google search and a single GPT-5 inference item by item (based on 2024 estimates), and the vast gap between the two sets of data is shocking.
Figure 6 Computing comparison between a single Google search and a single query to GPT-5 (2024 estimate). Source: Estimated by the author.
The most critical metric here is the floating-point operations (FLOP). A single Google search requires approximately 1 billion to 10 billion FLOPs; whereas a single GPT-5 inference can reach 100 trillion to 1,000 trillion FLOPs, a gap of fully 10,000 to 1,000,000 times between the two.
In addition, traditional search relies solely on CPUs to complete index matching; generative AI, however, requires 150 to 400 GPUs to execute ultra-large-scale matrix operations. The data showing that energy consumption per task is 30 to 100 times higher, and execution cost per task is 500 to 3,000 times higher, also corroborates the explosive growth in computing demand.
For users, searching and asking AI appear to be no different, but in terms of underlying computing scale, generative AI is 10,000 to 1,000,000 times higher. As the usage behavior of billions of people globally continues to migrate to generative AI, the required computing resources and memory capacity will expand at astronomical levels.
The Root Cause of the Sharp Expansion in Capital Expenditure by Hyperscalers
Over 1 billion users globally repeatedly initiate requests daily with computing volumes 10,000 to even 1,000,000 times higher. The giant tech companies such as GAFAM that undertake these massive computing demands are what the industry calls hyperscalers. Figure 7 shows the capital expenditure trends of the four leading hyperscalers (Amazon, Google, Microsoft, Meta).
Figure 7 Massive capital expenditures of the four leading hyperscalers. Source: Compiled and drawn by the author based on their respective financial reports and TrendForce public data.
The combined capital expenditure of the four companies grew from $21 billion in 2015 to $91 billion in 2020; it reached $217 billion in 2024 after the release of ChatGPT, and further climbed to $355 billion in 2025; institutions forecast it will exceed $755 billion in 2026. Calculated at an exchange rate of 160 yen to 1 US dollar, the scale is approximately 120 trillion yen, which has already surpassed Japan's total general account annual national budget.
In particular, the growth rate over the past year is extremely abnormal, and describing it as an "abnormal surge" is not an exaggeration. It should be noted that the above only includes statistics for the four leading companies. If investments related to AI by OpenAI, other cloud service providers, and various governments are included, the total scale of funds directed towards AI infrastructure will be even more massive. These huge amounts of funds are mainly used to build AI data centers and procure supporting semiconductor hardware.
AI Data Centers Continuously Devour Memory Storage Like "Black Holes"
The logical chain outlined above directly points to the internal mechanism behind the soaring memory prices. Figure 8 illustrates the complete structure formed by this round of price increases in a schematic diagram: major hyperscalers continuously build AI data centers, which act like giant "black holes" constantly absorbing various types of hardware—GPUs as AI chips, High Bandwidth Memory (HBM) stacking 8 to 12 layers of DRAM, and SSDs equipped with NAND flash memory used as storage media, flowing into data centers endlessly.
Figure 8 Why DRAM and NAND Prices Continue to Skyrocket
Meanwhile, DRAM and NAND for consumer electronics such as PCs, smartphones, and gaming consoles are facing severe shortages. Memory manufacturers have a capacity ceiling, and the limited capacity is mostly prioritized for the AI data center sector, ultimately resulting in an extreme shortage of memory chips diverted to the consumer terminal market.
Demand significantly exceeds supply, and memory prices naturally rise. This is the objective law of market operation and also the essential reason for the soaring prices of DRAM and NAND in Figure 4 above. The price increase further pushes up the market size measured by revenue, ultimately forming the abnormal expansion situation of the DRAM and NAND market sizes presented in Figures 2 and 3.
Can This Round of Extraordinary Growth Sustain?
The preceding text has sorted out the entire logic of the abnormal expansion of the memory market size triggered by generative AI. Next, we need to explore the core issue: how long will this growth cycle last? Judging from two dimensions—historical patterns and physical constraints—the author believes that this boom will reach a turning point in the near future.
Argument 1: Historical Patterns of the Memory Industry—Never Achieved Five Consecutive Years of Year-on-Year Positive Growth
Figure 9 shows the year-on-year growth rate trend of the memory market from 1991 to 2028. The chart marks the key events driving the various cycles of the memory industry: the release of Windows 95, the dot-com bubble, the Lehman crisis, the memory boom cycle, pandemic-driven demand, and the current AI boom.
