Issue 002 · The HBM chokepoint

We like SK Hynix. We think memory is one of the genuine chokepoints of this AI build-out, and we hold that view more firmly than most of the desks now upgrading the stock. We are still not buying it at ₩2,521,000. It is up 248% this year and 587% over twelve months, and the easy money has been made.

Both of those things are true at once. The demand for high-bandwidth memory is real, it is locked in under contract, and it is sold out for 2026. None of that is in dispute. The question we actually care about is narrower: at this price, what is the market already paying you to expect? Our answer is that it is paying for the peak — peak earnings, peak valuation, and peak market share, all in the same quarter.

So here is what we would do instead of buying it today. We are not chasing it, and we are not shorting a business this good. We wait. Section 06 sets out the price at which we become buyers; Section 07 sets out the two things that would make us sell rather than buy. Until one of those prints, the position is simple: add on weakness, not a fresh buy at today’s quote, and not a lazy HOLD.

Every note we write reasons from the same prior: the AI build-out is a default trajectory, not a forecast. Effective compute — very roughly, the useful computing power thrown at AI — is rising about half an order of magnitude a year. Annual AI investment is climbing fast toward a $7.6T baseline over 2026 to 2031, with hyperscaler capex alone guided to ~$700B in 2026. The surge is front-loaded into this decade. We pass through more OOMs (orders of magnitude — 10× jumps in scale) between 2024 and 2030 than in all the decades that follow. The exponential is in motion, the clusters are already signed, and we put the burden of proof on the sceptic.

Inside a build-out this large, the money is made at the bottlenecks. Whenever one component is the thing everyone is short of, whoever controls it earns outsized profits — for exactly as long as they remain the only real source. First it was the GPUs themselves. Now, increasingly, it is the memory that feeds them.

Memory is now a chokepoint, by name. Situational Awareness names HBM and CoWoS advanced packaging as “already key bottlenecks… the primary constraint,” and projects memory rising to potentially more than 20% of cluster capex after the GPU itself. We hold that view. SK Hynix sits at the memory layer, in the HBM sub-layer, and it owns the largest single share of the most strategic product in the AI supply chain.

That is the bull case. The honest question, and the one this note actually answers, is what comes after. A chokepoint is only as good as its monopoly. The rent gets competed away wherever the constraint is contestable. So for any memory name the decisive question is not “is demand strong,” because we assume it is, but “how single-sourced is the specific thing this company sells, and for how long?”

For SK Hynix, the answer is: very single-sourced today, but visibly less so tomorrow. That is the entire note.

This is the chart that matters. SK Hynix’s HBM bit-share has already gone from 62% in Q2 2025 to ~57% in Q4 2025, and every force pulling it lower from here is now public. The same sell-side models that have SK Hynix leading today have its share eroding further as Samsung’s HBM4 ramps. The lead is melting on the industry’s own numbers, not on bear-case speculation.

Three things put the melt in motion. First, Samsung began mass-producing HBM4 in February 2026 and has shipped early samples to NVIDIA. Second, Micron sold out its 2026 HBM capacity and is ramping HBM4 a quarter early in Q2 2026. Third, NVIDIA certified all three vendors for HBM4 supply to the Vera Rubin platform on 5 June 2026. The stock fell roughly 10% on the multi-sourcing news.

None of this means SK Hynix is a bad business. It means the lead is a lead, not a lock. The difference matters for valuation. A moat is a permanent feature of a business; a lead is a snapshot, and snapshots decay.

The lazy version of the chokepoint thesis says: SK Hynix makes the memory, SK Hynix owns the chokepoint, the rent accrues to SK Hynix. The honest version says the chokepoint is a triad, and the durable rent is migrating upstream.

