Today, I learned something...

Table of Contents

• 1. DRAM Shortage: Why Prices are Rising
  • 1.1 Background
  • 1.2 Cyclicality of the Memory Market
  • 1.3 Current AI-driven Cycle
  • 1.4 In a Nutshell
  • 1.5 Impact on Consumers
  • 1.6 Some Statistics
  • 1.7 Conclusion

1. DRAM Shortage: Why Prices are Rising

A colleague mentioned today that there is an acute shortage of DRAM in the market, and that he is concerned about increasing prices for consumer tech like smartphones and laptop. I decided to dig a little deeper, and found this article: SemiAnalysis

1.1 Background

DRAMs have gotten cheaper over the years because of increasing technological innovation. This innovation was mostly due to two laws: Moore's Law (more and smaller transistors) and Dennard Scaling (power density remains constant, even as transistors shrink in size). As a result of these innovations, shrinking transistors reduced cost per bit and capacitor engineering kept sufficient charge to maintain signal integrity (ensuring power density is constant).

These innovations have mostly become limited. The article quotes a statistic: "Over the past decade, DRAM density has increased by only ~2× in total, versus roughly ~100× per decade during the industry’s peak scaling era." As a result, the once secondary constraints are now the main bottlenecks: bitlines and sense amplifiers.

The article explains these bottlenecks with a bucket analogy: Imagine a DRAM cell is a tiny bucket that holds electricity instead of water. Each bucket stores a bit of data by holding a small electrical charge. These buckets, over the years, have been made smaller to fit more memory on a chip. But, we've now reached a point that these buckets have become too tall and narrow. So, these buckets now hold very few electrons.

The problem is that when the system tries to read the data, it has to detect this faint electrical signal (due to few electrons) and distinguish it from noise. The signal is transmitted through wires that connect these cells (bitline) and read through tiny sensors (sense amplifiers). The signal has become so weak that even small variations in manufacturing or temperature can cause errors.

Thus, Dennard Scaling no longer holds, and Moore's Law has also slowed dramatically for DRAM. The physical limits of shrinking these structures mean cost-per-bit reductions have also slowed down. Now, DRAM pricing is no longer as dependent on technology-driven cost reductions, but more on capacity additions and cyclical supply-demand dynamics.

1.2 Cyclicality of the Memory Market

Next, the article describes in some detail as to how the memory industry (and consequently, DRAM pricing) is driven by cyclicality. You can read the article for the full detail, but the overall gist is: supply is inelastic, but demand is extremely volatile. It takes years to build a new DRAM supply line, but the demand for it can fluctuate daily.

The demand growth can be relatively stable, reflected by small increments in memory content per device. However, during "inflection periods" — when new computing platforms or architectures emerge — memory demand can grow explosively.

Generally, the sequence in such "inflection periods" panned out like this:

1. Unexpected demand catches memory suppliers off guard;
2. Prices increase because of the unexpected demand;
3. This results in elevated profitability for the suppliers;
4. The memory suppliers are more willing to invest in capital and expand capacity aggressively;
5. This results in oversupply and subsequent market corrections (supply increases beyond demand, prices decrease to remain competitive)

Some examples the articles cite:

1. 1993 Windows PC supercycle: Windows switched from text to graphical interfaces, which needed 4x more memory per PC. Supply was constrained after a bad downturn in the late 1980s. Prices surged, margins exceeded 50%. Then everyone built new fabs, 50 fab construction plans were announced in 1995-96 alone, and prices crashed over 60%.

2. 2010 smartphone/cloud cycle: iPhones and Android phones exploded in popularity while cloud computing (Google, Amazon, Facebook data centers) took off. Both needed lots of memory. But this cycle was shorter and weaker because mobile memory (LPDDR) was cheaper and more price-sensitive. Prices peaked and fell quickly.

3. 2017-2018 server supercycle: Data centers needed way more memory per server for virtualization and cloud workloads. Companies made record profits. Then supply caught up, demand normalized, and prices crashed again.

4. COVID-era cycle (2020-2021): Lockdowns caused a sudden surge in demand for PCs, servers, and consumer electronics. Companies panic-ordered, placing years' worth of orders at once ("double and triple ordering"). Supply was constrained by labor shortages and logistics disruptions. This cycle was important because it taught the industry to be more disciplined about spending, setting the stage for what's happening now.

