Context The earlier post Will AI Memory Demand Exceed Expectations? put demand into a base case of 45%, a beat case of 35% and a miss case of 20%. This piece re-tests the same question through the words and actions of the industry’s top supplier. It reads the public remarks SK Hynix Chairman Chey Tae-won made from the June 2, 2026 Computex through the July 17 Jeju Forum, together with the Nasdaq listing and capacity-expansion decisions he carried out in between, side by side in one place.
TL;DR
- Chairman Chey meets Jensen Huang and TSMC management directly and watches SK Hynix’s order book in real time, a level of access that public-market investors cannot easily reach. But the domain where his information edge is strongest is the direction of demand and supply constraints, not fair value and long-term margins.
- Two months of remarks compress into one sentence. AI memory volume and the duration of demand are stronger than expected, but keeping the ecosystem alive requires expanding supply and cutting token costs, so today’s scarcity pricing and excess margins should not be made permanent.
- His remarks therefore strengthen the company case for SK Hynix, but they do not equally strengthen the stock case that today’s margins persist for long.
- Actions say more than words. He issued roughly KRW 40 trillion of new shares near the highs, said capacity would double, and set token-cost reduction as the ultimate goal. All three are consistent with a volume-conversion strategy that presumes ASP has already passed its peak.
- The call is hold, add only on conditions, do not chase. Adding to the position is approved not by management remarks but by confirmation that 2028 earnings hold up, that long-term contracts actually convert into volume, and that ROIC holds after the expansion.
- The single indicator that decides direction has not changed. It is the DRAM contract price.
The industry keeps growing through 2030, the supplier’s margin peak has passed, and the stock price only trends up on a 2-3 year time axis. Run Chairman Chey’s remarks and actions through an interest filter, and what remains lines up exactly with the position we had already staked out: industry strength coexisting with a valuation boundary.
1. Why Read These Remarks
Chairman Chey sees material on customers’ long-term supply requests, actual production constraints, and government and Big Tech capital deployment before analyst reports do. He negotiates directly with NVIDIA, TSMC and US hyperscalers, and has access to internal yield, equipment, packaging, workforce and power plans. That makes his words premium material for confirming the direction of demand.
At the same time, an interest filter has to be applied. He is SK Hynix’s largest shareholder, a seller who just issued roughly KRW 40 trillion of new shares, and the party who made reassuring remarks right after the stock fell sharply from its highs. [Inference] The fact that his information edge is real, and the fact that the incentives of the person delivering that information are not neutral, both hold at once. That is why direction should be trusted while intensity is discounted.
This piece follows one rule. Separate information edge from the speaker’s interests, separate the company case from the stock case, and separate volume growth from price durability.
2. Remarks Timeline
From May 19 to July 19, 2026, this section gathers, in chronological order, the public remarks and actions capable of moving an investment call. The information grade runs higher the closer a source sits to a primary-source interview, and lower the more it relies on secondhand event coverage or secondary commentary.
