1.5 Trillion US Dollars, Storage Giants Step In to "Prop Up the Market"

Memory prices are experiencing a sharp increase.

According to the latest forecast by US investment bank Jefferies, memory prices will rise by 40% to 50% quarter-on-quarter in the third quarter of 2026 and another 30% to 40% in the fourth quarter. Subsequently, prices will increase by 40% to 45% year-on-year throughout 2027, and the upward trend will not slow down until at least 2028.

On the supply side, the existing production capacities of the world's three major DRAM (Dynamic Random Access Memory) manufacturers, Samsung, SK Hynix, and Micron, are already stretched thin. Half of their current total production capacity is locked in by "long-term agreements."

Capacity expansion is also underway during the same period.

Micron has previously invested over $150 billion in new production capacities in markets such as the United States. Samsung Electronics and the SK Group recently jointly announced a ten-year investment plan focusing on semiconductor and artificial intelligence infrastructure, with a total scale of 2,000 trillion won, approximately equivalent to $1.3 trillion. The combined investment of the three memory manufacturers is expected to exceed $1.5 trillion.

Coincidentally, as the three giants hold over 90% of the market share, they have also faced antitrust lawsuits. US law firm Hagens Berman has accused them of colluding to manipulate prices and artificially create shortages.

01 Memory Manufacturers Start "Burning Money"

According to local media reports, Samsung and the SK Group's investment projects in artificial intelligence infrastructure are spread across multiple fields, including semiconductors, displays, batteries, and data centers, with semiconductors being the main focus.

Samsung plans to build 4 to 5 front-end (wafer manufacturing) wafer fabs in Gwangju and South Jeolla Province in southwestern South Korea, with an investment of approximately 300 trillion won. Another 6 wafer fabs will be built in the Yongin Semiconductor Cluster, with an investment of 360 trillion won. The original plan was to be completed in 2048, but it is now planned to be advanced to 2034 - 2035. An investment of over 56 trillion won will be made in the advanced packaging bases in Cheonan and Wenyang, South Chungcheong Province, which are positioned as R & D and production centers for packaging.

Beyond semiconductors, over 350 trillion won will be invested in AI data centers, with a key location in Asan, South Chungcheong Province.

SK Hynix plans to build four or five front-end wafer fabs in Gwangju and expand its NAND flash memory factory in Cheongju, North Chungcheong Province, with a related investment scale of approximately 600 trillion won.

Before this, SK Hynix had already launched some capacity expansion projects, but the company did not clarify whether these projects are included in the newly announced investment plan.

For example, SK Group Chairman Choi Tae-won said on June 11 that the M15X wafer fab in Cheongju is planned to start operations in the second half of 2026, with an initial monthly production capacity of 40,000 wafers, expected to increase to approximately 80,000 wafers in 2027.

In addition, the Yongin Semiconductor Cluster is also accelerating. Phase I is expected to be completed in early 2027 and will consist of 6 clean rooms. Each clean room will gradually increase its monthly production capacity by 60,000 wafers. Just the first factory can increase the monthly production capacity of DRAM by 360,000 wafers before the first half of 2030.

Micron's capacity expansion is being promoted globally, with the US market as the core.

During the earnings conference call for the third fiscal quarter, CFO Mark Murphy provided capital expenditure guidance: approximately $10 billion in the fourth fiscal quarter and $27 billion for the entire fiscal year 2026. He also previewed that the quarterly capital expenditure in fiscal year 2027 will be higher than that in the fourth fiscal quarter of 2026, with more than half of the year-on-year increase coming from construction capital expenditure.

Currently, Micron is building two new wafer fabs, ID1 and ID2, at its headquarters campus in Boise, Idaho, USA, with a total investment of $50 billion. It is one of the largest clean rooms ever built in the United States. ID1 is expected to produce the first batch of DRAM wafers by mid-2027, including the basic chips for HBM. ID2 is expected to start production by the end of 2028.

An investment of $100 billion in a wafer fab cluster in Syracuse, New York, is the largest single private investment in the history of New York State. Construction began in January 2026.

The factory in Manassas, Virginia, recently started producing 1-alpha DDR4 technology, mainly serving the demand for mature process products in the automotive, industrial, medical, aerospace, and defense markets.

In Asia, a new DRAM wafer fab in Hiroshima, Japan, with an investment of $9.6 billion, has introduced EUV equipment. The existing wafer fab acquired in Taiwan, China, is expected to achieve mass product shipments by mid-2027, about one quarter earlier than expected. Construction of a second clean room of the same scale nearby has begun, which will support EUV equipment.

Singapore will become Micron's center of excellence for advanced packaging, focusing on expanding HBM packaging capacity. It is expected to start contributing revenue in the first half of 2027. The ten-year investment in a new NAND wafer fab locally is approximately $24 billion.

Mehrotra specifically mentioned that Micron recently signed a multi-year EUV supply agreement with ASML, paving the way for the next-generation 1-delta node and more advanced processes.

The investment arrangements of the three companies have one thing in common: they are all expanding production, but they are spreading the burden over a long period. None of them is betting on a short-term boom.

Sanjay explained during Micron's third fiscal quarter conference call that the entire industry is working hard to increase supply, but the speed of capacity expansion is limited by the following factors: the long construction cycle of wafer fabs, the increasingly slow equipment delivery, the shortage of skilled workers, and the inadequate energy and water infrastructure.

He collectively referred to these factors as "structural constraints." Under such constraints, the tight supply situation is unlikely to ease in the next two or three years.

