The ledger remembers what the bubble forgets.
A 19-fold profit surge. A 500% stock rally. And yet, the closest thing to a top-tier signal in any cyclical market is the moment when sellers become more rational than buyers.
On July 29, SK hynix will report earnings. Samsung follows on July 30. Both will likely post record profits thanks to HBM3E and whispers of HBM4 allocation. But the technicals are already flashing red: Chaikin Money Flow for SK hynix sits at -0.139. Samsung’s MFI is 42. The capital that chased the narrative is rotating out before the numbers hit the tape.
This is not a bearish article. It is a structural one.
Because if you think semiconductor cycles are separate from crypto cycles, you have not looked at the liquidity map. The same risk-first framework that governs on-chain capital flows governs the massive cap-ex cycles of memory manufacturing. The same herd behavior that drives Bitcoin’s bull runs drives the allocation of billions into HBM capacity. And the same structural fragility that faces a DeFi protocol with a single-dependency oracle faces a memory maker whose largest customer is one chip designer.
The difference is that memory makers have balance sheets. But that does not protect them from the cycle. It just delays the reckoning.
Let’s trace the liquidity.
I. The Hook: Earnings, Technicals, and the Divergence
Most people believe that record earnings justify record stock prices. Data disagrees.
In the last 30 days, three AI–memory stocks—Samsung, SK hynix, and SanDisk (Western Digital)—all fell despite being at or near all-time high profitability. Samsung’s operating profit is up 19x year-over-year. Its stock dropped 7% last week. SK hynix’s ROE hit 61%, higher than any year in its history. Its money flow turned negative.
SanDisk is the most extreme case: a 500% rally from the 2023 low, now showing a -0.07 Chaikin Money Flow and an MFI of 36. The percentage of stocks trading above the 50-day moving average collapsed to 30%.
This is not a correction. It is a signal. A signal that the market is pricing the peak of the cycle before management confirms it.
For a macro watcher, this is familiar territory. It is the same pattern we saw in DeFi summer 2021: TVL records, but on-chain velocity declining. The best time to sell is when the numbers are perfect and the crowd is still buying.
The crowd is not buying these names. Retail participation is missing. The big money has rotated out.
II. The Context: HBM as the New Oil, But with a Short Shelf Life
High Bandwidth Memory (HBM) is the physical bottleneck for AI training chips. A single H100 GPU requires roughly 80 GB of HBM, and the next generation B200 will likely require 192 GB. The industry cannot make it fast enough. SK hynix controls roughly 50% of the HBM3E market, Samsung 40%, and Micron the rest.
But this is not a technology question. It is a capacity question. HBM production uses the same fabs that make DRAM. The decision to allocate wafer starts to HBM versus conventional DRAM is a financial one. Every wafer for HBM is a bet that AI demand will stay high. If that bet goes wrong, the oversupply of conventional DRAM will crush margins across the board.
SanDisk is NAND. NAND is not HBM. Its price rally has been driven by AI data center hoarding of SSD storage, but NAND cycles are shorter and sharper. A 500% rally in a commodity memory product is not sustainable. It is a liquidity event.
III. The Core: Structural Liquidity Fragmentation in AI Memory
Let me reframe this from my lens: a Data Science graduate who spent years auditing token distribution models and DeFi liquidity pools. The same principles apply.
In 2017, I wrote a Python script to track Golem’s token emission schedule against real-time DEX liquidity. I found a 15% discrepancy in the claimed distribution. The lesson: liquidity is not depth, it is just delayed panic.
Today, the HBM market faces an analogous fragmentation. Three players control the supply. One customer—Nvidia—absorbs more than 60% of HBM output. That is single-point-of-failure dependency. It is the same risk as a DeFi protocol with a single oracle: if the oracle fails, the entire market panics.
What is the oracle here? Nvidia’s capital expenditure guidance. If Nvidia’s hyperscaler customers—Microsoft, Meta, Google, Amazon—reduce their AI CapEx by even 10%, Nvidia’s orders drop. HBM demand drops. The memory makers are left with excess capacity and falling prices.
This is not a theory. It is a historical pattern. In 2018, after the crypto mining boom, NAND prices collapsed because the mining demand vanished. The same could happen with HBM if AI CapEx cycles moderate.
But there is a deeper structural issue. The industry is deploying massive capital to expand HBM capacity. Samsung and SK hynix are spending $20–30 billion each per year on CapEx. Much of that is for HBM-specific packaging: TSV (through-silicon vias), micro-bumps, hybrid bonding. This is not easily convertible to other uses.
Imagine a Layer-2 solution that raised a billion dollars to build a sequencer, and then the L1 changed its fee model overnight. That is the risk memory makers are taking. The infrastructure is purpose-built for AI memory. If AI demand plateaus, the CapEx becomes stranded.
IV. The Contrarian Angle: Decoupling Is a Myth
Some analysts argue that HBM and NAND are now structurally decoupled from the traditional memory cycle because AI demand is secular, not cyclical. They point to the CAGR of 50%+ and the fact that HBM now accounts for 40% of a GPU’s BOM cost.