| Date | Venue | Confirmed key remarks and actions | Information grade |
|---|---|---|---|
| 2026-06-02 | Computex, Taipei | Total wafer capacity to double within 5 years, possible bottleneck through 2030, new fab lead time of at least 3 years, supply whatever is needed without pre-calculating a capex ceiling, Jensen Huang requested increased HBM4E wafer output, best-ever cooperation with TSMC | A-/B+ |
| 2026-06-11 | Nikkei interview | Evaluated Japan as an overseas fab candidate, emphasized the power, materials and equipment ecosystem, reviewing a Japan AI factory, accelerating Yongin completion | B+ |
| 2026-06-14 | Icheon Forum | The essence of AX is operational improvement, one agent per person, emphasized a full stack spanning memory, data centers, energy and electrification | A |
| 2026-06-29 | Great Leap megaproject | 15GW AI data center, KRW 1,100 trillion memory investment roadmap, target of completing 4 Yongin fabs by 2033, expanding the Cheongju and southwest production belt | A |
| 2026-07-10 | Nasdaq listing, CNBC, Six Five, New York press conference | ADR listing of roughly $26.5bn in new shares (largest ever by a foreign company), agents and robots are changing the structure of demand, customers are requesting supply and long-term plans that far exceed SK’s own expectations, cutting token costs, funding and geopolitics are the key five-year risks | A-/B+ |
| 2026-07-10 ~ 14 | Six Five, Bloomberg | The biggest five-year worry is a break in AI funding supply, the stock market may have a bubble while AI technology is real and will be reflected in actual share prices in 2-3 years, if memory prices rise too much they will hurt traditional industries such as retail and autos | A- |
| 2026-07-17 | Jeju Forum | Given time the stock price trends up, so hold rather than trade (while also saying he does not know), AI is still 4 years old, Korea cannot beat the US on token cost and quality so it should target niches in infrastructure and apps | B |
Per CNBC’s reporting, the HBM volume customers requested was 5-6x SK’s own expectations, and next year’s semiconductor demand growth rate was cited at a minimum of +50-60%. [Blocked] These figures, however, rest on event coverage rather than a primary transcript, and the 5-6x figure is the customers’ wish list, not confirmed orders.
3. The Recurring Mental Model
The thinking that runs through two months of remarks stacks into three layers.
Structural Thesis: The End of the Cycle
[Fact] He sees the era in which memory demand was tied to population and hardware unit counts as over. AI agents, robots and physical AI consume memory regardless of headcount. [Inference] He discards the model that calculated memory’s ceiling by multiplying PC and smartphone shipment volumes by content per unit, and switches to a model that multiplies the number of AI workloads by state, context and model memory per workload, then multiplies again by execution frequency. His arithmetic is that demand growth outruns the 3-year pace of fab construction all the way through 2030. The one micro-level piece of evidence he repeats is KV caching, the shift in the makeup of demand toward inference occupying memory persistently.
Moving Up the Value Chain: Turning a Forecast into a Business
[Fact] His ultimate goal is not capacity itself but cutting the cost per token. He stresses that AI services are too expensive and that chips, data centers and technological innovation must bring the cost down. [Inference] He is trying to shift cyclical exposure from chip sales into an annuity-like structure built on memory-as-a-service (MaaS) and the $1 trillion data-center concept. Rather than leaving the end of the cycle as a forecast, he treats it as a task to be engineered directly into the business structure.
Admitting the Risk: Funding and Geopolitics Over Technology
[Fact] As the biggest risk over the next five years, he named not technology but a break in the supply of AI funding, along with geopolitical and energy-price shocks. [Inference] Even if AI demand is technically valid, if hyperscalers, governments and capital markets fail to keep funding it, orders and fab schedules slip together. This is the single most important falsification condition he has admitted himself.
Underneath these three layers, three views run through consistently: efficiency gains read as a Jevons effect rather than demand destruction (memory per task falls, but total usage grows even faster); the bottleneck moving from a single component to the entire stack including power, water and packaging; and customer survival ranked ahead of the supplier’s own long-term profit. The last view is decisive. The logic that there is no long-term volume if customers collapse is equivalent to a declaration that he will not make a quarterly operating margin in the 70% range permanent.
4. What to Believe and What to Discount
Even remarks from the same person carry different levels of reliability. The more an admission runs against his own interest, the more it deserves belief; the more an optimistic claim favors his own position, the more it should be discounted.
Most credible: admissions against his own interest. His remark that traditional industries get hurt if prices rise too much is the top supplier executive admitting, on his own, the axis of ASP self-cannibalization. It mirrors TSMC management’s comment that component prices hit set demand. His remark that the stock market may have a bubble and that it will only be reflected in the actual share price 2-3 years later amounts to the man himself conceding our own view that industry strength and a valuation boundary coexist. His remark that a funding break is the biggest worry, and his remark that he does not know either, also belong in this category.
