The memory chip shortage persists: with over 5.6 trillion invested in AI, capacity won't catch up until 2028
Some time ago, I was chatting with a friend who designs hardware solutions. He works with numerous smart cockpit and edge computing device manufacturers, and the most headache-inducing issue over the past six months has been the inability to secure memory chips. Whether it is DRAM dedicated to industrial equipment or high-speed flash memory for AI servers, original manufacturers have all implemented quota-based supply, with order lead times extended directly by three to six months.
Coincidentally, Ingenic Semiconductor's recent external communication on the supply chain status corroborated the industry-wide shortage. The company's management frankly stated that its core automotive-grade and industrial-grade memory products have long been in a state of short supply. Foundry capacity has been locked in by customers' long-term contracts. Even with continuously rising order demand, there is no way to rapidly expand production and deliver. According to data calculated by major global institutions, the capital expenditure scale of the entire AI industry chain will reach the 5.6 trillion level in the coming years. The massive computing power construction has spawned massive storage demand, but the landing time of new global storage capacity is concentrated in 2028, making the contradiction of supply-demand mismatch difficult to resolve in the short term.
This entire article is compiled based on data from TrendForce, public research minutes from various semiconductor companies, and industry capital expenditure announcements. All data are traceable. It breaks down the supply and demand logic of the memory industry in plain language that ordinary people can understand, without blindly being bullish or deliberately instilling panic, and objectively sorts out the opportunities and potential risks.
I. Understanding the Underlying Reality of the Industry-Wide Memory Chip Shortage Through Ingenic Semiconductor's Current Operations
Many people only know that memory chip prices are rising, but few understand why goods are unavailable even if you have money. Ingenic Semiconductor's business model can intuitively reflect the supply constraints of the entire industry chain.
1. Asset-Light Design Model: Capacity Completely Constrained by Wafer Foundries
Ingenic Semiconductor is a typical Fabless chip design company, responsible only for chip R&D and design, with all production processes outsourced to wafer foundries. It does not have its own manufacturing production lines. The company's core products are divided into two major segments: one is niche DRAM dedicated to smart vehicles and industrial control, and the other is memory supporting AI edge devices. Demand for these two types of products has exploded in recent years.
However, constrained by the limited overall capacity of foundries, even with a steady stream of orders from downstream customers, the company can only schedule production according to the limited capacity allocated by the wafer foundries. Internal company research information shows that all current memory capacity has been locked in by long-term supply agreements signed by automakers and computing power enterprises. New customers wanting to secure goods can only queue up and wait for idle capacity, significantly increasing the difficulty for small and medium-sized hardware manufacturers to obtain supplies.
I know a procurement manager for automotive components who reported multiple times in the first half of this year that it is difficult for small and medium-sized Tier 1 manufacturers to secure stable memory supplies. They can only continuously increase their procurement budgets. Meanwhile, leading automakers, relying on large long-term agreements, can prioritize locking in capacity. The trend of industry resources concentrating towards large customers is very obvious.
2. Continuous Explosion in Demand in Niche Tracks Further Exacerbates Supply Tightness
Unlike consumer electronics memory, the automotive-grade and industrial memory sectors that Ingenic Semiconductor focuses on have extremely high entry barriers. The complete certification cycle for a single product is as long as two to three years. It is difficult for new players to quickly enter the track, and the incremental market supply is inherently limited.
The penetration rate of new energy vehicles is rising year by year. The number of memory chips equipped in a single smart model is more than three times that of traditional fuel vehicles. Smart cockpits and autonomous driving modules all require stable DRAM support. Meanwhile, the popularization of edge AI devices and industrial computing terminals continues to open up new demand spaces. Under the superposition of dual demand, the shortage in the niche memory track even exceeds that of general-purpose server memory.
3. Industry Price Increase Trend Continues, but Growth Ceiling is Locked by Capacity
According to contract price data released by TrendForce, DRAM contract prices in the second quarter of 2026 increased by over 50% month-on-month. Although the increase in the third quarter has somewhat converged, it still maintains an upward range of 13% to 18%. Companies can improve gross profit margins through product price increases, but constrained by the upper limit of wafer capacity, shipment volumes cannot expand synchronously. Performance growth can only rely on the price dividend, making it difficult to achieve leapfrog growth through volume expansion.
