South Korean stocks are now tradable! Invest globally with one app
Key Takeaways from SK Hynix Chairman's Latest Interview: The Memory Cycle Is Set for a Complete Overhaul by 2027 ❣️
In a recent exclusive interview with CNBC, Chey Tae-won, Chairman of the SK Group and SK Hynix, provided a core forecast for the memory industry over the next two years: 2027 will see the largest supply deficit in this cycle. Confident in this trend, SK Hynix has unveiled a massive $720 billion capacity expansion plan.
Over the weekend, I conducted a deep dive into the full interview. Combining this with the current low-level consolidation in the US memory sector, I will break it down in plain language: the AI-driven memory rally operates on completely different logic than the cycles of the past decade.
1. Understanding the Core: AI Has Fundamentally Rewritten Memory Demand Logic
In the past, analyzing the memory cycle was straightforward: it simply followed trends in smartphones and PCs.
With only billions of users globally—each typically owning one phone and one computer—the demand ceiling was fixed, offering extremely limited growth. Consequently, the industry was trapped in a vicious cycle of 'price hikes leading to capacity expansion, resulting in oversupply, followed by price drops and production cuts.'
But now, with the advent of AI, the industry's foundational logic has been completely upended.
Chairman Chey used a vivid analogy: current AI is like a four-year-old child that will experience exponential growth in the future.
The most critical data point: the number of global AI agents is projected to surge by 77 times over the next five years.
Unlike traditional devices, each AI agent requires dedicated memory and storage resources. It is no longer about humans using devices; instead, countless AI programs run data, store information, and perform inference 24/7. This multiplier effect has completely shattered the previous ceiling on memory demand.
Demand is exploding, but supply is tightly constrained, creating a rigid structural gap:
1. Wafer fab capacity cycles are extremely long: It takes at least 4-5 years from planning and construction to mass production for a new factory, making it impossible to rapidly expand capacity to address urgent shortages;
2. Severe supply-demand mismatch currently exists: Major cloud providers' memory demand is set to nearly double next year, but capacity cannot keep pace. The 2027 capacity of the three giants—Samsung, SK Hynix, and Micron—is already largely allocated. Customers generally receive only 60-70% of their requested volume.
This has triggered the current chip inflation:
High-end HBM memory is in extreme shortage and is particularly capacity-intensive; under the same wafer area, HBM consumes more than four times the capacity of standard DRAM.
Put simply: AI high-end storage is aggressively crowding out traditional storage capacity for smartphones, PCs, and automobiles.
Currently, the share of DRAM and NAND in the bill of materials (BOM) for smartphones and computers has jumped from the previous 25% to nearly 50%. The cost pressure on terminal hardware will continue to be passed down the supply chain.
II. New Strategies by Industry Giants: Moving Away from Blind Capacity Expansion and Using Long-Term Contracts to Eliminate Extreme Cycles
Many people see SK Hynix's KRW 720 billion capacity expansion and their first reaction is: Are they going down the old path of 'overcapacity' again?
In fact, that is completely wrong. The giants have shifted their business model this time and are no longer blindly betting on market trends.
The current core industry solution: universal adoption of long-term supply agreements.
Previously, manufacturers feared price declines on their own. Now, there is rigid demand from both upstream and downstream players: cloud providers fear chip shortages and supply disruptions, while original manufacturers fear sharp cyclical downturns. Both sides are locking in multi-year contracts, using mechanisms such as annual price reviews, long-term renewals, and even joint ventures for factory construction to convert volatile cyclical pricing into stable, predictable cash flows.
Furthermore, do not assume that capacity expansion can be executed at will; the industry has significant hidden barriers.
Choi Tae-won also spoke frankly this time: wafer fabs do not lack land, electricity, or subsidies; what they lack most is a complete industrial chain ecosystem.
A mature wafer fab requires support from 600–700 suppliers of materials, chemicals, and equipment.
Rebuilding this ecosystem overseas costs twice as much as doing so domestically. This means that while there are many paper plans for capacity expansion, the release of actual effective capacity is far slower and smaller than the market expects. The tight supply situation cannot be resolved in the short term.
3. SK Hynix's Hidden Advantage: Competing Not Just on Technology, but on Team Strength
Currently, among the three memory giants—Samsung, Micron, and SK Hynix—there is virtually no generational gap in technology or financial scale.
SK Hynix’s ability to break through against the trend is not primarily due to hardware advantages, but rather the combat effectiveness of the team forged during its restructuring.
Teams that have undergone debt restructuring possess stronger risk awareness, higher execution efficiency, and better collaboration, with an intense desire for performance growth. This constitutes a hidden core competitiveness that surpasses technology and capital.
4. The current primary contradiction: peak industry prosperity versus extreme market pessimism
This represents the most fragmented segment with the greatest expectation gap in the current US memory stock sector:
The industry clearly forecasts a major shortage by 2027, yet capital market valuations have plummeted to historic lows.
