Why have DDR5 memory prices skyrocketed?

01

DRAM prices surge

In November, the most notable event in the semiconductor industry was likely the official entry of DRAM and flash memory prices into a new upward cycle. While there has been no shortage of related news (such as "LPDDR supply tightens, AI drives prices to surge"), it is still necessary to further analyze the underlying reasons from an industry perspective.

Previously, the three major DRAM manufacturers—SK Hynix, Samsung, and Micron—successively announced the cessation of DDR4 production. Subsequently, ChangXin Memory Technologies (CXMT) was also reported to be halting DDR4 wafer starts. This series of exit actions quickly turned DDR4 into a scarce commodity, leading to a sudden spike in market prices and the emergence of a rare "Bit Cross" phenomenon—where the price of the newer-generation DRAM (DDR5) fell below that of the older generation (DDR4) for the same capacity. Even now, DDR4 prices remain elevated, indicating that the four manufacturers' plans to phase out DDR4 production remain unchanged. Given the current supply-demand dynamics of DDR5, the discontinuation of DDR4 is essentially a trend-driven move, and there likely aren’t any more realistic alternatives.

What is truly surprising is the trend of DDR5. As previously discussed, DRAMeXchange tracked the spot price fluctuations of DDR4 (16Gb 3200) and DDR5 (16Gb 4800/5600) between April and June. Extending this timeline now reveals that DDR5 has once again experienced a significant surge, resulting in a new "Bit Cross" phenomenon with DDR4. More critically, the slope of this recent price increase is excessively steep, representing a clear abnormal fluctuation from an industry perspective.

Figure 1: DDR4/DDR5 Spot Prices. Source: Compiled by the author based on DRAMExchange data.

The price trend at the consumer end is even more striking. Taking Crucial, Micron’s consumer brand, as an example, the price trajectory of its DDR5-5600 16GB memory (model CT16G56C46U5) on Amazon Japan has been particularly extreme:

Before November 4th, the price remained below 10,000 yen. After that, it began surging, first to 11,480 yen, then to 16,880 yen. By early December, it had climbed to nearly 22,000 yen, and by December 6th, it reached 24,000 yen. In other words, the price nearly doubled in less than a month, and the product is currently showing as out of stock. Judging from historical trends in the DRAM market, this is clearly not a normal structural increase.

                        Figure 2: Price trend data obtained via Keepa

02

The current sharp rise

is close to "panic buying".

So, what triggered this round of price increases? There is no definitive information at the moment. Based on observed signs, this market trend differs from the earlier rise in DDR4 prices and instead resembles an industry-wide "panic buying" or "preemptive stockpiling." Similar to the rush for toilet paper during past oil crises or the scramble for masks at the beginning of the pandemic, there may not necessarily be an actual shortage in the market. However, emotionally driven hoarding in advance can create artificial scarcity.

Reported driving factors within the industry generally fall into several categories: First, major server manufacturers are jointly increasing their inventory levels of DDR5 memory and enterprise-grade NVMe SSDs to meet surging demand for AI servers. Second, GPU and AI processor manufacturers are ramping up their procurement of HBM (High Bandwidth Memory) to align with future delivery schedules. Third, some foreign media have reported that NVIDIA is "shifting to using LPDDR," positioning it as a customer on a scale similar to large smartphone manufacturers.

The last point is the most perplexing, but foreign media reports did mention that NVIDIA has "recently" shifted to using LPDDR. In reality, NVIDIA's Grace CPU already incorporates a significant amount of LPDDR5X, and its next-generation Vera CPU is also highly likely to continue using LPDDR. This shift is not a "recent" change. Nevertheless, as the AI market rapidly expands, it is entirely logical for companies to seek to secure the supply of key components as early as possible.

Against the backdrop of concentrated explosive demand for AI servers (including some edge-side AI servers), contracts for storage chips and memory in the coming year have been signed in large volumes and ahead of schedule. Initially, DRAM manufacturers celebrated the positive news of "next year's contracts locked in early with confirmed volumes." However, as the contracted quantities continued to pile up, manufacturers realized that these orders were approaching or even exceeding their existing production capacity. Resources for the consumer-grade and DIY markets were further squeezed out, leading to a rapid tightening of spot supply and a simultaneous surge in prices.

