Over the past 18 months, SK Hynix has seen its market capitalization triple, driven by an insatiable demand for high-bandwidth memory (HBM) used in AI training chips. Its operating margin surged from near zero to over 40%, a shift that would have been unthinkable during the 2022 downturn when memory prices collapsed by 70%. This dramatic recovery has sparked a heated debate: Has the memory chip industry finally broken the boom-bust cycle that has defined it for decades? The answer, as with most complex systems, is nuanced. The industry is indeed undergoing a structural transformation, but the curse of cyclicality is not dead; it has merely mutated.
To understand this mutation, we must first acknowledge the historical pattern. The memory chip industry—dominated by Samsung, SK Hynix, and Micron—has long been a textbook example of a commodity cycle. Capital-intensive fabs take years to build, and when demand surges, manufacturers race to add capacity. But because supply is lumpy and demand can shift abruptly, the industry repeatedly swings from severe shortage to devastating oversupply. The 2021-2022 shortage gave way to a 2023 glut that wiped out billions in profits. The conventional wisdom was that consolidation over the last decade—from six major DRAM players to three—would dampen volatility by promoting capital discipline. Yet the 2023 downturn was one of the sharpest on record, suggesting consolidation alone is insufficient.
Enter artificial intelligence. The explosion of large language models (LLMs) and AI inference at scale has created a new, seemingly insatiable demand for memory, particularly HBM. HBM3e, the current generation, stacks multiple DRAM dies vertically, connected via advanced packaging technologies like TSV and hybrid bonding. It offers unprecedented bandwidth for GPUs, and it commands a price premium of 300-500% over standard DRAM. For a concentrated oligopoly already adept at managing supply, this high-margin product is a dream—a demand shock that appears structurally driven, not cyclical.
Hype burns out; robustness remains in the ledger. But is the AI demand signal robust, or is it another hype cycle wearing the mask of structural change? To answer that, we need to examine the current state of the industry through a technical lens.
The technology roadmap reveals that the big three are investing heavily in 1β and 1γ DRAM nodes, as well as in 200+ layer 3D NAND. These advanced nodes rely on extreme ultraviolet (EUV) lithography and ultra-high aspect ratio etching, equipment that is both expensive and scarce. The gate count per chip is rising exponentially, and with it, the capital intensity. SK Hynix alone is spending over $15 billion on a new HBM fab in South Korea, while Samsung is investing $17 billion in its Texas facility. This is the same pattern that has triggered past booms: massive capital expenditure in pursuit of expected future demand. The difference today is that the demand is anchored to AI, which is still in its early innings. But as an open-source evangelist who has spent years auditing governance mechanisms and tokenomics, I recognize the danger of placing all faith in a single narrative.
We audit the logic, for humans will always err. Let us audit the AI-capacity flywheel. Currently, HBM is in significant shortage, with lead times extending to 12 months or more. CoWoS packaging, which integrates HBM with GPUs, is the bottleneck. All three memory giants are racing to build HBM packaging lines, using hybrid bonding and fine-pitch micro-bumps. The technology is complex, and yields are still being ramped. However, assume yields improve and capacity comes online over the next 18 months. What happens if AI training demand plateaus, as some researchers predict, due to diminishing returns from scaling LLMs? Or if inference workloads shift to more efficient architectures that require less memory? The consequence would be a sudden oversupply of HBM, which is essentially commodity DRAM packaged differently. The price premium would collapse, and the billions invested in dedicated fabs would become a drag on earnings. This is the classic boom-bust mechanism, now with an AI wrapper.
Code is the only law that does not sleep. The code in this case is the capital allocation decisions made by the oligopoly. Historically, the big three have been adept at signaling discipline while secretly expanding capacity. During the 2022 downturn, they all announced cuts to capital expenditure, yet actual spending only declined by a fraction because of long-term commitments. Today, with government subsidies (the CHIPS Act, South Korean tax credits) incentivizing domestic production, the incentive to expand is even stronger. Geopolitics adds another layer: national security concerns are pushing Samsung and SK Hynix to build factories in the United States, a higher-cost region that demands higher utilization to break even. This is a recipe for oversupply, not stability.