I disagree. Architecture outlasts anxiety, but architecture is built on balance sheets.
The secular thesis assumes that AI training demand will grow linearly with compute. But compute itself is subject to efficiency gains. If NVIDIA’s next architecture reduces HBM requirements by 30% per epoch (a plausible engineering outcome), the demand for HBM per GPU drops. The total market may still grow, but the margin on each chip shrinks.
More importantly, the decoupling thesis ignores the supply side. Samsung, SK hynix, and Micron are all racing to deliver HBM4 by 2026–2027. When multiple players rush capacity, the market trends toward oversupply. That is a law of physics in cyclical industries, no matter how innovative the product.
SanDisk’s NAND decoupling thesis is even weaker. NAND is a commodity. AI data centers need SSD storage, but that demand could be met by a 10% increase in capacity from existing fabs. The 500% rally was speculation on AI storage demand, not structural necessity.
V. The Takeaway: Positioning for the Cycle, Not the Narrative
If you hold a position in any of these names, the question is not whether the technology is good. It is whether the current price already accounts for the peak.
The data suggests it does. Money flow is negative. Retail participation is absent. The stocks are falling on good news. These are classic signs of institutional distribution.
But this does not mean the cycle is over. It means the easy money has been made. For a macro watcher, the next move is not to buy the dip without a catalyst. The next move is to identify the catalyst that resets expectations. That catalyst could be:
- HBM4 official orders from NVIDIA (a positive surprise, but already priced in).
- A CapEx cut from a hyperscaler (a negative surprise, not priced in).
- A technology breakthrough that reduces HBM need (unlikely, but would be a massive negative).
The most likely scenario is a sideways grind until the earnings reports clarify the HBM4 timeline. After earnings, the market will have a new set of data points. If guidance is strong, expect a relief rally of 10–15% in SK hynix and 5–10% in Samsung. If guidance is weak, the sell-off will accelerate.
SanDisk is a binary bet. If AI storage demand materializes in the form of large enterprise orders, the rally could resume. If not, a 50% correction from current levels is plausible.
VI. Structural Parallels to Crypto
I see three direct lessons from this cycle for crypto participants:
- Liquidity fragmentation kills margins. Just as HBM capacity is split among three players, crypto liquidity is split across dozens of L2s. The same small user base gets sliced thinner. The result: lower transaction fee revenue and worse user experience. The HBM market shows what happens when capacity grows faster than demand. Crypto L2s are on the same trajectory.
- Single-customer dependency is a structural risk. SK hynix with 70% revenue from Nvidia is the same as a DeFi protocol with 80% TVL from one whale. The whale leaves, the protocol dies. In crypto, we call it a “bank run.” In memory, we call it a “client diversification problem.” The mechanics are identical.
- Macro moves first; the chain reacts later. The sell signal on these memory stocks appeared weeks before the earnings reports. On-chain data for Bitcoin and Ethereum shows similar pre-emptive behavior during market tops. The ledger remembers what the bubble forgets. The memory cycle is telling us that AI demand, like Bitcoin demand, is not infinite. It is just currently underestimated.
VII. Risk-First Framework for the Next 12 Months
Based on my own experience modeling DeFi liquidity stress tests in 2020 and stablecoin peg risks in 2022, I apply the same framework here.
| Risk Factor | Probability (12-month) | Impact if Realized | |-------------|------------------------|--------------------| | HBM oversupply (2027) | 30% | 30-50% downside in SK hynix, 20-30% in Samsung | | Nvidia CapEx cut | 20% | Immediate HBM order reduction, 20-40% downside in SK hynix and Samsung | | Samsung HBM4 certification | 25% | Positive for Samsung, negative for SK hynix (20-40% downside for SK hynix single-stock) | | NAND price decline | 40% | 50%+ downside for SanDisk from current levels |
The highest probability risk is NAND price normalization, which is already priced into SanDisk’s technicals but not its valuation. The highest impact risk is an Nvidia CapEx cut, which would crater the entire AI memory complex.
VIII. Final Thought: The Cycle Is Not Dead, Just Delayed
I have been observing semiconductor cycles for 17 years. Each time, the narrative changes—AI, cloud, mobile, crypto. But the structural pattern remains the same:
- A breakthrough product creates a demand shock.
- Suppliers race to add capacity.
- Demand growth slows due to efficiency gains or macro headwinds.
- Supply catches up, margins compress, stocks correct.
- The cycle resets.
We are between stages 2 and 3. The demand shock is real. The capacity additions are massive. The efficiency gains are coming. The margin compression is not yet visible, but the stock market is pricing it in.
The crypto analogy is obvious: we are in the post-halving accumulation phase for Bitcoin, with ETFs providing demand. But the infrastructure (L2s, sidechains) is overbuilt for current user growth. The divergence between narrative and network activity is the same divergence between HBM earnings and stock prices.
Liquidity is not depth. It is just delayed panic.
When the panic comes—whether from a CapEx cut, an oversupply, or a macro shock—the market will remember that the finest technology cannot support a valuation that assumes perpetual exponential growth.
The ledger remembers. The balance sheet does too.