The ultimate outcome is that even Micron has decided to withdraw from its Crucial consumer-grade business, redirecting its limited production capacity entirely toward enterprise and AI server applications. This decision reflects genuine production pressures and explains why spot market prices have skyrocketed within such a short period.

03

Reanalysis of HBM/LPDDR/DDR Demand

However, the author's biggest question here is: While such large-scale advance purchasing is feasible, can GPUs actually be supplied as planned?

Returning to the logic chain behind this price increase, NVIDIA's next-generation GB300, Rubin/Rubin Ultra, or AMD's Instinct MI350 and MI400 series will all incorporate significant amounts of HBM3E/HBM4. Currently, the only suppliers capable of providing HBM3 and later specifications are SK Hynix, Samsung, and Micron, all of which utilize their most advanced DRAM processes to manufacture these products. Although the underlying DRAM cell structure remains the same, the DRAM die for DDR, LPDDR, and HBM are entirely different. Therefore, as demand for HBM continues to grow, the production capacity allocated to DDR/LPDDR is inevitably squeezed, making it understandable why DDR/LPDDR have become increasingly tight.

However, when it comes to HBM itself, supply plans must be locked in through long-term contracts between memory manufacturers and NVIDIA/AMD. Both parties sign corresponding memory contracts based on GPU production schedules, making it unlikely for HBM to experience "sudden demand spikes" in the short term. Even if a company suddenly announces, "We are doubling GPU production," memory manufacturers cannot temporarily increase HBM supply capacity. In other words, any changes in HBM demand should have already been factored into the plans of memory manufacturers—otherwise, it would be unusual.

Next, let’s look at LPDDR. As mentioned earlier, NVIDIA’s Grace and the upcoming Vera will both use LPDDR. Therefore, if the shipment pace of Grace/Vera accelerates, it is plausible for LPDDR supply to tighten. However, similar to HBM, LPDDR is directly soldered onto the board in a BGA package rather than shipped as modules like DDR5 DIMMs. Thus, this demand must also be secured through long-term contracts between NVIDIA and memory manufacturers. Logically, memory manufacturers should have already factored this shipment plan into their calculations.

Lastly, there's DDR. NVIDIA's Grace Blackwell or Vera Rubin architectures themselves do not use DDR5. However, for manufacturers offering solutions like the "HGX B200/B300 (Blackwell only, without Grace) + white-label x86 servers," DDR5 is essential. On AMD’s side, EPYC server CPUs and the Instinct MI300/MI400 configurations also rely on DDR5, meaning the number of players involved in DDR5 allocation is even larger.

In this context, AMD itself is not responsible for system integration (even for large-scale systems like Frontier, El Capitan, or more recent ones such as Helios and Herder, they are delivered in collaboration with global enterprise technology providers like HPE). Therefore, the procurement of DDR5 is handled by partners. If there has indeed been a sudden and significant increase in DDR5 demand, it could only suggest that certain developments have occurred along this supply chain.

04

Has the GPU supply been increased?

The scenario that immediately comes to mind is roughly as follows.

For instance, consider five partner companies, A through E, which initially received notifications from the GPU manufacturer stating that "each company will be allocated 10,000 GPUs per month until Q1 of next year." Based on this expectation, they formulated their business plans and proceeded with component procurement. However, the situation becomes entirely different if the GPU manufacturer later sends another notification indicating that "overall production capacity will increase by an additional 10,000 units per month moving forward."

How would companies A through E interpret this? While it remains uncertain whether the entire additional 10,000 GPUs could be secured by a single company, waiting until the actual allocation is confirmed only to realize that "components are insufficient, and servers cannot be shipped on schedule" would undoubtedly result in significant opportunity loss.