Simply put, the current memory industry is in a pattern of "high price, low volume." Price increases can bolster profits, but before the capacity bottleneck is resolved, it is difficult for the overall market supply scale to increase significantly.
II. 5.6 Trillion in CapEx Across the Entire AI Industry Chain Spawns an Epic Demand Gap for Memory Chips
Global tech companies, cloud computing vendors, and computing power service providers will cumulatively invest 5.6 trillion in AI infrastructure over the coming years. This huge expenditure will not only flow to GPU chips. As the core carrier for computing power operation, memory chips will see demand growth far exceeding market expectations.
1. Storage Consumption per AI Server Far Exceeds Traditional Equipment
A regular data center server has limited memory capacity, whereas servers used for large model training and AI inference require 8 to 10 times the DRAM configuration of traditional servers, with supporting flash memory capacity reaching more than three times that of regular models.
In 2026, the DRAM capacity consumed by global AI servers already accounts for 53% of the global total capacity, compared to just 33% two years ago. With the continuous implementation of AI agents and local edge-side large models, mobile phones, PCs, and industrial terminals will all add AI computing modules. Storage demand will spread from the single server track to all categories of electronic devices, and the demand increment will continue to be released.
2. Capacity Shift by Leading Memory Original Manufacturers Squeezes General-Purpose Memory Supply
Samsung, SK Hynix, and Micron account for over 90% of global memory capacity. In pursuit of higher profits, these three manufacturers have proactively adjusted capacity allocation, directing 70% to 80% of their advanced wafer capacity entirely to AI-specific memory products such as HBM and high-end server DDR5.
HBM high-bandwidth memory has complex processes. The wafer area consumed by a single chip is equivalent to that of three regular DRAM chips, significantly reducing the number of chips that can be produced under the same capacity. Manufacturers prioritizing the production of high-margin AI memory directly compresses the production share of general-purpose memory for mobile phones, computers, and vehicles, further amplifying the market shortage problem.
3. Continuous Intensification of Global Computing Power Construction Grants Demand Long-Term Rigidity
Overseas cloud vendors such as Microsoft, Google, and Amazon, along with major domestic computing power platforms, are continuously expanding their data centers and have all signed three- to five-year long-term memory supply agreements to lock in capacity for the coming years. Institutional calculations show that the storage increment brought by AI from 2026 to 2028 will maintain a growth rate of over 45% annually, while the global memory supply growth rate during the same period is only 7% to 16%. The supply-demand gap continues to widen.
III. Why Won't New Capacity Be Concentrated and Released Until 2028? Three Hard Cycles Cannot Be Avoided
Many investors may wonder: since the profit margin for memory is sufficient, why don't major manufacturers immediately build factories on a large scale to expand production and alleviate the market shortage? The core reason is that semiconductor manufacturing has multiple rigid time barriers, and no company can skip the cycle to quickly land capacity.
1. Landing Cycle for New Wafer Factories Is as Long as 2-4 Years
For a 12-inch memory wafer factory, from land construction, cleanroom setup, and semiconductor equipment installation, to process debugging and yield ramp-up, the complete cycle requires at least 24 months. The construction cycle for some high-end HBM-exclusive production lines is close to four years.
For the expansion projects launched by major manufacturers in 2026, the equipment delivery cycle alone takes 18 months. Coupled with long-term process debugging, all new large-scale capacity will uniformly achieve stable shipments between late 2027 and 2028. The market will maintain a tight supply pattern in both 2026 and 2027.
2. Bottlenecks in the Supply of Supporting Semiconductor Equipment and Materials
Memory production lines require hundreds of supporting equipment and materials, including lithography machines, etching equipment, and specialty electronic gases. Currently, global semiconductor equipment capacity is equally saturated, with equipment manufacturers' order backlogs fully loaded, making it impossible to quickly deliver new equipment. Even if memory companies secure funding to plan new fabs, they cannot proceed with production without core equipment, further extending the capacity expansion cycle.