Figure 9 Year-on-year growth rate of the memory storage market. Source: Compiled and drawn by the author based on WSTS data and WSTS Spring 2026 semiconductor forecast data.
Looking at the curve, it is not hard to find that the memory market has long alternated between growth and recession in short cycles. Even if high growth rates of 43% or 52% are achieved in a given year, the following year often sees a sharp decline of 30% or 50%.
One highly reference-worthy historical empirical pattern: looking back at the trends over the past three decades, the memory market has never experienced five consecutive years of year-on-year positive growth. No matter how strong the demand or how high the boom, it will eventually turn into negative growth within a few years due to overcapacity or a phased peak in demand. This is the silicon cycle that the memory industry finds hard to escape.
Under this round of AI boom, institutions forecast the memory market growth rate to reach 79.3% in 2024, and an unprecedented 249.5% in 2026. However, if this historical pattern is used for deduction, this round of abnormal high growth is highly likely to come to a phased halt in 2027 to 2028.
Argument 2: Demand Grows Exponentially, While Supply Infrastructure Can Only Expand Linearly—Encountering the "Capacity Ceiling" Barrier
Besides the historical cycle patterns, another constraint that cannot be ignored comes from the physical level. Figure 10 illustrates the relationship between chip demand and the carrying capacity of the infrastructure supporting the demand.
Figure 10 Chip demand grows exponentially, while infrastructure supply capacity grows linearly
Driven by generative AI, chip (especially memory) demand shows an exponential climb, accompanied by the superimposed effects of user scale expansion and the continuous scaling up of AI models. As shown by the green curve in the figure, the chip demand curve shoots up like a rocket.
However, the infrastructure that can match the demand—semiconductor factories, supporting power, cooling water, and the capacity of various production equipment—can only expand linearly as indicated by the red straight line in the figure. Building a new wafer fab takes years; power grid expansion, water resource allocation, and equipment manufacturer capacity expansion all require long preparation cycles and are subject to multiple constraints. From a mathematical logic perspective, the exponential growth curve and the linear growth curve will inevitably diverge significantly in the end.
This will give rise to the situation described in the upper right of Figure 10: there is demand, but production cannot be expanded. Even if AI data centers continue to crave memory, once the bottlenecks formed by power, water sources, and production equipment are hit, the supply scale will be hard to continue increasing. Ironically, this round of growth slowdown does not stem from weakening demand, but is constrained by the physical limits of supply. In any case, linearly expanding infrastructure cannot sustain exponentially climbing demand in the long term.
Conclusion: Recognize the Internal Logic of This Boom and Prepare for the Cycle Turning Point
Since the advent of generative AI, the global semiconductor market has seen a steep upward trend, with the memory sector showing the most prominent growth. DRAM and NAND have simultaneously achieved extraordinary expansion, with the direct trigger being a nearly tenfold surge in spot prices.
Tracing back to the source, the underlying logic of the price increase lies in the shift of user behavior from traditional search to generative AI, while the computing volume behind the two differs by 10,000 to 1,000,000 times. The massive demand for computing power has driven a substantial expansion in capital expenditure by hyperscalers; AI data centers, like "black holes," continuously absorb GPUs, HBM, and SSDs, while at the same time, supply diversion has led to a severe shortage of memory in the consumer electronics sector.
However, this round of growth cannot last forever. First, from a historical pattern perspective, the memory market has never recorded five consecutive years of positive growth. Second, linearly expanding supply infrastructure is difficult to accommodate exponentially growing demand, and the physical barrier of "having demand but being unable to produce" will eventually emerge. Based on these two points, the author judges that the current round of extraordinary growth will reach a critical turning point in 2027–2028. At that time, will history repeat itself with the silicon cycle starting to decline again, or will the AI boom gradually cool down dragged by infrastructure bottlenecks? Even if demand has long-term structural support, the physical upper limit on the supply side remains difficult to break through.
Regardless of the final trajectory, calmly clarifying the internal logic behind the current AI boom is an indispensable preparation for welcoming the cycle turning point. We are currently in the steepest upward phase of the memory super cycle.
*Disclaimer: This article is created by the original author. The content represents their personal views. Our repost is solely for sharing and discussion and does not imply our endorsement or agreement. If you have any objections, please contact the backend.