Middle tier: real readings that may be inflated. The claim that customers requested 5-6x and exceeded expectations could be an actual reading of the order book, but it could also be a figure inflated by duplicate bookings during a shortage. The double-ordering seen in past telecom-equipment boom cycles stands as a warning that still applies.
Lowest tier: position talk. The phrase that the cycle has ended is language heard at every peak, and telling investors to hold because the price trends upward is a reassuring statement made by the largest shareholder right after a sharp drop. Neither should be used as investment grounds.
The information edge itself breaks down by domain as follows.
| Domain | Information edge | Reason |
|---|---|---|
| Direction of customers’ supply requests | Very high | Direct negotiation with NVIDIA, TSMC, Big Tech and AI companies |
| HBM roadmap and production bottlenecks | Very high | Access to internal yield, equipment, packaging, workforce and power plans |
| Global footprint and policy pressure | Very high | Direct contact with the Korean, US and Japanese governments, customers and capital markets |
| Five-year capital-deployment commitment | High | Doubling capacity and regional investment plans reflect actual board and funding decisions |
| Actual demand behind the customers’ 5-6x request | Medium | Could be an unconstrained request, duplicate bookings, or negotiation-driven demand |
| Precise scale of the shortage through 2030 | Medium | Sensitive to fab yield, bit growth, customer capex and efficiency gains |
| SK Hynix fair value | Low | Conflict of interest from the listing, large-scale capex and managing shareholder expectations |
| Durability of the current excess margin | Low | He himself has set supply expansion and cost cuts as the goal |
5. Actions Speak Louder Than Words
Actions carry incentive-weighted signal and are more honest than words. Every action taken over the two months pointed in the same direction.
Roughly KRW 40 trillion of new shares issued near the highs. Per press reporting, the Nasdaq offering price was set at roughly a 2.9% premium to the Korean closing price, and the first-day closing price was $168.49, up about 13.1%. [Inference] Executing the largest equity raise in the company’s history, in new shares priced above the domestic closing price, is observationally equivalent to a capital-allocation judgment that the stock was priced well for issuance at that moment. Then, starting July 13, the first Korean trading day right after the listing, a cascade of double-digit single-day declines for the stock began. The new-share supply, the listing catalyst running dry, and ADR arbitrage overlap in timing, but causation is not asserted here.
Supplying whatever is needed, no capex ceiling. This is the man himself directly disproving the defensive argument that disciplined expansion protects price. Because doubling capacity includes commodity product too, accelerating the expansion reads almost like a confession that he cannot make money on price anymore, so he will make it on volume instead.
A $1 trillion data center, memory-as-a-service, and token-cost cuts. All three are business shifts that presume chip ASP will decline. It means he has staked his bet on lowering costs to grow total demand.
6. The Most Important Separation: Volume and Margin
The market mostly reads his remarks as saying the shortage lasts a long time. What matters more is separating a volume supercycle from a margin supercycle.
| Item | What the remarks imply | Investment read |
|---|---|---|
| Memory bit demand | Structural increase | Strongly positive |
| Customer supply requests | Exceed current capacity | Positive, but contracts need verification |
| Supply lead time | 3-5 years | Price protection through 2026-2027 |
| Long-term ASP | Needs to ease for customer sustainability | Neutral to negative |
| Long-term operating margin | Can decline from expansion, depreciation and price normalization | Negative, under watch |
| Market share | Defended by large-scale capacity and HBM execution | Positive |
| Shareholder ROIC | Sensitive to investment scale and funding | Undetermined |
[Inference] The odds have risen that SK Hynix becomes a company whose revenue and bit volume grow for a long time, but the evidence that it stays a company with today’s extreme margins for a long time has actually weakened. That is why a low PER alone cannot establish that the stock is undervalued. The center of gravity has moved entirely from whether a collapse happens to how long peak margin can last.
