The arrival of new memory factories will not quickly ease the shortages affecting servers, computers, and electronic devices. An analysis attributed to Bank of America and reported by specialized media suggests that SK hynix may only add about one-sixth of the additional capacity the market initially expected by 2028.
This estimate does not come from a public forecast by the company and should be considered an analyst’s projection, not a confirmed schedule. Still, it aligns with the message conveyed by SK hynix’s CEO, Kwak Noh-jung: 2027 could be the most challenging year for memory supply, and demand is expected to continue outstripping supply beyond 2030 despite announced investments.
The key points of SK hynix’s capacity in 20 seconds
- The supposed addition of only one-sixth of the capacity projected until 2028 is an estimate, not an official guide.
- SK hynix anticipates an especially intense shortage during 2027.
- Demand for HBM for AI accelerators is competing with conventional DRAM.
- Building a factory does not mean production begins immediately.
- New plants require land, electricity, water, clean rooms, and specialized equipment.
- South Korea aims to double its DRAM production within five years.
- Samsung and SK hynix announced four new factories in the southwestern part of the country.
- Many of these projects will not contribute significant capacity before the next decade.
- Samsung, SK hynix, and Micron face a class-action lawsuit over alleged market manipulation.
- Industrial delays may support the plaintiffs’ narrative but do not by themselves prove price collusion.
South Korea unveiled an investment program at the end of June totaling 800 trillion won, roughly $518 billion, including two new Samsung factories and two more from SK hynix in the southwest. Samsung has indicated Gwangju as the location for its future complex, while SK hynix still needs to finalize the site and secure the necessary infrastructure.
The political goal is to double national DRAM production in five years, but the timelines for these new facilities do not align easily with that target. SK Group’s president, Chey Tae-won, noted that developing the Yongin complex took nine years and that any new factory requires large amounts of land, electricity, water, and skilled workers. The announced centers in the southwest are seen as long-term investments, with significant capacity likely during the 2030s.
Why does it take years for a new factory to start producing memory?
Major investment announcements are often expressed in cumulative figures over a decade or more. This money does not translate directly into available wafers.
Before installing a single machine, manufacturers must acquire land, obtain permits, prepare electrical connections, and ensure sufficient water supply. Construction includes clean rooms with strict control over particles, temperature, humidity, and vibrations.
Next comes the installation of lithography, deposition, etching, metrology, and inspection tools. Some have delivery timelines of several years and require supplier personnel for commissioning. A modern factory does not operate as a single line: wafers go through hundreds of steps and are repeatedly processed by different equipment.
Commercial production begins gradually. The initial batches help adjust processes and improve yield, i.e., the percentage of usable chips obtained from each wafer. Reaching stable volume may take several quarters.
Therefore, it is useful to distinguish three concepts often confused:
| Concept | What it represents |
|---|---|
| Announced investment | Planned capital for land, buildings, machinery, and infrastructure |
| Wafers capacity | The number of wafers that can be processed per month |
| Bits produced | The actual amount of usable memory obtained after fabrication and testing |
A company can reduce its monthly wafer capacity while manufacturing more memory if it uses a process that allows more bits per wafer. It can also shut down old lines while opening newer, more advanced ones, so net capacity may grow less than the size of the new facility suggests.
This is one of the points in the analysis attributed to Bank of America. Expansions must account for the retirement or modernization of old plants, transition periods, and time needed to improve process yields.
This context helps explain why manufacturers can announce massive investments while still anticipating shortages. The market needs memory during 2026 and 2027, but a significant part of the new factories will not be available until several years later.
SK hynix is also considering locations outside South Korea for future wafer plants, including the U.S., Japan, and Southeast Asia. These options are still under evaluation and respond to needs beyond the current price cycle.
HBM is shifting the distribution of DRAM factories
The growth of AI has not only increased demand for memory but has also changed which types of memory are prioritized for manufacturing.
HBM used with GPUs and other accelerators is built by stacking multiple DRAM chips and connecting them via vertical links. Each package can incorporate numerous layers, along with advanced bonding, testing, and encapsulation processes.
This results in much higher bandwidth than conventional memory but requires more silicon and a more complex manufacturing chain. Manufacturers allocate their most advanced lines to products needed by Nvidia, AMD, cloud providers, and major data center operators.
