What Samsung, SK Hynix, and Micron's capex mean for the future of the memory trade?

The artificial intelligence boom has transformed high-bandwidth memory (HBM) chips into one of the most valuable components in the semiconductor industry, delivering record profits and soaring shareholder returns for the handful of companies capable of producing them.

Samsung Electronics, SK Hynix and Micron have emerged as the biggest beneficiaries of the AI infrastructure race, with hyperscale cloud providers scrambling to secure memory chips needed to power next-generation AI accelerators such as Nvidia's Rubin platform and Google's Ironwood Tensor Processing Unit.

With demand far outpacing supply, manufacturers are naturally resorting to expanding capacity.

However, there is a catch.

The industry in the spotlight has been burned before by such exuberance followed by expansion of manufacturing capacity, followed by oversupply, and then a crash.

Therefore, as once again billions are being committed in capex by these memory makers, investors are asking whether the industry is laying the groundwork for another painful memory chip oversupply cycle, as that would mean that the AI memory trade boom might be more short-lived than is being expected.

AI demand has created an unprecedented supply crunch

Unlike previous technology cycles, the current AI wave has created an extraordinary imbalance between demand and supply.

According to TrendForce, data centres are expected to consume more than 70% of all high-end memory chips produced in 2026, leaving limited supply for traditional markets such as personal computers, smartphones and enterprise hardware.

HBM has become one of the most critical components inside AI servers because it enables processors to move enormous volumes of data at extremely high speeds, a requirement for training and running increasingly sophisticated AI models.

That shortage has allowed memory manufacturers to command premium pricing while posting record earnings.

However, history suggests that such periods rarely last indefinitely.

Over the past two decades, the memory industry has repeatedly followed the same pattern: demand surges, prices rise sharply, manufacturers rush to build new fabrication plants, production eventually floods the market, and prices collapse.

The most recent example came during 2022 and 2023, when Micron and SK Hynix suffered billions of dollars in losses after expanding production based on pandemic-era demand that faded far more quickly than expected.

For investors riding the lucrative AI memory trade, an unexpected pivot to oversupply risks compressing the explosive margins and premium valuations that chipmakers like SK Hynix, Micron, and Samsung have enjoyed.

Chipmakers are committing hundreds of billions of dollars

Memory manufacturers are nevertheless accelerating expansion plans to meet AI demand.

SK Hynix said on Friday its board has approved about 54.3 trillion won ($38.3 billion) of investments through 2031, including 35.2 trillion won for the second phase of its Yongin fabrication complex and another 19.1 trillion won for its M17 NAND facility in Cheongju.

The company had previously announced plans to invest 600 trillion won in the Yongin semiconductor cluster, and another 100 trillion won to expand operations in Cheongju.

Yongin's Y2 fabrication plant will become the second of four DRAM facilities planned for the cluster.

Construction is expected to begin in July next year, with production of HBM and next-generation DRAM chips scheduled to start in mid-2029.

Construction of the M17 NAND flash memory plant will begin in early 2027, with operations expected to start in late 2028.

Samsung is also expanding aggressively.

Together with SK Hynix, the company plans to invest more than $500 billion in a new semiconductor manufacturing hub in southwestern South Korea, part of a government-backed initiative aimed at spreading AI-related industrial investment beyond the Seoul region.

Samsung has also accelerated its Yongin fabrication schedule, aiming to begin operations in 2029 instead of the previously expected 2030-2031 timeframe.

Micron, meanwhile, has committed at least $200 billion toward expanding memory manufacturing and research in the United States while also building a $24 billion semiconductor facility in Singapore.

According to Deloitte, combined capital expenditure by Micron, Samsung and SK Hynix is projected to increase nearly 340% between 2024 and 2027.

Why analysts believe this cycle could be different

Despite the scale of planned investments, many analysts argue this cycle differs fundamentally from previous memory booms.

According to IG, today's market is being supported by structural rather than purely cyclical demand.

"First, Samsung, SK Hynix and Micron — the three companies that have long dominated DRAM — are showing more capital discipline than in past cycles, resisting the urge to race each other into speculative overbuilding. Second, AI-optimised memory is simply harder to produce: Bank of America (BofA) estimates it requires three to four times the production capacity of conventional memory per unit, and the new capacity needed to meet that demand faces physical limits — cleanroom space, power and water supply — that cannot be solved simply by spending more," wrote Fabien Yip, market analyst at IG.

Counterpoint Research director MS Hwang believes meaningful new production capacity is unlikely to arrive before 2028.

"The point at which all of these expansion plans will simultaneously contribute to actual production output is expected to be 2028 at the earliest," Hwang told Benzinga.

"None of this new capacity reaches volume production before 2027 — supporting pricing in the near term, but also meaning several large expansions land in the market at roughly the same time in 2028 and beyond, the point at which the sector's current supply discipline will face its real test," Yip added.

AI spending remains the biggest variable

The industry's biggest risk remains the pace of AI infrastructure investment itself.

Memory stocks experienced sharp volatility throughout July as investors questioned whether hyperscale cloud providers were spending too aggressively on AI infrastructure.

The worry is that with the spend impacting balance sheets now, with lower reserves of cash, any potential pause in data centre building would mean the sky-high demand for memory and other semiconductor companies' products would suddenly drop.

Deloitte estimates that hyperscalers will devote roughly 30% of their data-centre capital expenditure to memory in 2026, rising to 36% in 2027.

"The current memory supply tightness and elevated prices may persist until 2029 or even 2030, assuming continued demand among hyperscalers for memory chips. Other customers that need memory for devices such as PCs, smartphones, and other consumer electronics, as well as for non-AI data centres, will likely also need to contend with high memory prices," Deloitte said.

Bank of America analyst Vivek Arya also believes concerns about a near-term downturn have become excessive.

Arya acknowledged that memory prices and profit margins are likely to normalize eventually as additional supply enters the market between mid-2027 and 2028.

However, he argued that the recent sell-off reflects investor positioning ahead of a potential downturn rather than any meaningful deterioration in demand.

"Hyperscaler spending continues to rise despite higher component costs, suggesting semis/memory pricing power," Arya wrote.

He noted that GPU rental rates remain near record highs and that no major cloud provider has indicated memory shortages are slowing AI deployments.

Following Micron's latest earnings, Bank of America extended its projected AI memory supercycle through the end of 2027, with a possible extension into 2030.

Among the supporting factors cited were Micron's HBM capacity being fully sold through 2027, Kioxia's NAND production already committed into 2027 and 2028, and warnings from SK Hynix's chairman that global memory supply could remain roughly 20% below demand through 2030.

Arya said even under a bearish scenario in which DRAM and NAND prices decline in line with previous industry downturns, earnings could still remain near $100 per share.

That would be substantially higher than Micron's previous cycle peak of around $12 per share recorded in 2018.

"We do not think the industry's underlying cyclicality has disappeared, but the mechanism behind it has changed. Long-term agreements now lock in pricing and allocation years in advance rather than quarter to quarter, giving buyers and sellers more visibility than in past cycles, when a shortage could reverse within a few quarters. That doesn't make memory a one-way trade — it makes the eventual turn slower to arrive, and potentially harder to time," Yip said.