7. Where the Stock and Consensus Stand Today
The numbers below confirm where things stand. All figures follow the local database and consensus, with the reference date noted.
| Item | Value | Reference date |
|---|---|---|
| Closing price | KRW 1,842,000 | 2026-07-16 |
| Return since 2026-05-19 | +5.6% | 2026-07-16 |
| Vs. 2-month high | -36.9% | High of KRW 2,919,000 |
| FY2026 PER | 5.85x | 2026-07-16 |
| FY2027 PER | 4.20x | 2026-07-16 |
| FY2027 EPS | KRW 438,114 | 2026-07-16 |
| FY2028 EPS | KRW 427,332 | 2026-07-16 |
| Average target price | KRW 3,547,917 | 2026-07-16 |
| Upside to target price | +92.6% | Own calculation |
[Inference] The drop from the recent high is a large -36.9%, but the stock is still up +5.6% versus May 19. This is not an early undervaluation zone but an adjustment zone that re-examines terminal margin after a large re-rating has already happened. An FY2027 PER of 4.20x looks cheap, but that is the market discounting the possibility that 2027 earnings mark the peak. The average target price rose from KRW 2,189,167 on May 19 to KRW 3,547,917 on July 16, up about 62.1%, and a target price that jumped this quickly in the wake of results should not be used as an expected return as-is. The fact that FY2028 EPS is 2.5% below FY2027 shows that consensus itself already reflects a stall and normalization in 2028 earnings. On a book-value multiple, trailing PBR is about 3.8x, meaning his upward trend is justified only on a 2-3 year time axis.
8. Scenarios
The probabilities below are not statistical values. They are inferences for investment decision-making that reflect public information.
| Scenario | Probability | Industry path | Implication for SK Hynix stock |
|---|---|---|---|
| A. Demand overshoot with supply discipline | 40% | AI workloads surge, long-term contracts convert into real volume, 2028 capacity also gets absorbed | Earnings and multiple both hold, strongest upside |
| B. Volume growth with margin normalization | 40% | Volume stays strong, but supply expansion, customer price pressure and rising depreciation weigh in | PER and margin compress even as revenue grows, the path most likely to be underrated right now |
| C. Funding and geopolitical shock | 20% | AI capex funding is disrupted, power and construction are delayed, duplicate orders get corrected | Orders, prices and supply-demand all get impaired at once, the case needs to be scaled back |
Two months of remarks raise the combined probability of A and B, meaning strong volume persistence. But what separates A from B is not demand, it is supply discipline and ROIC. His statement that he will supply whatever is needed adds more weight to B than to A.
9. Investment Execution
The call is hold, add only on conditions, do not chase.
There are four reasons not to buy more right now. Management’s remarks are concentrated right around the listing and the large-scale capital raise; a Q1 operating margin of 72% and an extremely low FY2027 PER are numbers that carry substantial normalization risk; doubling capacity over five years defends market share while simultaneously creating depreciation, dilution and ROIC risk; and the customers’ 5-6x request is not a confirmed order.
Adding to the position is considered once at least three of the following four conditions are met.
- FY2028 earnings consensus stops falling relative to FY2027 and turns upward.
- Customer qualification, volume, upfront payments or long-term contracts for HBM4E and next-generation HBM are confirmed through official disclosures.
- Even as wafer and bit supply rise in 2027-2028, DRAM and HBM contract prices and customer inventories remain stable.
- Mid-term ROIC, free cash flow and shareholder-return policy are not impaired even after the large-scale capex announcement.
Conversely, if two or more of the following occur, cut the stock case rather than the company case. Customer requests fail to convert into orders and long-term contracts and bookings get canceled instead; forecasts repeatedly confirm that 2028 DRAM bit supply growth exceeds demand growth by 5pp or more; share or premium erodes from yield and qualification delays on HBM4 and HBM4E; earnings get revised down on depreciation and localization costs while capex keeps rising; or two or more hyperscalers cut their AI capex or server installation plans.