02 HBM Becomes a Strategic Bargaining Chip

Compared with traditional DRAM (Dynamic Random Access Memory), it is HBM (High-Bandwidth Memory) that really keeps downstream manufacturers awake at night.

The demand logic for this stacked high-bandwidth memory is different from all previous memory categories.

In June, when Jensen Huang met with SK Group Chairman Choi Tae-won in Seoul, he directly stated that SK Hynix's plan to double its memory wafer production capacity by 2030 is far from enough. "We have been purchasing from SK Hynix, with an annual purchase volume of billions of dollars, and this figure will increase significantly."

According to the calculation logic of analyst fin in "Endgame Deduction of AI Semiconductors," the HBM capacity demand per GPU increases by approximately 40% annually, the production capacity increment on the supply side is 14%, and the growth rate of DRAM unit density is only 9%. The gap between demand and supply is widening.

This means that the supply growth rate of DRAM is only 24%, with a growth rate difference of 16%.

More importantly, traditional DRAM is a standardized product. In a counter-cyclical situation, demand drops sharply, followed by a sharp decline in prices, and the industry faces huge losses. HBM avoids this trap.

Fin believes that although the upper layers of HBM are still DRAM, the Base die has customized attributes. Customers need to sign long-term agreements with memory manufacturers to lock in production capacity. At the same time, the demand for AI is more predictable. A price drop will actually stimulate customers to increase their configuration, providing a bottom support. In addition, due to the rapid technological replacement and the fast depreciation of old products, memory manufacturers tend to compete for the next-generation technology qualification certification of fabless customers rather than engaging in price wars in the existing market.

According to fin's analysis, HBM has changed from the traditional cycle of "making money for one year and losing money for three years" to a growth cycle of "making more money during the upswing and not losing much during the downswing." He believes that as long as the attention mechanism in the Transformer architecture is not disrupted, the exponential demand for HBM will not stop.

03 Who Is Competing for South Korean Production Capacity?

In June, Jensen Huang held a "dinner party" in South Korea with SK Group Chairman Choi Tae-won, SK Hynix CEO Kwon Oh-hyun, and others.

After dinner, Jensen Huang confirmed to the on-site media that NVIDIA's newly launched Vera CPU will use SK Hynix DRAM; the two sides are preparing for "super-large-scale cooperation" in the second half of this year and next year.

On the same day, NVIDIA and SK Hynix officially announced a multi-year technical cooperation agreement covering AI supercomputers, extending to robotics, digital twins, and semiconductor manufacturing.

Jensen Huang is not the only one eyeing South Korea's memory production capacity.

In March this year, AMD CEO Lisa Su completed her first visit to South Korea since taking office in 2014. Her first stop was Samsung Electronics' campus in Pyeongtaek, Gyeonggi Province.

The memorandum of understanding signed between the two sides covers multiple aspects: Samsung will supply HBM4 memory for AMD's next-generation AI accelerator MI455X and develop a customized DRAM solution for the sixth-generation EPYC processor codenamed "Venice."

The fact that industry giants are flocking to Seoul is essentially because memory is changing from a general component that can be purchased on demand to a strategic resource that requires the CEO to personally negotiate for.

04 Micron Upsets the Apple Cart, and Apple Gets Bitten Back

In this round of price hikes, Micron has taken a tougher stance than Samsung and SK Hynix.

After reporting record-breaking third fiscal quarter results, Micron Chief Commercial Officer Sumit Sadana, in an interview with the Wall Street Journal, rarely brought up old grievances.

He revealed that Micron was unable to invest in capacity expansion during the previous industry downturn, partly because some customers were "very aggressive in pricing" and "purchased at the lowest price," squeezing the suppliers' gross profit margins into negative territory. He said that Micron had told those customers at the time that "this approach was unconstructive." Due to extremely poor pricing and profit margins, many industry investments were directly halted in 2023.

Although not named, the customers Micron was "complaining" about are believed to be Apple.

Apple has long been known for its tough bargaining with suppliers. It locks in low prices through long-term purchase contracts, minimizing its own costs while raising the prices of its end products.

However, Apple's tough stance in the supply chain, which indirectly led to the memory shortage, is also biting back at the company itself.

Apple CEO Tim Cook recently admitted that this round of memory shortage is a "once-in-a-century flood," and price hikes are "inevitable." Shortly after his statement, Apple implemented comprehensive price increases for product lines such as MacBook, iPad, Apple TV, and Vision Pro. On that day, Apple's market value evaporated by $265 billion.

According to the Financial Times, the company is quietly lobbying Washington to obtain permission to purchase memory from Yangtze Memory Technologies Co., a Chinese memory chip manufacturer. In Apple's view, Yangtze Memory is not deeply involved in the supply competition for HBM used in AI and can theoretically provide cheaper memory.

Market research firm Gartner analyst Ranjit Atwal commented on this: "Even Apple is not immune. Despite their expertise and long-term planning, this situation is beyond their ability to limit the impact."

Now, everyone is predicting that the memory shortage will end by 2028 at the earliest. However, once the power structure is reversed, it is difficult to go back to the original state. In the memory market, buyers can no longer have the same "unrestricted access" as before - the "commodity era" of memory is coming to an end, and end manufacturers are not fully adapted to this trend yet.

Special translator Jin Lu also contributed to this article

This article is from "Tencent Technology", author: Su Yang, editor: Xu Qingyang. Republished by 36Kr with permission.