1/ The HSBC HBM report is making the rounds. Their argument: AI demand creates an 'HBM super cycle,' SK Hynix is the bottleneck, and we're not at the peak yet. It’s the narrative that fits the current market euphoria perfectly. That’s precisely why I’m skeptical.
2/ The report is well-structured. It identifies the key driver: the demand for High Bandwidth Memory from NVIDIA and AMD is insatiable. It correctly points out that HBM is now the physical bottleneck, not just GPU compute. SK Hynix’s 50-55% market share and its alliance with TSMC are real moats.
3/ But any macro observer knows the seeds of the next downturn are planted during the current expansion. The report’s fundamental flaw is that it frames this as a purely technological, demand-driven cycle. It treats the scarcity as a permanent feature, not a temporary, capital-intensive phase. This is a classic liquidity trap for investors.
4/ Let’s start with the 'super cycle' label. In semiconductor history, a super cycle implies a structural shift in demand that outpaces supply for a prolonged period. Memory is cyclical by nature—boom and bust. The current HBM boom is driven by a single, concentrated demand source: AI training. This is a high-beta game, not a new steady state.
5/ The report overlooks the massive, looming supply response. SK Hynix, Samsung, and Micron are pouring tens of billions into HBM capacity. SK Hynix alone has announced a multi-billion dollar plan to double HBM capacity by 2027. The capital expenditure (Capex) cycle is aggressive. When these fabs come online in 2026, we are not facing a shortage—we are facing a glut.
6/ The key metric to watch isn't market share or technology; it's the free cash flow yield. Today, SK Hynix has negative free cash flow due to those massive Capex investments. The 'super cycle' narrative is built on the assumption that future cash flows will justify these investments. This is a high-stakes bet on AI demand staying exponential.
7/ What happens when the first bit of bad news hits? If a major cloud provider (a hyperscaler) cuts its 2026 Capex guidance, or if an AI killer app fails to materialize, the carefully constructed 'scarcity' narrative collapses. The high-margin, high-demand scenario reverses, and we are left with enormous depreciation costs chewing through margins. This is the classic 'value destruction through expansion.'
8/ The report also minimizes the competitive threat. It assumes SK Hynix will maintain its lead. Memory is a game of leapfrog. Samsung is massive and aggressive. Micron has a clear roadmap. In memory, technology leads are measured in quarters, not years. Once Samsung or Micron ramps equivalent HBM products, the pricing power evaporates. The market becomes a commodity market again.
9/ From a macro-liquidity perspective, this report is a classic 'late-cycle' sell-side call. It arrives when the stock has already rallied significantly, and it’s trying to justify the current valuation. The thesis is 'things are good, and they’ll get even better.' That is not a risk management framework; it’s a narrative for momentum.
10/ Let’s examine the 'contrarian missing piece' the report glosses over: the potential for a demand cliff. The current AI boom is heavily capex-funded. If the return on that capital (ROIC) doesn’t materialize for the hyperscalers, they will pull back. There is no current evidence that the ROI is there for the massive model training. This is speculative spending, not consumer demand.
11/ I’ve seen this movie before. In 2017, I audited ICO smart contracts that promised 'decentralized computation'—they all failed because the user demand didn't match the infrastructure spend. The exact same dynamic is playing out in AI: a massive build-out of compute capacity (HBM, GPUs) in anticipation of use cases that may not arrive for years, if at all.
12/ The report’s 'agentic AI' argument is the deus ex machina. It assumes a new, unproven market will emerge to absorb even more memory. This is a hope, not a thesis. Relying on the next big technological breakthrough to justify current valuations is the hallmark of a speculative bubble, not a structural growth story.
13/ From an institutional yield skepticism standpoint, the HBM 'super cycle' is being used to sell a yield story. 'Invest in HBM, get AI growth.'. This is a trap. The true yield for SK Hynix is negative today because of Capex. Investors are paying for future potential, not current cash flow. In a higher-for-longer rate environment, future cash flows are worth less today.
14/ The most dangerous phrase in the report? It’s not 'super cycle.' It's the implicit assumption that the current supply-demand imbalance is the new normal. A structural view of capital flows tells us that high returns attract new capital. That capital is now flooding into HBM. Supply will respond. The bottleneck will be broken.
15/ The takeaway for sophisticated investors is clear: the HBM narrative is a momentum play, not a deep-value, structural growth opportunity. The risks are enormous: oversupply (2026+), competitive erosion, demand cliff, and a punishing Capex cycle. The market is pricing the best-case scenario. That’s a recipe for mean reversion.
16/ My professional experience, from DeFi Summer to the Terra collapse, has taught me one iron law: Liquidity isn't PnL, it's existence. When the demand narrative shifts, the liquidity flooding into HBM will dry up overnight. The companies left holding the multi-billion-dollar fabs will be the ones trapped. The 'super cycle' is just a cycle, and every cycle ends. I will be watching for the first Capex cut.
17/ My own research, based on cross-border payment flows and capital market liquidity, aligns with a more cautious view. The real opportunity isn't in the commodity silicon itself, but in the middleware and the design tools that enable the efficient consumption of memory. A thesis that relies solely on the physical scarcity of an emerging commodity is a thesis based on sand.