This decision impacts memory for traditional servers, computers, and consumer devices. Although HBM and DDR are not interchangeable products, they share part of the industrial capacity, tools, and manufacturing expertise.
The lawsuit filed in California claims that Samsung, SK hynix, and Micron allegedly used the shift toward HBM as a cover to reduce DDR3 and DDR4 production in a coordinated manner. Plaintiffs argue that this strategy contributed to roughly a 700% increase in the price of conventional DRAM since 2022. These allegations are still unproven in court.
The move toward HBM also reflects a clear commercial logic. Products for AI accelerators offer higher margins and are sold through supply contracts with major clients. Maintaining capacity for older memories may be less attractive once companies have pending orders for advanced products.
That does not mean all DDR scarcity stems from HBM. Line closures, process shifts, increased server memory demand, and manufacturers’ caution following previous oversupply cycles also play roles.
The memory sector has traditionally alternated periods of shortages and high prices with capacity expansions that eventually lead to surpluses. The current problem is that capacity buildup takes longer than the growth in demand driven by AI.
Some investors fear that the factories planned for 2027 and 2028 will create oversupply. SK hynix management remains optimistic, believing that even these expansions will not be enough to meet market needs in the coming years.
For infrastructure buyers, this uncertainty complicates planning. A data center operator must decide now how much memory they will need over several years, without knowing if prices will stay high or if a massive capacity increase will lead to a correction.
PC and mobile manufacturers face similar challenges. They can reduce RAM content, raise prices, or accept lower margins. In AI servers, memory is no longer a secondary component: it determines how many models can be loaded, the size of contexts, and the ability to handle multiple inferences simultaneously.
Delays do not prove price manipulation
The new class-action lawsuit was filed on 06/25/2026 in the Northern District of California. Its 17 plaintiffs accuse Samsung, SK hynix, and Micron of coordinating the supply and pricing of conventional DRAM since 2022. They seek class-action status, damages, and measures to halt the complained practices.
Information about the slow capacity growth can benefit the plaintiffs by strengthening two parts of their argument: substantial barriers to new competitors entering the market, and the high level of control maintained by the three leading manufacturers over global supply.
However, a factory taking ten years to complete does not prove illegal collusion. It could also reflect the real limitations of a capital-, energy-, water-, and labor-intensive industry.
This difference was decisive in a previous case. A 2018 lawsuit accused the same companies of coordinating production cuts. The case was dismissed, and the decision confirmed in 2022, because the alleged conduct could be explained by independent business decisions and did not sufficiently demonstrate an agreement among competitors.
The new complaint aims to overcome this hurdle by using the simultaneous shift to HBM as an additional element. The court will need to determine whether companies made parallel decisions because of shared economic incentives or if some form of coordination occurred.
SK hynix delays could serve both sides. Plaintiffs might present them as evidence of continued supply restrictions, while the defense could argue that rapid capacity expansion was technically impossible without any collusion.
The figure of one-sixth of capacity is striking but does not alone resolve this debate. First, it would be necessary to understand what initial capacity the estimate considered, which factories are included, how much results from shutting old lines, and how final production of bits is calculated.
The clearest conclusion for the industry is not legal but industrial: additional memory capacity will arrive later than what customers need. As AI investment continues to grow, SK hynix, Samsung, and Micron will hold considerable power to decide what products they produce and which clients they serve.
Frequently Asked Questions
Has SK hynix confirmed it will only add one-sixth of the planned capacity?
No. The figure comes from an analysis attributed to Bank of America and disseminated by specialized media. The company has not incorporated this into an official public forecast.
Why does HBM manufacturing affect DDR memory?
Both products use DRAM chips and share part of the industrial infrastructure. Prioritizing HBM can reduce resources available for conventional memories.
When might memory prices drop?
There is no certainty. Some analysts expect more capacity between 2027 and 2028, while SK hynix believes demand will continue to outstrip supply beyond 2030.
Does demand imply that manufacturers have manipulated prices?
Not necessarily. The lawsuit contains allegations that need to be proven. Price increases and parallel production decisions alone are not sufficient to establish illegal collusion.
