KB Securities has published a number that should stop a room. Between 2026 and 2027, the combined operating profit of Samsung Electronics and SK Hynix is forecast at 157.1 trillion won. Across the preceding twenty-six years โ 2000 through 2025 โ those same two companies cumulatively generated 74.8 trillion won. A two-year window priced at 2.1 times the entire operating history of the most consequential memory oligopoly the world has ever assembled. I have read a great many sell-side notes in my career; most are extrapolation wearing a suit. This one is different, not because the arithmetic is heroic, but because the note attaches a structural mechanism to it: roughly seventy to eighty percent of the memory both firms intend to produce in 2027 is already committed under five-year supply agreements with major global technology buyers. That is not a forecast of demand. That is a description of a market that has quietly stopped behaving like a market.
The underlying figures are precise and worth restating without adornment. KB Securities projects Samsung Electronics at 55.5 trillion won of operating profit in 2027, a fifty-one percent year-on-year increase, and SK Hynix at 38.9 trillion won, up fifty percent. The stated driver is AI accelerator demand compressing memory supply to the point where allocation replaces negotiation. Both names now trade at price-to-earnings ratios near three times. Foreign ownership across the pair has fallen to its lowest level since the 2008 global financial crisis. KB's conclusion is that the market is underestimating long-term profit growth potential.
Here the discipline has to begin, because two of those facts are being read as one. Three times earnings and a seventeen-year low in foreign participation are not evidence of a mispricing in semiconductor cash flows. They are evidence that the marginal buyer of Korean large-cap equity has changed nationality. Those are different claims, requiring different evidence, and conflating them is the most common failure mode in cross-asset analysis. A low multiple is a description of who is willing to hold the asset, not a verdict on what the asset earns.
The genuinely new information in the note is not the 157.1 trillion won. It is the seventy-to-eighty percent contracted share. Merchant memory has historically been the purest expression of spot-market cyclicality available in public equity โ a commodity where a twelve-week inventory swing could destroy a year of margin. If a majority of 2027 output is locked into five-year agreements, the volatility profile of the vendor changes at the level of capital structure, not at the level of quarterly guidance. Fixed-price volume converts a trading instrument into something closer to an infrastructure annuity. That is a defensible structural argument, and it is the one the market appears to be ignoring.
But contracts are not physics. In 2018, with DRAM pricing at its peak, hyperscalers signed supply agreements of comparable ambition. When capex digestion arrived in the first half of 2019, those agreements were renegotiated, deferred, or absorbed into inventory write-downs. Take-or-pay clauses exist on paper and dissolve in practice when the counterparty is a trillion-dollar buyer with twelve quarters of leverage. A five-year agreement is a five-year obligation only for the party with less bargaining power, and memory vendors have never been that party at the top of a cycle. The durability of these agreements at scale is untested. The blockchain remembers; the architect forgets โ and the memory industry has forgotten 2019 twice already.
Run the normalization instead of the forecast. Samsung's semiconductor division has, across the 2015 to 2019 window, operated at margins materially below what a 55.5 trillion won figure implies for 2027. If operating margin reverts to that mean while revenue holds, the profit line compresses toward the high twenties in trillions, and the three-times multiple quietly becomes six. Six times normalized earnings for a conglomerate with a lagging foundry business and a consumer electronics division exposed to handset saturation is not a bargain. It is approximately fair. Nothing in the note survives the mean-reversion assumption, and the note does not test it. That omission is not dishonest; it is structural. Sell-side models price the cycle they can see.
The foreign ownership data deserves equal skepticism, because it is being deployed as sentiment when it is actually plumbing. Korean large-cap equity has been subject to persistent passive outflows, a won trading near cyclical lows against the dollar, and the perennial governance discount that has kept KOSPI below book value for most of a decade. Foreign ownership at a 2008 low is the arithmetic consequence of those flows, not a signal that offshore capital has independently concluded Samsung is expensive. Positioning artifacts and valuation verdicts look identical on a chart and mean opposite things.
This is where the blockchain lens becomes useful rather than decorative, and it is the part of the analysis that semiconductor desks systematically omit. The marginal buyer of Samsung common stock inside Korea and the marginal buyer of Korean-listed digital assets are, to a first approximation, the same cohort โ the same retail funding base, the same brokerage applications, the same leverage facilities, the same behavioral reflex to rotate between semiconductors and tokens depending on which is producing momentum that week. When foreign institutional ownership falls to a seventeen-year low, the price of the equity and the price of the token converge on a single domestic liquidity pool. That correlation is not visible in any DRAM model. It is visible the moment you stop treating the two markets as separate universes.
The segmentation mechanics are familiar to anyone who has studied the Kimchi premium. Capital controls and settlement friction create parallel price discovery for the same underlying asset in two jurisdictions, and the spread persists not because of information asymmetry but because of transfer cost. Korean equity now exhibits a mild version of the same structure: offshore capital cannot easily express a view, domestic capital is the only price-setter, and the resulting multiple reflects the constraints of the domestic pool rather than the cash flows of the underlying business. Three times earnings is what a stock costs when the buyers who would pay eight times cannot get capital across the border efficiently.
