Economic Daily: AI demand squeezes capacity, memory chip price increases transmit to consumer electronics

The main body of the article

Newly purchased laptops have become more expensive, and the flagship models from phone manufacturers have also gone up in price—the main force behind this is the battle for memory chip resources driven by artificial intelligence (AI), and the price changes you and I feel are merely the cost at the very end of this "war of electronic devices."

The global memory chip market is now showing "fire and ice": on one hand, South Korean semiconductor chip manufacturer SK Hynix raised $26.5 billion through an American Depositary Shares (ADS) offering, making it the largest initial public offering (IPO) by a foreign company in the United States; Samsung Electronics' quarterly operating profit surged 18-fold, and tech giants are stepping up fundraising to expand memory chip production. On the other hand, Apple has been forced to raise prices on multiple products because of the chip shortage, and Lenovo and Dell have followed suit. The production side is booming, yet the consumer side has no chips to use—supply is clearly expanding, so why can't it fill the demand gap?

To unravel this mystery, we first need to meet the protagonists behind the scenes and understand the three basic types of memory chips: dynamic random access memory (DRAM), flash memory (NAND), and high bandwidth memory (HBM).

To use a simple analogy: DRAM is like a device's "workbench," responsible for temporarily processing the data at hand; NAND is the "file cabinet," where photos on phones and documents on computers are stored for the long term. These two types of chips mainly serve the phones, computers, and tablets we commonly use.

HBM can be seen as a super-enhanced version of DRAM. It stacks multiple DRAM chips vertically, achieving extremely fast data throughput. Therefore, in most cases, it does not directly serve ordinary consumers, but rather AI chips. If the graphics processing unit (GPU) is the brain of AI, then HBM is the highway that delivers data to the brain—if the road is not wide enough or fast enough, even the smartest brain can only sit idle.

The tipping point came from the explosion of generative AI in recent years. Training a large model with 100 billion parameters requires tens of thousands of GPUs running day and night, and every advanced GPU must be paired with several HBM units as a "close-range data warehouse." How fierce is this demand explosion? By 2026, the global HBM market is expected to exceed $60 billion. SK Hynix and Micron have publicly announced that their HBM capacity for 2026 is completely sold out, and customers have even locked in supply for 2027 and even 2028 in advance.

In an extremely tight seller's market, HBM prices continue to soar. According to a report by the internet data information network DigiTimes, the market unit price of next-generation HBM4 in the second half of 2026 will be about $2/GB (gigabyte), and by 2027 it is expected to rise to $4/GB to $5/GB, or even higher. JPMorgan recently released a research report estimating that the global average selling price of HBM will rise 54% year on year in 2027 and further rise 25% in 2028. It is precisely this huge profit that has prompted major memory manufacturers such as Samsung, SK Hynix, and Micron to urgently "switch tracks" in unison—shifting large volumes of advanced capacity originally used to produce memory for phones and computers to HBM.

What is even more critical is that HBM is inherently capacity-hungry: its chip area is 2 to 3 times that of ordinary memory of the same capacity, which means that under limited wafer capacity, for every 1GB of HBM produced for AI, about 3GB of general-purpose memory output must be given up. This is not a simple reallocation of capacity, but an irreversible capacity crowding-out driven by excess profits.

At present, about 70% of the world's advanced memory capacity has been swallowed by AI data centers, and consumer electronics such as phones and computers can only compete for the remaining 30%. The extreme imbalance between supply and demand is being transmitted layer by layer along the industrial chain: in the first half of this year alone, contract prices for general-purpose memory nearly doubled, and flash memory rose by more than 50%. Calculations by several investment banks show that this year the supply-demand gap for various memory chips generally exceeds 4%, all hitting new highs in more than 10 years.

Consumer electronics devices are the real victims of the wave of price increases. Memory chips' share of the total material cost of a smartphone has soared from the previous 10% to 15% range to more than 30%, and for some flagship models with large storage it is even approaching 50%. The procurement price of a single 12GB phone memory chip has risen from 200 yuan to nearly 600 yuan, and 1TB flash memory chips have also soared from just over 200 yuan to 600 yuan. No wonder Apple described the current supply-demand tension as a "once-in-a-century flood."

The memory chip industry has always had strong cyclicality. It is the most cyclical category in the semiconductor industry, with highly standardized products and strong substitutability, carrying attributes similar to commodities, and prices are extremely sensitive to changes in supply and demand. Over the past 10-plus years, the industry has experienced a price cycle every three to four years; almost every surge stemmed from short-term mismatches, and every plunge stemmed from manufacturers expanding production in unison. However, this cycle is completely different from previous ones, and its particularity is deeply rooted in the structural rupture brought by AI.

First, the nature of demand has changed. The driving force of past cycles was consumer electronics replacement waves and capacity upgrades, such as the explosive growth of smartphones and the increase in memory capacity of Android phones; after reaching the demand ceiling, prices would naturally fall back. This time, however, the core engine is AI training and inference, which tech giants regard as the most important strategic race of the next 10 years and are willing to compress short-term profits to preserve computing power investment. HBM orders thus show "infrastructure-like" rigidity—as long as large models keep getting bigger and AI applications keep expanding, strong demand for it will be hard to fade.

Second, capacity flexibility has disappeared. After traditional memory capacity expansion, it can be converted into supply relatively quickly, and production lines can also be flexibly allocated among different product categories. But HBM production relies heavily on advanced packaging, and yield ramp-up is difficult. A newly built dedicated wafer fab takes at least two years from construction to mass production, and once in production its lines are almost irreversibly locked in; even if demand changes, they cannot easily be switched back to producing memory for phones or computers. In other words, supply is becoming "hard and slow," essentially losing its ability to adjust to cyclical fluctuations.

Third, the pricing mechanism has been distorted. At present, the HBM capacity of the world's major memory manufacturers for the next one to two years has been fully locked up in advance by cloud providers and device brands through multi-year long-term agreements, leaving almost nothing available for the spot market. This means contract prices have lost room for short-term adjustment, and the market's price discovery function has been undermined. The market has thus split into "two parallel worlds": AI-specific high-performance memory remains priced high and continues to earn excess profits; consumer-grade memory, forced to follow the rise because of the supply collapse, passes almost all of the price-increase pressure one-way onto ordinary consumers.

At present, this structural split is still intensifying—an unprecedented capacity-expansion race has begun, but almost all of the new capacity is aimed at HBM and server DRAM, not the consumer-grade memory needed for phones and computers. The three major memory manufacturers—Samsung, SK Hynix and Micron—will spend a combined total of more than $50 billion in capital expenditure in 2026, with Samsung alone investing more than 110 trillion won. However, because HBM crowds out wafer capacity, supply of consumer-grade DRAM and NAND may not ease but may instead tighten further.

Today, the memory industry is standing at a new crossroads. As all manufacturers go all out to expand capacity, and as newly built wafer fabs are put into production one after another after 2027, will this AI-driven "super cycle" become the longest boom in the history of the memory industry, or will it once again become a cruel footnote to the idea that "the cycle never disappears"? Or will a "third script" emerge—HBM and consumer-grade memory parting ways from then on, with one continuing to boil under the push of AI while the other, under long-term suppression of capacity, bids a final farewell to the era of cheapness? On these questions, the industry has yet to reach a clear conclusion. The future has arrived, but uncertainties remain.

(Source: Economic Daily) Mid-article embedded advertisement