ADATA President Chen Li-bai believes that the global demand for memory will continue to outpace supply over the next ten years, even if Samsung Electronics, SK hynix, Micron, and Chinese manufacturers expand their capacity. This statement represents one of the industry’s most aggressive forecasts and should not be confused with a consensus estimate for all types of DRAM and NAND Flash.
30-Second Key Points on Memory Shortage
- ADATA believes data centers and AI-enabled devices will see demand grow faster than production for another decade.
- HBM, DDR5 for servers, and enterprise SSDs are shifting capacity from computers, mobile devices, and traditional products.
- It takes years to bring new factories into production, and some capacity is already reserved through long-term contracts.
- TrendForce agrees on short-term tension but expects overall NAND shortages to ease starting in the second half of 2027.
Chen argues that it’s still too early to talk about an AI bubble. His view is that the market is focusing too much on immediate investments by major cloud providers and not enough on the future expansion of AI into businesses, governments, homes, vehicles, robots, factories, and peripheral devices.
ADATA’s position provides valuable insight into channel orders and pricing, although it needs context. The company is a leading global provider of branded DRAM modules and SSDs but does not manufacture the advanced memory wafers integrated into its products. Its forecast reflects the opinion of a particular segment of the supply chain, not an independent estimate of global capacity.
AI Is No Longer Solely Dependent on HBM
Discussions about shortages often focus on high-bandwidth memory, known as HBM, used alongside NVIDIA, AMD, and other accelerators. However, an AI data center requires many more layers of memory and storage.
Servers incorporate large amounts of DDR5 and RDIMM modules for CPUs. Models, datasets, vector bases, and caches are stored in enterprise SSDs. Additionally, agent systems maintain longer contexts, repeatedly call tools, and generate information that must be preserved.
TrendForce estimates that the growth of agent-based AI is causing a structural demand expansion. The firm increased its global memory market forecast for 2027 to over $1.28 trillion, with $903.3 billion from DRAM and nearly $379.4 billion from NAND Flash. These revenue estimates are driven not only by volume but also by sharp price increases, not just the number of chips sold.
Manufacturing HBM adds pressure because it requires large chips, complex stacking processes, and more wafer capacity per useful bit than conventional DRAM. TrendForce anticipates that HBM production will account for about 30% of the total wafer input from the top three manufacturers by the end of 2027, though it will contribute only around 13% of total bits produced.
This gap helps explain why HBM growth may indirectly reduce the availability of DDR5, LPDDR, and other memories. Manufacturers prioritize their most advanced processes for the most in-demand and profitable products, while computer, mobile, and industrial electronics customers compete for remaining capacity.
Something similar occurs with NAND. Suppliers are prioritizing large-capacity enterprise SSDs for data centers. During Q2 2026, TrendForce expected contractual NAND prices to rise between 70% and 75% quarter-over-quarter, even as PC and smartphone manufacturers cut capacity to keep costs down.
| Segment | Main Demand Source | Projected Situation |
|---|---|---|
| HBM | GPU and AI accelerators | Strong pressure until 2027 |
| DDR5 and RDIMM | Servers and data center processors | Limited supply and high prices |
| LPDDR | Mobile devices, laptops, and low-power servers | Competing for capacity with AI products |
| Enterprise NAND | SSD for AI, analytics, and cloud | Shortages throughout 2026 |
| Consumer NAND | PCs, mobiles, and devices | Lower priority for manufacturers |
| Mature SLC and MLC | Automotive, industrial, and embedded systems | Risk of prolonged structural shortage |
Peripheral demand further supports ADATA’s thesis. Markets such as software-defined vehicles, robotics, smart homes, satellites, automated factories, and industrial equipment require localized memory, even if some processing occurs in the cloud. Not all these markets will grow at Chen’s predicted rate, but they expand the problem beyond large data centers.
Building Factories Won’t Solve Shortages Overnight
Samsung, SK hynix, and Micron are increasing investments, but a memory factory takes several years to build, install equipment, validate processes, and reach acceptable manufacturing performance.
Micron announced in July that it will increase investment in the U.S. to over $250 billion through 2035. The company has just begun vertical construction on its new York plant and aims to produce around 40% of its DRAM in the U.S. long-term. This capacity will not hit the market all at once.
Moreover, part of future production is being secured in advance through long-term supply agreements. Micron has strategic supply deals with Ford and General Motors for DRAM, NAND, and other memories over extended cycles. SK hynix also has multi-year agreements with NVIDIA to develop and supply upcoming generations of memory for its AI platforms.