Moreover, the blockchain and crypto industry offers a cautionary parallel. The Ethereum network’s transition to proof-of-stake was celebrated as a permanent demand driver for staking tokens, yet the price of ETH remains heavily cyclical, tied to broader risk appetite. Similarly, the memory industry’s AI demand is tied to the valuation of tech giants like Nvidia, Microsoft, and Google. If those companies’ capex cycles turn, memory will be hit hard. I have seen this pattern in the DeFi summer of 2020: every project assumed governance token demand was structural, but when the hype faded, liquidity evaporated.
Faith in people is costly; faith in math is free. Let us look at the math of the memory industry’s financial structure. The three giants are currently generating strong free cash flow from HBM sales, but they are also burning billions on new fabs. Their combined capital expenditure in 2024 is expected to exceed $100 billion, dwarfing depreciation. This means they are net borrowers, depending on the equity market’s faith in their AI thesis. If that faith wavers, they could face a funding crisis, which would force them to cut prices to maintain utilization, reigniting the cycle. The oligopoly’s discipline works only when all players believe the narrative. History shows that one maverick—often Samsung—has broken ranks to grab market share, triggering price wars.
A contrarian view often emerges from the margins. What if the real enemy of stability is not volatility of demand but the concentration of supply? The oligopoly is now so tight that any disruption—a natural disaster, a labor strike, or a geopolitical flashpoint—could send prices soaring, followed by an overreaction in investment, which then leads to a bust. This is exactly what happened after the 2011 Thailand floods, which caused a DRAM shortage and massive investment, leading to a severe downturn in 2012. The structure is the same, only the catalyst changes.
I seek the signal amidst the noise of the crowd. The crowd is currently shouting that this time is different because AI is a secular trend, not a cyclical one. Yet secular trends can still flow through cyclical industries. The personal computer era was secular, but the memory industry experienced violent cycles throughout the 1990s and 2000s. The smartphone era was secular, yet from 2016 to 2018, DRAM prices surged, then crashed. The current AI mania may be the biggest secular trend yet, but its impact on memory demand is mediated by the same forces: time lags in capacity, capital expenditure, and the herd behavior of executives.
From a blockchain perspective, there is an additional risk: the rise of decentralized AI. Projects like Bittensor and Gensyn aim to use distributed computing resources for AI training and inference, potentially altering the hardware demand profile. If AI shifts to a more edge-based model, the memory requirement could become more fragmented, favoring high-volume, lower-margin products over the high-end HBM that drives current profits. This would be a boon for middle-tier memory manufacturers but a bust for the current premium segment. Yet the big three have staked their future on HBM. Diversification into edge memory would be a slow pivot, given the long life of existing fabs.
Finally, we must consider the human element: the executives themselves. They are incentivized by stock options and bonuses tied to revenue growth and market share. When analysts and investors demand growth, they will invest, regardless of the long-term cycle. The same psychological bias that leads crypto traders to buy at the top also leads memory executives to approve new fabs at the peak of the hype cycle. The curse is as much about human nature as it is about technology.
Open source is a covenant, not just a license. In blockchain, open-source code ensures transparency and prevents hidden agendas. In the memory industry, there is no such transparency. The three giants provide limited detail on specific capacity plans, and their forward guidance is often rosy. Investors are left to infer from equipment orders and supply chain data, an imperfect signal. The industry would benefit from a more open approach, perhaps through a shared forecasting mechanism, but that conflicts with competitive advantage.
In conclusion, the memory chipmakers have not escaped the boom-bust curse. They have merely entered a new phase where the boom is driven by AI, and the inevitable bust will be blamed on something else. The key is to recognize that the structural changes—consolidation, AI demand, and advanced packaging—are real, but they are not antidotes. They are new variables in an old equation. The industry’s fate still hinges on the delicate balance between technological progress and human overreaction. We audit the logic and find that the code remains the same. The only question is when the next major cycle turns, and whether the oligopoly can absorb the shock without a catastrophic implosion. As an open-source evangelist, I believe that transparency and decentralized governance offer better resilience. However, the memory oligopoly is unlikely to adopt such models voluntarily. The cycle, therefore, will continue.