3. Manufacturers Learn from Past Losses and Reject Disorderly and Aggressive Expansion
From 2022 to 2023, the memory industry experienced a deep downturn, with the three major overseas original manufacturers suffering combined losses exceeding 100 billion yuan. The industry has completely abandoned the old model of blindly expanding capacity to seize market share. Today, all capital expenditure plans are tied to long-term orders from leading customers. There will be no blind construction of new production lines detached from actual demand. The pace of expansion has become conservative, and there will be no concentrated burst of short-term capacity.
Multiple foreign institutions uniformly predict that 2027 will be the year with the largest supply-demand gap in the memory chip market. It will not be until 2028, when a large number of newly built production lines achieve stable mass production, that the market shortage will be substantially alleviated.
IV. Objectively View Industry Opportunities: Four Types of Potential Market Risks That Cannot Be Ignored
Long-term tight supply and demand do not mean the industry will only rise and never fall. Whether you are a hardware practitioner or an investor focusing on the semiconductor sector, you need to rationally identify the variables in the market and avoid one-sided optimistic thinking.
First, the risk of downward revision in AI capital expenditure. The 5.6 trillion industry investment is a medium- to long-term estimate. If global technology companies cut their budgets for computing power construction, the growth rate of memory demand will slow down accordingly, directly affecting product price trends.
Second, competition brought by the continuous release of domestic memory capacity. CXMT and YMTC continue to expand production, and domestic niche memory manufacturers are constantly breaking through technologically. After 2028, when domestic capacity is implemented on a large scale, it will divert market share from overseas original manufacturers and compress the overall profit margin of the industry.
Third, the continued sluggishness in consumer electronics demand. Demand in the mobile phone and PC markets has remained sluggish for a long time, and general-purpose memory lacks sufficient support from terminal consumption. Once the incremental demand from AI fails to offset the decline in the consumer segment, the driving force for industry price increases will weaken.
Fourth, uncertainties brought by the iteration of high-end memory technologies. HBM and new memory technologies continue to iterate, and the value of old-process production lines will continue to shrink. If enterprises lag in the pace of technological upgrades, they will gradually lose market competitiveness.
V. Reconstruction of Industry Cycle Logic: Core Insights for Ordinary People to Understand the Memory Sector
In the past, people viewed memory chips simply as a strong cyclical sector, with market trends fluctuating along with the consumption demand for mobile phones and computers. However, the emergence of AI computing power has completely reconstructed the underlying logic of the industry.
Memory is no longer just a supporting component for consumer electronics; it has become an indispensable core infrastructure for the development of the AI industry. Demand has shifted from short-term cyclical consumption to long-term rigid demand for computing power. Companies like Beijing Ingenic, which focus on automotive-grade and industrial AI memory, avoid the fiercely competitive general-purpose track. Relying on certification barriers in segmented tracks and stable long-term orders, they can navigate short-term industry fluctuations.
For practitioners in the real economy, signing long-term supply agreements with original manufacturers in advance to lock in memory procurement prices and capacity can effectively hedge against the cost pressure brought by rising raw material prices. For investors focusing on the semiconductor sector, they should not just stare at short-term price increase trends. They need to combine the capacity release cycle and downstream demand structure to judge the medium- to long-term growth space of enterprises and avoid stocks driven purely by thematic speculation.
Conclusion
Based on feedback from industrial chain enterprises, global AI capital expenditure plans, and wafer fab expansion cycles, the current shortage of memory chips will persist. The incremental memory demand brought by massive computing power construction cannot be met by sufficient capacity in the short term. The inflection point of industry supply and demand is clearly locked in 2028.
AI has completely changed the development logic of the memory industry. The cyclical market trends once driven by consumer electronics have now switched to a long-term growth track driven by computing power. I would like to ask everyone reading this: when you come into contact with digital hardware in your daily life, have you felt the cost increase brought by rising memory prices? Do you think that after the concentrated release of capacity in 2028, the memory industry will return to a downward price cycle? Welcome to leave your views in the comment section and objectively discuss the development trends of the semiconductor industrial chain together.