Semiconductor shares always fall before the cycle turns. This time, three things may change the shape of the decline.

American memory-chip shares wobbled again ahead of Monday's open. Yet three shock-absorbers that did not exist in 2018—high-bandwidth memory sold on annual contracts, long-term supply deals that spread downside risk to customers, and a step-change in shareholder returns—may limit how far this downturn goes. Commodity DRAM could still follow the familiar cycle script, but it would be a mistake to assume the whole industry will travel the same road.

Three lines that capture this story

The rule of thumb that chip shares peak before the industry does has held again. In 2018, SK Hynix's share price topped out in May, roughly five months before DRAM contract prices began falling in October. This time, however, the shares have already fallen further from their peak than they did in that earlier cycle.

Three things are different now. HBM is sold on annual contracts; customers have signed long-term agreements that oblige them to share the pain of any downturn (as The Economist put it, customers will share the hangover); and both South Korean chipmakers have committed to materially higher shareholder returns.

The defences are not yet complete. SK Hynix's third-quarter results on 27 October—when additional return-of-capital measures are expected—and Samsung Electronics' board meeting at the end of the month are the first proper tests.

What happened on Friday, and what to watch on Monday

South Korea's KOSPI index rose 2.05% on 16 October to close at 5,668.05, snapping a seven-session losing streak. Samsung Electronics gained 2.62% to 215,500 won; SK Hynix added 0.79% to 1,272,000 won—both bouncing after plunges of 9.48% and 11.06% respectively the previous day. Foreign investors ended eight consecutive sessions of net selling and turned net buyers (a provisional 340 billion won as of 3:35pm).

The trouble came after Seoul closed. In American trading on 16 October, Micron, SanDisk, and SK Hynix's ADR all fell roughly 3%, reportedly on news that China's CXMT had made progress in high-density memory design. Goldman Sachs pushed back the same day, arguing that the market was overestimating CXMT's pace of technological catch-up and that the risk-reward on both Samsung and SK Hynix looked attractive. Because the American selloff occurred after the Korean market had closed, Seoul's memory stocks are likely to open lower on Monday—by how much cannot be known before the market opens. The more important variable is whether the foreign buying that emerged on Friday persists.

Shares have always moved first

It is worth acknowledging this industry's long-established pattern. The 2018 SK Hynix cycle is the clearest example.

The shares were already falling while earnings were still rising. The same dynamic could easily be playing out now. (Comparable figures for Samsung Electronics over the same period were not independently verified for this article.)

Markets are applying the old playbook to current prices

What is striking this time is that the share-price decline has already exceeded that precedent. In October, Samsung Electronics has fallen 19.74% and SK Hynix 28.38%; measured from their June peaks, the drops are 40.55% and 56.42% respectively (based on 16 October closing prices). Combined, the two companies have shed roughly 678 trillion won in market capitalisation this month alone. In 2018, a 40% fall from peak only arrived after DRAM prices and profits had already turned down. This time, the share prices have moved first—and sharply—even as third-quarter operating profits are forecast to reach 10.74 trillion won for Samsung (up 782.5% year-on-year) and roughly 7.7 trillion won for SK Hynix (based on consensus estimates).

Our reading is that the market has largely priced in a repeat of the old cycle's magnitude and speed. That is a judgement call, not a fact. The shares moving ahead of earnings does not guarantee that earnings will hold up. But for the old scenario to replay fully, the industry's structure would need to be the same as before—and it is not.

For context: applying the 16 October closing prices to 12-month forward EPS and book-value-per-share estimates (Samsung approximately 66,300 won and 157,300 won; SK Hynix approximately 449,800 won and 858,000 won) yields a forward price-to-earnings ratio of roughly 3.3 times for Samsung and 2.8 times for SK Hynix, and a price-to-book ratio of roughly 1.4 times and 1.5 times respectively. These figures do not reflect any subsequent consensus revisions. KB Securities has put Samsung's 2027 implied P/E at 3.1 times and SK Hynix's at 2.6 times. A low P/E does not signal a floor by itself, but as long as earnings forecasts hold, it suggests that a great deal of bad news is already in the price.

What is different this time: ① HBM

In the 2018 cycle, high-bandwidth memory (HBM)—the specialised, high-performance chips stacked in layers for use in AI servers—was not large enough to drive the industry's fortunes. Today it forms a distinct tier of the memory business. A Samsung Electronics executive has said that HBM wafers account for roughly 20% of global DRAM production capacity and could approach 30% next year; TrendForce sees the share rising from 18% in 2025 to around 30% by 2027.

The crucial difference is how HBM is sold. Unlike commodity DRAM, which trades on spot markets or short-term contracts, HBM is contracted annually—meaning its price adjusts far more slowly to demand shifts (TrendForce). Much of the volume is also pre-sold. Samsung said at its first-quarter results in April that all available HBM4 supply had been sold and that HBM revenues this year would more than triple year-on-year. SK Hynix is reported to have signed long-term supply agreements with around ten major customers. By market share, Counterpoint Research estimates SK Hynix held roughly 50% of the HBM market in the second quarter, Samsung 33%, and Micron 18% (figures are as reported; rounding means they sum to slightly over 100%).

