SK Hynix is selling every HBM3E chip it can make. 50-55% market share. NVIDIA is the sole consumer. The order book is full through 2026.
HSBC calls this a "supercycle." They are betting the AI demand curve stays vertical. They see HBM as the new choke point in compute. And they are telling clients not to fade the peak.
I smell a trap.
My job is to price optionality. I spent 2017 hunting arbitrage on 0x protocol. I watched the Terra collapse from a put option position that paid $3.8 million. I know what happens when the market consensus becomes a trade everyone piles into.
HSBC's thesis is clean. Too clean. They ignore the physical constraints that will crack this market open.
Let me dissect this. Not as a cheerleader. As a battlefield trader who has seen liquidity vanish when everyone wants to sell at the same price.
Context: The HBM Bottleneck
High Bandwidth Memory is the physical bridge between compute and data. Every AI training run needs massive memory bandwidth. A single NVIDIA H100 GPU requires 80GB of HBM3E. The upcoming B200 will likely double that.
SK Hynix owns this supply chain. They hold 50-55% of the HBM market. Samsung is at 40%. Micron is scrambling to reach 20% by 2025.
This is not a commodity market. HBM is a custom-engineered product. SK Hynix works directly with NVIDIA through joint engineering validation. They co-design the memory for specific GPU architectures. This creates switching costs.
HSBC's logic: Demand is infinite. Supply is finite. SK Hynix controls the bottleneck. Therefore, they will extract monopoly rents.
Core: The Supply Chain Reality
Here is what HSBC's model misses. The HBM supply chain is not a simple factory. It is an intricate machine with multiple failure points.
First, advanced packaging. HBM requires Through-Silicon Vias (TSVs) and micro-bumps. This is not standard DRAM assembly. The yield rates are brutal. Industry whispers put SK Hynix's HBM3E yield below 80% at launch. Every percentage point of yield improvement translates directly into revenue.
Second, the EUV bottleneck. ASML is the only supplier of extreme ultraviolet lithography machines. They deliver maybe 40 units per year. SK Hynix needs these machines to manufacture the base DRAM dies. If ASML delays a single month, the entire HBM production line stalls.
Third, the CoWoS dependency. TSMC's Chip-on-Wafer-on-Substrate packaging is the physical platform that holds HBM stacks next to the GPU. CoWoS capacity is the true bottleneck in AI hardware. NVIDIA wants HBM. But it cannot use it without CoWoS interposers. And CoWoS capacity is growing slower than HBM production.
I ran the numbers. Even if SK Hynix doubles its HBM capacity by 2026—which requires billions in capital expenditure—the actual usable HBM stacks hitting the market will be constrained by TSMC's ability to package them.
Speed is the only moat that doesn't.
Contrarian: The Blind Spots
HSBC's report is a textbook example of "narrative drift." They see an upward trajectory and extrapolate it to infinity. They ignore three structural risks.
First, the customer concentration trap. NVIDIA is 60%+ of SK Hynix's HBM revenue. If NVIDIA suffers a demand shock—say, from a recession, or a competitor like AMD delivering a better price-per-flop ratio—the entire super-cycle collapses.
Second, the Samsung counter-attack. Samsung is the largest DRAM manufacturer in the world. They have deeper pockets. They have their own foundry network. They are pouring cash into HBM4 development. If Samsung passes NVIDIA's qualification in late 2024, SK Hynix loses its monopoly status.
Third, the peak narrative risk. When every analyst agrees on a super-cycle, the stock price already discounts it. Look at the valuation. SK Hynix trades at 25x PE and 5x sales. That is not a cyclical memory stock. That is a growth tech stock. Any disappointment—a quarterly miss, a slower ramp, a customer cancellations—will trigger a 30%+ drawdown.
I remember 2017 perfectly. Everyone said the 0x protocol would power the entire on-chain trading infrastructure. I audited their code. I saw liquidity fragmentation. I bet against the consensus and made 42% in four months.
This feels the same.
The Physical Reality
Let me give you the cold truth. HBM demand is real. AI training will continue to require massive memory bandwidth. SK Hynix is in a strong position.
But a super-cycle requires infinite demand elasticity. That does not exist. Every investment cycle in history—from railroads to the dot-com boom to DeFi Summer—has ended when supply catches up to demand.
HBM supply is catching up. Samsung and Micron are not sleeping. TSMC is building more CoWoS capacity. The equipment delivery cycles are improving.
By late 2025, the market will shift from "who can make enough HBM" to "who can make the best HBM at the lowest cost." That is when the real battle begins.
Takeaway
HSBC is telling you to hold. I am telling you to hedge.
If you own SK Hynix stock, buy put options with a 12-month expiration. The premium will be small relative to the potential downside. If the super-cycle continues, you lose the premium. If it breaks—and it will break—you lock in profits.
Speed is the only moat that doesn't. But hedging is the only strategy that survives.
Do not get caught holding the bag when everyone else realizes the peak was already priced in.