These contracts improve visibility for manufacturers and customers but reduce the amount available for spot purchases. A new plant can boost global capacity but won’t necessarily prevent shortages in the short term if early years are already committed.
Chen also believes that major manufacturers will avoid repeats of the disorderly expansions of previous cycles. Memory has historically suffered from capacity excess, price collapses, and significant losses. Companies are incentivized to grow capacity cautiously and protect their margins rather than flood the market.
Process improvements can add more bits without significantly increasing wafer counts. Manufacturers can add layers to 3D NAND, improve yields, or migrate to QLC cells. In DRAM, they can introduce denser nodes. These measures increase production but also require halting and adapting lines, so their immediate effects are limited.
The Ten-Year Outlook Conflicts With Potential NAND Relief
The main caution regarding ADATA’s forecast is that “memory” encompasses different markets. A HBM shortage may coexist with a sufficient supply of consumer NAND. Additionally, advanced product shortages may occur while older chips used in automotive or industrial applications remain scarce.
TrendForce agrees that 2026 will likely remain tight, but its outlook for NAND is less severe. The firm expects bit supply growth to start surpassing demand in the second half of 2027, thanks to process improvements, gradual line expansions, and weakening smartphone and laptop markets.
This doesn’t necessarily contradict Chen’s thesis entirely. A market can move from deficit to a small surplus while maintaining high prices, low inventories, and specific category shortages. It also does not imply that DRAM will follow the same timeline as NAND.
Differences can be summarized as follows:
| Forecast | Timeline | Scope |
|---|---|---|
| ADATA | Approximately ten years | Memory broadly, driven by cloud and edge computing |
| TrendForce on DRAM | Short-term tension and rising prices through 2027 | HBM, DDR5, and server memory |
| TrendForce on advanced NAND | Possible improvement from late 2027 | Global bit supply-demand balance |
| TrendForce on mature SLC, MLC, and NOR | Prolonged shortages | Mature products for industrial, automotive, and embedded markets |
| Micron | Conditions tightening beyond 2026 | DRAM and NAND, but not over a full decade |
For mature products, shortages may last longer. TrendForce estimates that worldwide MLC NAND capacity will decrease by 41.7% during 2026, as several major providers exit or limit that technology. Expanding advanced NAND for AI SSDs will not automatically resolve the shortage of older chips still used in long-cycle equipment.
ADATA’s outlook has also grown more cautious over recent months. In March, Chen said that both DRAM and NAND would be scarce throughout 2026 but avoided predictions for 2027 or 2028. The new ten-year forecast underscores how much demand expectations have increased but also highlights the uncertainties inherent in such long-term predictions.
For companies and data centers, the takeaway should not be unlimited inventory hoarding. It’s more prudent to secure critical capacities, diversify suppliers, review refresh cycles, and design systems that do not rely on a single module or SSD generation.
Buyers should also differentiate risks by product. A server using standard DDR5 faces a different market than one with HBM4, an enterprise QLC SSD, or industrial SLC memory. Talking about a single “memory shortage” can obscure significant differences in availability, pricing, and substitution options.
ADATA’s forecast envisions a scenario where memory and electricity become the main physical constraints on AI expansion. The shortages in 2026 and 2027 support this view. Sustaining an uninterrupted global deficit until 2036 would require demand to grow faster than all new factories, efficiency improvements, and Chinese manufacturer expansions. While possible, that outcome remains far from certain.
Frequently Asked Questions
Has ADATA stated that DRAM and NAND shortages will last for ten years?
Their president referred to a prolonged broad memory shortage. The statement is often summarized as a decade of DRAM scarcity but does not provide a separate quantitative forecast for each DRAM and NAND category.
Why does HBM reduce the supply of conventional memory?
HBM uses large chips and complex stacking processes that consume more wafer capacity per bit. Prioritizing HBM means fewer resources are available for DDR5, LPDDR, and other types of memory.
When might NAND availability improve?
TrendForce expects the overall supply-demand balance to start improving in the second half of 2027. Some industrial and mature categories may remain scarce for longer.
Will new factories lower prices?
They will increase capacity but take years to reach significant output. Much of their volume is already under long-term contracts, so immediate price drops are not guaranteed.
Sources:
- ctee and statements by Chen Li-bai on memory demand trends, reported by tech media.
