The fork wasn't in the road; it was in the forecast. When SK Group Chairman Chey Tae-won told Korean media that the industry should 'expand capacity rather than restrict supply,' he wasn’t making a market call. He was performing an act of strategic theater. He is asking the entire memory ecosystem to bet on a future that is elastic—where AI demand pulls enough volume to justify today’s massive capital expenditure. It’s a beautiful narrative. But a forensic look at the physical constraints tells a different, more brittle story.
The Hype Cycle's New Bottleneck We are deep in the AI infrastructure build-out. The headlines are all about NVIDIA’s next chip or OpenAI’s next round. But the real war is being fought in the packaging line. SK Hynix, the current leader in High Bandwidth Memory (HBM), is the gatekeeper for the most critical component in an AI server: the memory stack. They are not just a supplier; they are the primary bottleneck between a GB200 order and a fully functional datacenter rack. This position gives them immense power, but it also makes them a target. Chairman Chey’s statement is a defensive signal as much as an offensive one.
The Core: Dissecting the Capacity Mirage The core of Chey’s argument is straightforward: demand for AI-specific memory (HBM) will grow 60-100% in 2025, while overall memory demand rises 50-60%. To meet this, he is pleading for an industry-wide capacity push. But let’s test this thesis against the reality of the wafer fab. The industry doesn’t just flip a switch. You need ASML’s High-NA EUV lithography machines, which have a 12-18 month delivery lead time. You need the advanced packaging equipment for Through-Silicon Vias (TSV) and hybrid bonding, a market where the supplier base is even more constrained.
Here is the first disconnect. Chey talks about "equipment, personnel, and construction timelines" as limiters. This is true. But the primary limiter is not just the number of fabs; it is the technical architecture of the HBM stack. An HBM3E die is not just a DRAM cell shrunk to 1bnm. It is a complex 3D structure requiring perfect alignment across eight or twelve layers. The yield on the TSV process, the bonding alignment, and the thermal management are where production actually stalls. These are physics problems, not just volume problems. Based on my analysis of multiple HBM projects, the capacity constraint isn't in the DRAM fabs themselves; it is in the back-end packaging capacity. This is the hidden assumption in Chey's call: he is implicitly asking the entire OSAT (Outsourced Semiconductor Assembly and Test) ecosystem—companies like ASE and Amkor—to also double their HBM-specific capacity. That is a far bigger ask.
The Contrarian View: Where the Bulls Got It Right I am a cold dissector, but I must acknowledge where Chey’s logic holds water. The contrarian angle here is not that he is wrong about demand, but that he is too optimistic about the elasticity of supply. He is correct that the current pricing environment is historically high. An HBM3E stack is selling for a premium roughly 5x that of an equivalent DDR5 module. He is also correct that maintaining this high price by limiting supply would be a short-term strategy that invites competition from Samsung and Micron, and potentially even a vertical integration move by NVIDIA. His "expand the pie" strategy is a classic defensive move: flood the market with capacity to build unassailable supply chain relationships.
The bulls are right that someone needs to build this capacity. The question is who pays for it, and when does the return arrive. Chey is betting that the AI investment cycle is a multi-year super-cycle, not a two-year bubble. He is effectively saying: "Let’s build the factory now, and the demand will come." This is the same logic that built the DRAM industry into a boom-bust cycle for decades. The difference this time is the explicit government backing (CHIPS Act) and the near-monopsony power of the buyer (NVIDIA).
The Takeaway: A Bet on a Future You Cannot See Chey Tae-won is not just selling memory. He is selling a narrative of inelastic demand. But the real risk is not a demand collapse; it is a supply-side bottleneck so severe that it prevents the narrative from materializing. We are not in a phase where capacity is a simple choice. We are in a phase where the delivery of a single High-NA EUV scanner can bottleneck an entire nation's AI ambitions. Cold hands dissect the heat of a hype cycle, and the heat here is not from demand. It is from the friction of a supply chain that is too hot to handle. The question is not 'will the demand come?' It is 'can the physical world keep up with the digital promise?'