📶 Here is a set of striking data for your reference:
SK Hynix has already recorded its highest-ever profits, yet its 12-month forward P/E ratio has dropped to as low as 3.3x.
For comparison: 2019 marked the trough of the previous memory cycle, during which SK Hynix posted losses for five consecutive quarters, yet its valuation remained slightly above 3x.
Currently, despite generating record profits, its valuation is comparable to that seen during its worst loss-making period.
The severe divergence between high P/B and extremely low P/E ratios essentially indicates that the market does not believe high profitability is sustainable and has already priced in a future cyclical downturn.
⚠️ The two major bearish arguments currently prevailing in the market, which are also the main sources of investor hesitation, are:
1. Downstream clients have begun proactively cutting costs, leading to a marginal slowdown in memory demand
Memory chip prices have surged too sharply, forcing downstream players to start hedging against risks as they can no longer absorb the costs.
For instance, NVIDIA’s new-generation server platform has directly reduced memory configurations per rack;
Meanwhile, cloud providers have locked in long-term agreements, capping any further price increases. The rise in spot prices for server DRAM has begun to moderate, with market concerns growing that prices are nearing their peak.
2. AI capital expenditure is approaching its limit, posing leverage risks
Major cloud providers’ AI investments are nearly catching up with their own operating cash flows.
We are even seeing a model where institutional investors step in to provide financing support, relying on leverage to sustain computing power procurement.
Market concerns: Once AI technology iterates, chips become obsolete, or customers default, financial leverage risks could backfire and impact the entire supply chain.
V. Key Divergence: DRAM and NAND trends will differ completely in 2027
A key reminder: 2027 will see a structural bull market, not a broad-based rally across all memory chips; the trends for these two segments will diverge significantly.
1. DRAM (HBM + Server Memory): Persistent shortage with the largest supply gap
The core shortage in 2027 will center on DRAM.
Industry roadmap: By the end of 2027, the three major manufacturers will allocate 30% of their wafer capacity to HBM (up from just 20% currently).
But here is the key point: Allocating 30% of wafer capacity will only yield a 13% increase in memory bit supply.
HBM’s extreme consumption of production capacity will continue to squeeze the output available for standard server DRAM.
The conclusion is clear: Supply and demand for server DRAM and HBM will be tightest in 2027, with strong market sentiment persisting throughout the year.
2. NAND Flash: Tight initially, easing later; likely divergence in the second half of the year
The investment thesis for NAND is significantly weaker.
Although AI inference and massive data storage are driving up demand for enterprise SSDs and high-capacity hard drives, NAND technology allows for relatively easy capacity expansion.
Total storage capacity can be rapidly increased by stacking more layers and upgrading process nodes.
Therefore, if demand from consumer electronics such as smartphones and PCs fails to keep pace, the NAND market could easily shift from shortage to surplus in the second half of 2027, posing a risk of price declines.
6. Core Conclusion: The cycle has not ended, but its underlying logic has undergone a fundamental upgrade.
Traditionally, the memory chip sector was viewed as a classic strong-cycle industry: price hikes → capacity expansion → oversupply → sharp price declines.
However, AI is now reshaping the entire industry:
1. Demand is driven by long-term structural growth from AI, rather than short-term fluctuations in consumer electronics;
2. Effective capacity is tightly constrained by ecosystem dynamics and cyclical factors, preventing rapid oversupply;
3. Long-term contract mechanisms have smoothed out extreme price volatility, significantly enhancing earnings stability.
Current market pessimism has already priced in all negative factors—such as peak prices, cycle reversal, and peaked capital expenditure—into single-digit valuations.
Yet the market has completely overlooked two major positive catalysts:
1. Unprecedented abundance of industry cash flow;
2. Industry giants are initiating substantial shareholder returns: SK Hynix has committed to allocating half of its free cash flow to dividends and buybacks over the next three years, while SanDisk is returning 100% of its excess cash flow to shareholders.
⭕️ Final Summary:
1. 2027 will be a year of severe shortage for DRAM/HBM, marking the peak of prosperity in this memory cycle;
2. The industry is no longer characterized by the boom-and-bust patterns of previous cycles; AI has fundamentally raised both the earnings center and the duration of the cycle;
3. The current sector performance is driven by an expectation gap featuring "high industrial prosperity + extremely undervalued valuations." Pessimistic sentiment has been fully released, providing a substantial margin of safety;
4. The market trend is structurally differentiated: maintain heavy positions in HBM and server DRAM, while viewing NAND in the second half of 2027 with caution.
The pendulum of the cycle never stops, but for this round of AI-driven memory rallies, you absolutely cannot rely on the old playbook from the past decade.
Risk Disclaimer: The above content only represents the author's view. It does not represent any position or investment advice of Futu. Futu makes no representation or warranty.Read more
Comments
to post a comment
8
4