Here is that picture in detail. HBM is built from three layers. The HBM stack itself, made by SK Hynix, Samsung, and Micron. The CoWoS advanced packaging that bonds the HBM to the logic die, made almost entirely by TSMC, fully booked with 52 to 78 week lead times, scaling from ~35,000 wafers per month in late 2024 to a projected 130,000 WPM by end-2026. And the logic base die that the HBM sits on, also made by TSMC, since HBM4 uses a 12nm TSMC-fabricated die designed by NVIDIA. NVIDIA occupies more than 70% of TSMC’s CoWoS-L capacity.

The rent in the triad splits roughly like this. SK Hynix captures the HBM bit-share and the premium pricing that bit-share commands. But the CoWoS gate, the base-die fab, and the EUV scanners behind all of it sit with TSMC and the equipment oligopoly (ASML, principally). NVIDIA, as monopsonist buyer, designs the HBM4 specification and dictates the base die. The bargaining power is asymmetric and tilts against SK Hynix over time. Downstream, hyperscaler long-term agreements pre-empt roughly half of SK Hynix’s revenue at contracted prices. That is a floor on volume, but it is also a ceiling on rent: contracted pricing renegotiates in the customer’s favour more often than the seller’s.

For readers who want the chokepoint with the least contestable rent, the upstream layers are arguably the cleaner trade. CoWoS and the base die sit with TSMC. EUV sits with ASML. Both have longer lock-in, fewer competitors, and less customer-concentration risk than any single HBM vendor. We hold SK Hynix for the idiosyncratic kicker, the purest HBM read-through, Korea-beta exposure, and a possible US-listing catalyst, with eyes open to share decay.

The Situational Awareness prior is a compute prior, not a memory-intensity prior. It says effective compute is rising 0.5 OOMs per year. It does not say HBM per unit of compute is rising. That part is a bet.

Two forces pull in opposite directions on how much HBM each unit of computing work needs. The bearish force is algorithmic efficiency — the steady progress that lets each new generation of models do more with less hardware. KV-cache compression, paged attention, multi-latent attention, mixture-of-experts, 4-bit and 8-bit quantisation: each of these is a technique that squeezes the memory needed per unit of work. A model that runs on half the HBM is, from SK Hynix’s point of view, a model that buys half the HBM.

The bullish force is Jevons. Long-context models, agentic workloads, reasoning chains, multi-modal inference. Each raises HBM per query, because the system has to hold more state in working memory while it thinks. The bet is that Jevons wins, that the demand for intelligence per dollar rises faster than algorithmic efficiency per dollar. We hold that bet, with caveats.

The honest horizon is not “structural for the decade.” It is structural 2025 to 2028, with a high-probability air-pocket, then cyclical re-assertion. The sell-side’s own DRAM price models already price the cycle turning by 2029, with DRAM ASPs modelled to fall as three-vendor capacity lands at once. If algorithmic efficiency compounds faster than Jevons, the air-pocket arrives earlier. If a hyperscaler capex pause of more than six months lands on the trip-wire of the ~3x capex-to-AI-services-revenue ratio, the air-pocket arrives in 2027. We are not pricing 2030. We are pricing the next three years.

This is the part of the chokepoint thesis that the consensus does not write about. The demand for memory is real. The demand for more memory per unit of compute is a separate question, and it is not guaranteed.

The price is paying for the peak on three independent axes.

Peak earnings. Q1 2026 operating profit of ₩37.6tn nearly matched all of FY2025 (₩47.2tn). DRAM ASPs rose mid-60% quarter on quarter in Q1; NAND ASPs rose mid-70%. The Q1 print is the most extreme in memory history, and the entire 6.6× forward P/E sits on a denominator the company itself will almost certainly not repeat. Market cap of ~₩1,786tn implies forward net income of roughly ₩271tn (mcap / 6.6×) versus ~₩79tn trailing (mcap / 22.5×). The price requires net income to roughly triple off the trailing base and hold there. The forward P/E is a peak-earnings denominator, not a cheap multiple.