The pricing outlook also looks favourable. Negotiations over HBM4—the next-generation product set to dominate in 2027—are said to have begun in the second quarter of this year. Samsung is reportedly asking for mid-to-high $4 per gigabit; those involved note this is an opening position, not a concluded price. Compared with the roughly $1.50 per gigabit cited for HBM3E, that would represent a more than threefold increase, according to the same reporting. TrendForce expects the blended average selling price across all HBM products to rise 121% by 2027 compared with this year (this does not mean every product rises by 121%). Some analysts also argue that HBM production is consuming so many standard DRAM wafers that commodity supply could tighten too, potentially disrupting the old cycle's logic even for mainstream chips.

The limits are real. HBM is the product most directly tied to the AI capital-spending boom, whose sustainability is precisely what this market correction is questioning. If AI investment falters, HBM will not be spared. The 2027 price is still being negotiated, and as Samsung and Micron ramp up HBM4 supply in the second half of this year, competitive pressure on pricing will build. HBM's precise contribution to overall profits was not confirmed for this article. The possibility that commodity DRAM follows the old cycle remains entirely open. But the fact that one axis of the business moves on annual contracts and pre-sold volumes is, by itself, enough to change the shape of any downturn.

What is different this time: ② shared pain

In a 15 October article asking whether the chip industry faces a great crash, The Economist identified three reasons why any downturn is unlikely to be as steep as the recent boom. First, oligopoly: the number of companies capable of making leading-edge memory has fallen from more than 20 in 1990 to just three—Samsung, SK Hynix, and Micron. China has produced a new entrant, but The Economist judged it unable to obtain the advanced equipment needed to rival the leaders. Second, contracts: chips used to be sold on spot markets or short-term deals, so customers could simply stop buying when demand fell, causing prices to collapse. Now, with supply constrained and suppliers few, customers are locking in multi-year agreements. Nvidia has committed more than $250 billion to secure supply through January 2029; Micron disclosed in September that take-or-pay contracts—most accompanied by upfront cash payments—lock in 35% of revenues (roughly $32 billion) through 2030. Third, balance-sheet strength: chipmakers now fund capital expenditure from operating cash flow rather than debt; Samsung's capex runs at roughly one-fifth of its operating cash flow, The Economist noted. The conclusion: if a downturn comes, customers will share the hangover.

South Korea's two champions are moving in the same direction. At its second-quarter results, Samsung said its target is to cover 60–70% of DRAM and NAND production capacity with long-term agreements (LTAs), that it had already signed deals with five global data-centre customers, and was in the final stages with five more large customers. The standard term is five years, with a rolling annual extension. SK Hynix operates similarly, on five-year base terms, with variations by customer, product, volume, and pricing mechanism.

Caveats apply. The Economist itself acknowledged that if AI spending stops, prices will fall, profits will shrink, and contracts struck at the top of the market will be reopened. The pain is shared, not eliminated. Micron's 35% means 65% remains unhedged. And DRAM prices being roughly six times their early-2023 lows implies a substantial margin for retreat. The confirmed public detail on LTA shares, upfront payments, and price-floor clauses at both Korean companies is limited. Customer pushback against sharp price increases is also cited as a headwind for the shares. For this reason, it is more accurate to say that the risk-sharing model is "expanding rapidly" than that it is "in place" at the Korean firms.

What is different this time: ③ shareholder returns

On 19 August, SK Hynix's board approved a policy of returning at least 50% of cumulative free cash flow over 2025–2027 to shareholders. It announced a buyback of 40 trillion won-worth of shares (roughly 3.3% of shares outstanding, or 24.07 million shares) between 20 August and 19 November, to be cancelled on completion, alongside a fixed annual dividend of 1,500 won per share. The scale of any additional returns will be disclosed with the third-quarter results on 27 October.

On 21 August, Samsung's board set a ceiling of 110 trillion won for shareholder returns this year and committed to a cash dividend of roughly 30 trillion won in the third quarter, including the regular quarterly dividend. Details will be finalised at the board meeting at the end of October. It would be the largest shareholder payout in South Korean corporate history.

Qualifications are needed here too. Of Samsung's 110 trillion won, only the 15 trillion won share buyback for employee compensation has been confirmed; the remaining dividend and cancellation details are deferred to a January board meeting. Some market participants had reportedly anticipated up to 200 trillion won, and the shares fell in after-hours trading immediately after the announcement. SK Hynix also suffered a roughly 10% drop in July after results disappointed those hoping for an immediate expansion of returns. Markets are demanding more—a signal that expectations remain high. It is also worth noting that Samsung's dividend income flows partly to the founding Lee family, which is reportedly using it to fund inheritance-tax obligations and increase its stake in the company.

The direction, however, is clear. Both companies have explicitly committed to returning more than half of free cash flow, and the question is whether cash generation is strong enough to sustain that commitment through a downturn. This is where the three factors converge: HBM and long-term contracts put a floor under earnings; shareholder-return commitments put a floor under cash distributions. If 27 October brings a larger-than-expected additional payout from SK Hynix, it could mark a turning point in how the market perceives the stock.

Signals that would force a rethink

We take the view that structural change will limit the depth of this decline—but the following developments would require a reassessment:

- Commodity DRAM contract prices turning negative, or evidence that CXMT's capacity expansion is materially affecting market prices - HBM4 price negotiations concluding well below Samsung's opening ask - Reports of customers seeking to renegotiate long-term contracts or reduce upfront payments - Sustained downward revisions to brokers' earnings estimates - A retreat in the scale or timing of shareholder returns - Widening stress in AI infrastructure financing (concerns about CoreWeave's creditworthiness were cited as a factor in the recent selloff)

Calendar: what to watch next week

Whether Friday's rebound marks a genuine turning point or merely a technical bounce will become clearer as these events unfold.