Also, one thing to note: "Industry has consolidated from 20+ suppliers in the 1990s to just 3-4 today (Samsung, SK Hynix, Micron), which means fewer players controlling supply." There are now simply fewer companies that can build new capacity, making the subsequent shortages harder to resolve (unless more fabs are built by more companies).

1.3 Current AI-driven Cycle

The current cycle is driven by AI, and the article argues that this cycle is different and predicts that it will be bigger and longer than the previous cycles. Key reasons include:

• AI systems (like those training large language models) need massive amounts of a special type of memory called HBM (High Bandwidth Memory), which is essentially multiple DRAM chips stacked on top of each other. HBM is much more expensive and complex to manufacture than regular DRAM.
• To make more HBM, manufacturers are reallocating their existing factory capacity away from regular DRAM. This creates a shortage of regular DRAM too, even for normal computers and phones.
• The HBM market is projected to grow from about $35 billion in 2025 to $100 billion by 2028, a roughly 40% annual growth rate.
• DRAM prices have risen 171% year-over-year, and DDR5 spot prices have quadrupled since September 2025.
• New fabs take years to build and billions of dollars to construct, so supply can't catch up quickly.
• The supply-demand imbalance is actually getting worse, not better.

The article suggests that this cycle can persist through 2027-28, potentially making it the longest supercycle in the industry's history.

1.4 In a Nutshell

AI's hunger for memory chips is so intense and enormous (not to mention, sudden) that the entire memory industry cannot build fabs fast enough (supply) to keep up with demand. This is creating a shortage that's driving prices for DRAM through the roof. This supercycle can potentially be the biggest in the memory industry's 50-year history, and we have not peaked yet, i.e., the prices for DRAM will keep increasing.

1.5 Impact on Consumers

What can be the impact on regular consumers? I found this article that explains the downstream effects: TechSpot Article.

Most consumer tech companies are scrambling to respond; they are under significant pressure to increase the prices of their products or reduce their profit margin. The computing giants (such as Amazon, Google, Microsoft) have locked multiple years contract with the memory industry suppliers. Understandably, the fabs are directing manufacturing for AI data centers (these deals are more lucrative) rather than towards general consumer tech. Some of the effects can be:

1. Phones, laptops, smart TVs, game consoles, or anything with DRAM or flash storage can cost more. That includes products of Dell, Lenovo, Apple, Samsung, and other similar consumer tech companies.
2. Some companies, instead of raising prices, may ship devices with less RAM to keep prices down.
3. Cloud services, streaming, and SaaS products all run on servers that need DRAM. As those costs rise, companies will eventually pass them on through subscription price increases or reduced free tiers.

1.6 Some Statistics

I gleaned some statistics and quotes from the above article + more recent articles, such as this one: Bloomberg: Rampant AI Demand for Memory Is Fueling a Growing Chip Crisis:

• Analysts at TrendForce expect average DRAM prices, including HBM variants, to rise between 50 and 55 percent in the final quarter of 2025 compared with the previous three months.
• Samsung, which along with SK Hynix controls more than 70 percent of the market, recently raised some memory prices by as much as 60 percent.
• Daniel Kim of Macquarie estimates price hikes in consumer electronics could reach between 10 and 20 percent in 2026.
• Citi's Peter Lee predicts chip stockpiling will get worse through 2026, not better.
• Phison CEO Khein-Seng Pua says that low-cost consumer devices are most vulnerable because they live and die by razor-thin margins
  • Example: Cost for an 8 GB eMMC component — a basic embedded storage part found in plenty of cost-controlled devices — rose from $1.32 in early 2025 to roughly $17.60 more recently.
• Ford CFO says DRAM Shortage may impact vehicle pricing.
• Sony Group Corp. is now considering pushing back the debut of its next PlayStation console to 2028 or even 2029.
• Nintendo Co. is contemplating raising the price of new Switch 2 console in 2026.
• Cisco Systems Inc. cited the memory squeeze when it gave a weak profit outlook last week that led to its worst share loss in nearly four years.
• Qualcomm Inc. and Arm Holdings Plc both warned of more fallout ahead.

1.7 Conclusion

I guess the only conclusion is to brace for consumer tech prices to increase. It is likely that prices will only increase at least for the next 2-3 years. Buying now might be cheaper than buying later.