Will AI Demand Elasticity Rewrite the Memory Cycle?

The market is pricing a memory bloodbath in 2028. Every sell-side model assumes a replay of 2019: capacity gluts, price crashes, and margin implosions. But one analyst at Citrini Research, Jukan, is arguing that the traditional memory cycle is dead, killed not by supply discipline but by a structural shift in demand elasticity from AI. His central claim is that AI workloads exhibit a price elasticity of demand of roughly 1.42. When HBM prices fall by 30%, AI-related HBM demand rises by 42%. That demand response, he argues, will prevent the classic overshoot-and-collapse pattern. Instead of a 50%+ profit decline in 2028, he models a relatively mild 15% drop. This is a provocative thesis. It suggests that memory stocks—currently trading on 10x PE with a deep cyclical discount—deserve a multiple expansion toward 15x or more, similar to growth semiconductor companies like ASML. But before we start repricing the sector, we need to stress-test the transmission mechanism. Does elastic demand at the application layer actually translate into higher volume for memory suppliers? The Elasticity Chain Jukan's 1.42 elasticity comes from observed API call volumes at large AI model providers. When inference costs fall by 30%, developers queue up more queries. That increases compute demand, which forces hyperscalers to buy more GPUs, which comes packaged with HBM. The logic is sound—until you insert NVIDIA into the middle. NVIDIA is not a price-taker. It has 80%+ market share in AI training and a growing inference franchise. When HBM costs fall, NVIDIA could absorb that margin rather than passing it through to cloud customers. If NVIDIA maintains its own profit margins, the lower HBM price never reaches the AI developer. The volume response at the memory level is muted. This is the hidden friction in the chain. The 1.42 elasticity applies to the last mile of AI consumption, not to the upstream memory market. The actual elasticity that HBM suppliers face is likely much lower, perhaps 0.5 to 0.8, depending on NVIDIA's pricing strategy. Historical precedent supports this. In 2016-2017, NAND prices fell by 40% but SSD demand only grew 25%—an elasticity of 0.6. The consumer and enterprise SSD market had its own pass-through layers (PC OEMs, cloud data center builders). HBM today is even more concentrated: a single customer, NVIDIA, buys 70-80% of all HBM3E output. That is a monopsony-like structure. Supply Discipline vs. Supply Diplomacy The other pillar of Jukan's thesis is that 2028 supply will be moderate because memory makers have learned from past mistakes. He points to capex discipline after 2022's crash. But look at the announced capacity plans. Samsung is building new HBM lines in Cheonan and Pyeongtaek, SK Hynix is expanding in Cheongju and M15X, and Micron is accelerating its Boise HBM fab. The cumulative capex over 2025-2028 will exceed $100 billion for the three players combined. DRAM fabs take 18-24 months to build and ramp. A decision made in early 2025 to meet NVIDIA demand in 2026 becomes supply in 2028. The industry is already in a race for market share in HBM4 and HBM4E. Each company wants to be the primary partner for NVIDIA's Rubin architectre in 2027 and beyond. That competitive dynamic overrides collective discipline. Furthermore, internal competition between Samsung and SK Hynix is intensifying. They are not a cartel. History shows that when two Korean giants fight for technological supremacy, they eventually cut prices to secure volume commitments. In 2023, Samsung deliberately undercut SK Hynix on HBM2E pricing to win a spot in NVIDIA's Hopper ramp. A replay is likely once HBM4 hits volume in 2027-2028. The risk is not just oversupply, but a price war disguised as "collaboration" with customers. NVIDIA will play them against each other. Geopolitical Brakes Jukan's model assumes smooth capacity ramps. But each HBM fab requires EUV lithography from ASML—single-supplier, 12-18 month lead times. Any disruption in export controls, equipment delivery, or even power availability in Korea could delay output by 6-12 months. That works in the other direction too. If supply is delayed, the 2028 glut never materializes. Prices remain elevated. Jukan's bear case becomes too pessimistic. But I'd argue geopoilitical risk is actually a brake on supply rather than an accelerator. The US CHIPS Act is incentivizing Micron to build capacity in America, but they face construction delays and labor shortages. Chinese DRAM makers like CXMT are years behind in HBM. The real supply overhang, if it comes, will be from Samsung and SK Hynix themselves, and it's already being hedged by massive buyback programs and willingness to cut legacy DRAM output to support HBM margins. The bottom line: the 2028 scenario is not a binary recession vs. expansion. It's a question of whether AI demand grows fast enough to fill the combined HBM capacity of three aggressive oligopolists. Valuation Implications If Jukan is right and the memory cycle is flattening, then memory stocks should trade at higher multiples. Current PE of 10-12x for Samsung and SK Hynix implies a deeply cyclical margin structure. A re-rating to 15-20x would add 50-100% upside even if earnings stay flat. But if the cycle is not dead—if 2028 brings a traditional profit collapse—then current prices are vulnerable. The market is already pricing in some cyclical discount. The risk/reward depends on which scenario is more probable. I believe the truth is in the middle. AI will provide a structural demand floor, but internal competition and technological parity will compress margins over time. The memory industry will not become a utility, but its peak profitability may be lower than in 2024. Stocks that trade on peak-cycle earnings will need to adjust. The best way to play this is not to bet on the macro cycle alone, but to pick the technology leader. In HBM, SK Hynix has the best execution and margins. Samsung is catching up but may need to cut prices to win share. Micron is a wildcard with US government backing. For now, the data does not support a panic sell. The elasticity chain is real, just attenuated. Monitor NVIDIA's capex guidance and memory makers' 2026 capex announcements—they are the leading indicators of the 2028 outcome. The cycle is not dead. It's just different. And different means new opportunities for those who understand the supply-demand mechanics, not just the narrative.