The read differs by memory layer.
| Layer | Read | Investment point |
|---|---|---|
| HBM | Strongest structural growth, also the highest expectations | Watch for leadership being maintained, and for HBM4 and HBM4E volume, yield and customer diversification rather than ASP |
| Server DDR and SOCAMM | Relatively undervalued | Watch agent and host-memory adoption spreading, and confirm the adopting platforms and actual shipments |
| LPDDR | Dual demand from mobile and AI servers | Benefits from low-power inference, watch for set demand erosion from high prices |
| NAND and eSSD | Second structural axis | Watch KV cache and RAG demand, checkpoint-storage demand, and confirm the high-capacity product mix and pricing |
| HBM packaging and equipment | Rising capex visibility | Exclude equipment names with lumpy revenue from the core position, prioritize recurring-revenue materials and consumables |
10. Checking the Counterarguments
The more a call leans bullish, the harder its counterarguments need to be stress-tested.
- Conflict of interest in the speaker. This is demand optimism voiced right after a Nasdaq listing and a large-scale capital raise. Even if the direction holds, the intensity may be overstated.
- The error in the 5-6x figure. Customers’ wished-for supply volume is not the same as demand that can actually be installed, powered and monetized.
- The trap in policy-driven capex. The stronger the national-security and regional-balance logic gets, the weaker the ROIC discipline on individual projects can become.
- Jevons failure. If the lower token cost from efficiency gains fails to lift usage enough, memory bit demand comes in below expectations.
- China’s double edge. The spread of lower-performance Chinese chips increases commodity bit volume, but it also creates supply competition and price declines in commodity DRAM and NAND.
- Customer in-housing. ASIC and memory optimization can reduce HBM content per unit and the supplier’s premium.
- Funding and geopolitics. This is the axis Chairman Chey himself named as the bigger five-year risk, ahead of technology.
11. The July 30 Earnings Readout
The next window for testing these remarks is earnings season around July 30, 2026. Four things to read.
- Whether Samsung Electronics’ earnings call acknowledges resistance in commodity demand. If it mirrors TSMC’s remarks on component pricing, the signal that ASP has passed its peak gets stronger.
- Whether the Q3 contract price holds the +40-50% increase that part of the market has floated. This is the direct discriminator for the Chairman’s thesis.
- Whether SK Hynix’s own call fleshes out the capacity-doubling roadmap and commodity expansion. If it does, the margin-normalization scenario gains weight.
- Whether memory-as-a-service and long-term supply-contract structures get mentioned. If they do, watch early for the phase where price risk converts into credit risk.
The single indicator that decides direction never changes. If the DRAM contract price rolls over, the market is rejecting the 2030-bottleneck thesis; if it holds, Chairman Chey’s arithmetic wins.
The stocks named in this post are examples for analysis and do not constitute a recommendation to buy or sell any specific security. Responsibility for investment decisions and their outcomes rests with the investor. Quoted remarks are based on press reporting and interviews, and primary transcripts were not obtained for some of them. The customers’ 5-6x request is a wish-list volume that has not been verified through confirmed orders, long-term contracts or upfront payments. The causal link between the July 13 sell-off and the Nasdaq listing is a timing-and-volume-consistent observation, not a proven causation. Scenario probabilities are not statistical values but inferences for investment judgment that reflect public information, and prices and consensus follow the local database as of July 16, 2026, with no later moves reflected.
Related posts
- Will AI Memory Demand Exceed Expectations? Reading the Over-Growth Odds Through Demand Scenarios and the Supply Map
- Are Semiconductors Cyclical, and What Is Fair Value? Pricing Samsung, SK Hynix and Micron with FCFE and Normalized Earnings
- The Real Debate in Semiconductors: Four Physical Clocks and One Stock-Price Clock
- SK Hynix Q2 Earnings Cut but Target Prices Held: Synthesizing the Mirae Asset and KIS Reports