Which raises the obvious product gap. Tokenized equity venues have spent three years building wrappers for United States listed instruments โ the mechanics are now well understood, the custody models are documented, and the compliance perimeter, while uneven, is at least defined. What does not exist at institutional scale is a tokenized wrapper for Korean large-cap exposure. The demand-side logic writes itself: an offshore allocator who wants Samsung at three times earnings without navigating KOSPI settlement, Korean tax withholding, and foreign ownership registration has no instrument. I state the limitation honestly rather than overclaiming โ this market does not exist at scale, and any analysis asserting otherwise is selling a narrative rather than a security. But the gap is real, it is persistent, and it is precisely the kind of gap that gets filled when a valuation dislocation lasts long enough to justify the build.
The decentralized compute sector offers a second mirror, and a cautionary one. GPU rental markets price compute in real time, and HBM is the binding physical constraint on every one of them. If memory supply genuinely tightens through 2027, the cost of training and inference rises across the entire decentralized compute stack. Yet most of those networks own no hardware. They are coordination layers over rented capacity. A rise in the input price compresses their margin, not the vendor's. The on-chain compute complex is therefore a second derivative on the same physical constraint, and it is priced as though it were the constraint itself.
This is the recurring error of the AI-token complex: bidding the narrative and ignoring the balance sheet. The beneficiaries of a memory supercycle are the parties who own fabs, lithography, and qualification slots. Tokenholders who own neither captured nothing in 2023, captured nothing in 2024, and will capture nothing in 2027 regardless of how many whitepapers cite HBM bandwidth. The blockchain remembers the fundraising; the architect forgets the margin. A protocol that routes compute does not participate in a supply squeeze โ it absorbs it.
What the bulls actually get right is narrower than the headline and more durable than the bears admit. The supplier base has genuinely consolidated. There are effectively three qualified HBM producers, and qualification cycles at hyperscale are measured in quarters, not weeks. Capex per wafer has risen to the point where a new entrant requires a decade of sovereign-backed patience, and the process node ladder is visibly slowing. That combination โ three suppliers, expensive entry, slow node cadence โ has not existed in this industry since the DRAM consolidation a decade ago. On that axis, the analysts anchoring to 2019 are wrong. The structure has changed. What has not changed is the behavior of buyers when inventory turns.
Here is the correct reading of a three-times multiple. It is a probability-weighted instrument, and it embeds something like a forty percent chance of a normalization event inside twenty-four months. That is not irrationality. That is the market assigning a price to the historical base rate of memory cycles ending badly, adjusted for the new contractual structure. KB is arguing the base rate has fallen. The market is arguing it has not fallen far enough to justify eight times earnings. Both positions are defensible, and neither is obviously mispriced.
My own record demands I say this plainly. In 2020 I published a structural breakdown of a leveraged farming protocol's oracle dependency, and the community called me a bear three days before a flash loan drained ten million dollars. The lesson I took was not that pessimism pays. It was that a correct model with a wrong temporal horizon is operationally indistinguishable from a wrong model. KB's 157.1 trillion won has exactly that defect. It is directionally defensible and temporally unhedged. A forecast that cannot specify when it would be falsified is not an analysis; it is a position with a spreadsheet attached.
The falsification conditions are concrete and observable. First, disclosure of the actual contract price ladder inside those five-year agreements โ if the locked pricing is fixed at 2026 spot levels, the vendors have sold their upside; if it is indexed, they have not. Second, the trajectory of hyperscaler capex guidance across the next four quarters, weighted toward the memory bill of materials rather than headline totals. Third, HBM4 qualification outcomes, which will determine whether the three-supplier structure holds or fractures. Fourth, and most revealing, whether foreign ownership reverses from its seventeen-year low. A genuine re-rating requires the marginal buyer to change nationality, and that flow is measurable before the earnings arrive.
For the crypto-facing reader, the actionable asymmetry is uncomfortable but clean. The trade is not the AI token. The trade is the constraint. Anything adjacent to memory supply, substrate, packaging, or the capital equipment that gates them is a direct claim on the bottleneck. Anything that merely routes compute is a tax on the bottleneck. Every cycle produces a cohort of instruments that look like exposure and behave like expense, and the current AI-token complex is the largest such cohort since the 2021 infrastructure boom. The blockchain remembers the 2021 fundraising rounds; the architect forgets that almost none of them survived the following winter.
The deeper lesson sits above both semiconductors and tokens. What KB Securities has actually documented is a mismatch between the timescale on which physical capacity is built and the timescale on which capital is priced. Fabs take three years and ten billion dollars. Memory qualification takes four quarters. Tokenized exposure can be assembled in an afternoon. When those clocks desynchronize, price discovery stops describing value and starts describing the patience of whoever is left holding.
That is the accountability question, and it should be directed at the sell-side rather than the market. If a two-year window genuinely outearns twenty-six years of operating history, the burden of proof is on the note to demonstrate why the historical base rate no longer applies โ with the contract terms, not with the volume commitments. Until that disclosure exists, the three-times multiple is not an underestimation. It is a discount for uncertainty that has not yet been resolved, applied by a market that has watched this industry break its own promises twice. The blockchain remembers; the architect forgets โ and the architects of this forecast have not yet told us what they have chosen to forget.