Whether the additional 10,000 GPUs will be entirely allocated to a single company or distributed among the five companies at 2,000 each remains unknown. However, from a practical business perspective, each company would naturally strive to secure as much of this additional share as possible and thus begin procuring extra components in advance. Compared to the cost of 10,000 GPUs, the expense of components is relatively low. It is only human nature to place additional orders despite the risk of failing to secure sufficient inventory through procurement, as the cost of missing out on this opportunity due to hesitation would be far greater.

Consequently, the procurement volume placed by companies A through E would no longer be limited to the components required to support servers for 10,000 GPUs. Instead, it would expand to a scale sufficient for 20,000 GPUs—essentially doubling. As a result, demand at the contract procurement level would suddenly surge dramatically. This scenario perfectly mirrors the script of the mask shortage during the pandemic: driven by fear of future scarcity, users (in this case, the partner companies) would compete to hoard inventory, plunging the entire supply chain into imbalance and chaos.

As seen in Figure 1, this round of price spikes indeed exhibits highly abnormal characteristics. Rather than a gradual buildup of demand based on planning, it appears to be a sudden, panic-driven surge in demand that has left memory manufacturers struggling to respond.

The discontinuation of the Crucial business can be seen as the most typical example. In other words, contract orders have suddenly accumulated to such an extent that even maintaining the production volume required to sustain the Crucial business has become unfeasible. Additionally, despite official denials, rumors such as "Samsung refused LPDDR contracts submitted by its smartphone division (reportedly due to higher external sales prices, which could improve Samsung’s overall profitability)" or reports like "In the past, NVIDIA typically provided consumer GPU manufacturers with a GPU + GDDR package, but now it has shifted to offering only the GPU itself (with GDDR supply risks and price pressures transferred to partners due to soaring prices and procurement challenges)" continue to circulate. These phenomena likely reflect the level of chaos brought about by this market turbulence.

05

Can the supply of GPUs really increase?

However, a straightforward question arises: Can NVIDIA and AMD truly supply enough GPUs to meet this sudden surge in demand? For now, this ultimately remains a structural issue tied to TSMC's N4/N3 process capacity. Unless TSMC significantly ramps up production at these nodes, this contradiction will be difficult to resolve.

Yet, based on currently available information, the situation does not seem to be heading in a positive direction. While it is undeniable that the operation of the Arizona fab in the U.S. has increased production capacity compared to the past, the reality is that even if chips can be manufactured in Arizona, the advanced "CoWoS (Chip on Wafer on Substrate)" packaging can only be handled in Taiwan, China. Therefore, the dies produced in Arizona must still be sent back to Taiwan, China, for packaging. As long as the capacity in this critical step cannot be expanded, the final supply of GPUs cannot substantially increase.

Based on the previous example, increasing production from 50,000 units per month to 60,000 may still be feasible within a certain range. However, scaling it up to 100,000 units is almost unrealistic at this stage, and signs of resolving this issue in the short term appear highly unlikely.

Therefore, the likely scenario that unfolds next can be imagined as nearly identical to the situation with masks during the pandemic. Partner companies, based on expanded contract volumes, procure large quantities of memory products, but the actual supply of GPUs fails to keep pace. As a result, memory products continue to accumulate as inventory. The critical question, however, is whether retaining such massive inventory across fiscal years is realistically viable—a challenge that could prove to be exceptionally harsh.

From this perspective, it can be inferred that companies might attempt to offload these accumulated inventories in various ways around the end of their fiscal years, leading to an

 influx into the market and subsequently causing a decline in prices. For companies operating on a calendar year as their fiscal cycle, the current inventory levels may not have

 reached particularly high levels (after all, while contracts have been signed, the physical delivery of DDR5 spot products may have only just begun). As a result, their financial

 reports for the 2025 fiscal year could likely maintain a status similar to the past. However, sustaining this into the 2026 fiscal year might prove considerably challenging,

 prompting these companies to explore ways to reduce inventory around that time. Conversely, it would be entirely unsurprising if prices remain elevated until that point.