Peak P/B. ~10.9× trailing book, outside SK Hynix’s historic ~1 to 2.5× band by a factor of four or more. The normalised ROE of 53.18% must persist to justify the multiple, and ROE is a mean-reverting series in this industry. Forward P/B re-rates to roughly 2.1× as the FY2026E book builds (current ~₩232k BVPS plus retained earnings), which is more defensible, but still assumes that ROE persists. Re-rating from trailing to forward is a real effect, but it does not save the entry: it makes the over-statement less severe.

Peak HBM share. ~62% in Q2 2025, ~57% in Q4 2025, with three-vendor HBM4 capacity ramping through 2026 and 2027. The lead is melting. SK Hynix is about to start earning peak margins on a falling share of the most strategic product it sells.

The sensitivity grid is the cleanest tell. Our base case sits ~31% below the current price. The bull case barely clears it (+9%). The bear is severe (−60%). And the bulls’ own published bear scenarios, including the larger Korean and global houses, sit at ₩1.2M to ₩2.0M, all below the current price. You can be right on the thesis and lose 30% on the entry. The asymmetry is the wrong way.

ACCUMULATE means we have a price at which we will be a buyer, and we have published it.

Trigger to add: a pullback toward the ~₩1.7M base case (a drawdown of roughly a third from the current price), provided two conditions hold. (1) SK Hynix’s HBM4 share is still at or above 40%, indicating the moat has not collapsed. (2) No HBM contract-price cut has printed, indicating the supercycle has not ended.

The next two earnings prints (Q3 and Q4 2026) are the binary read on HBM4 share, not on 2027 numbers. We will be buyers at the price the oligopoly produces, not at the price the consensus is paying. Pay the base case for the structural bull’s upside.

What it is worth now, and what a re-rating would change. On our base case the shares are worth about ₩1.7M today. At ₩2,521,000 the market is paying roughly 45% above that — the same gap the sensitivity grid shows as our base sitting ~31% below the current price — and only the bull case (₩2,740,000) justifies it. A genuine re-rating would move our number, not just the price: if HBM holds its premium, the US listing lands, and the market accepts that memory has partly de-cyclicalised, fair value migrates up toward the bull. We put ~60% on the base-case path — an air-pocket and a pullback toward ₩1.7M inside 12 to 24 months — and ~40% on the re-rate path. At those odds, today’s price is already paying you for the re-rate before it has happened.

We will change the position the day any of the following fires. We are stating it now, in writing, so that the discipline is public.

The kill-switch has already partially fired. NVIDIA’s 5 June 2026 three-vendor certification and Samsung’s Q1 2026 DRAM share lead (38% versus SK Hynix at 29%) are facts, not forecasts. We are not exiting. We are flagging that the conditions for a re-rating are now on the page, not in the future.

Why hold, and not move to underweight today? Three reasons. First, the business is genuinely excellent and the order book is sold out, so shorting it now is a bet against contracted demand. Second, the timing of the air-pocket is unknowable, and being early on a cyclical short is indistinguishable from being wrong. Third, the two events that would actually break the thesis have not printed — the conditions for a re-rating are visible, but they are not yet facts. Against the other ideas on our sheet, SK Hynix earns its place as the cleanest, highest-share read on the memory chokepoint, with an idiosyncratic US-listing catalyst on top. What it does not earn, at this entry, is fresh capital.

If (a) or (b) prints, we move to UNDERWEIGHT on the position, sell the cyclical, and revisit only when the air-pocket has cleared.

Bull (what we believe)

Bear (what we also believe)

Both halves are real. The whole call is which half the price is paying for.

HBM ASPs. Not revenue. Not share. Not earnings. The contract price of HBM, quarter on quarter. Volume misses can be explained by mix, customer timing, or inventory. A price cut means the cycle has turned.

The first house or channel check that reports an HBM ASP cut QoQ is the print we publish a position change on. That is the discipline. The chokepoint is real, the rent is contestable, and the entry is the peak. We will be buyers at the price the oligopoly produces, not at the price the consensus is paying.