The 2016-2026 capex cycle, ten years of returns from Samsung, SK Hynix and Micron, and a comparison with Apple at its peak
The memory 3 are in a window, unprecedented in any prior memory cycle, of executing record capex and holding record return capacity at the same time. The capacity itself is shared, but the timing, ratio and method of payment differ -- and that difference has already begun showing up in first-week-of-August stock prices.
Record quarterly results were confirmed -- Samsung's operating income of 89.5 trillion won, SK Hynix's 60.5 trillion won, and Micron's $33.7 billion (about 47.8 trillion won), combined roughly 198 trillion won -- alongside Micron's December 9, 2026 buyback restart, SK Hynix's flagged third-quarter return announcement, and Samsung's CFO signaling its next policy.
Return capacity itself is ample at all three. On policy specificity, Micron (full amount, every year, on a fixed date) is ahead of Samsung and SK Hynix (half, settled after three years), and that gap showed up in first-week-of-August stock moves (Micron +7.1% vs. Samsung -12.0% and SK Hynix -17.2%). In a downturn scenario, Samsung's regular dividend of 9.8 trillion won and SK Hynix's fixed dividend provide an explicit floor, while Micron's dividend floor is effectively immaterial against its market cap.
Micron's momentum-index exclusion in the third week of September (more than $2.3 billion of rebalancing sales) -> Micron's buyback restart on December 9 -> SK Hynix's additional return announcement in Q3 -> Samsung's next three-year policy (2027-29, expected around year-end/early next year) -> whether the historical pattern of a market collapse the year after a capex peak repeats in 2027.
Before discussing capacity, the history capex needs to be laid out first. What extinguished shareholder returns at the three memory makers over the past decade was not weak demand -- it was capex. Overlaying the ten-year trajectory of combined capex with DRAM market size reveals the shape of the cycle.
Ten-year capex trajectory of the memory 3 -- Samsung DS division facility investment, SK Hynix and Micron annual capex, converted to trillion won
DRAM market size vs. combined capex -- capex is a lagging function of price. Both market collapses came the year after a capex peak
In the second half of 2016, server demand and supply constraints combined to lift the DRAM market from $41 billion in 2016 to $99 billion in 2018 -- a 2.4x jump in two years. All three companies plowed profits into capex. Samsung's semiconductor facility investment hit a then-record 27.3 trillion won in 2017; SK Hynix raised capex 65% to 17.0 trillion won in 2018; Micron's fiscal 2018 capex reached $8.9 billion, also a 2.4x rise in two years. Combined capex peaked in 2018 at 53 trillion won, and the DRAM market collapsed the very next year, 2019 (-37%) -- a lag created by the 12-18 month lead time between fab construction and market supply.
It was in this window that Micron made its May 2018 capital return pledge: 50% of returned plus a $10 billion buyback. At the time of the announcement, fiscal 2018 FCF was $8.5 billion; by fiscal 2020, FCF had fallen to effectively zero. The buyback program, after spending $2.7 billion in fiscal 2019, became largely dormant. It was the first case of a return policy designed around peak FCF disappearing along with the cycle.
When the market re-expanded on stay-at-home demand in 2020-21 (reaching $94 billion in 2021), capex rose again. Samsung set consecutive records at 43.6 trillion won in 2021 and 47.9 trillion won in 2022; SK Hynix spent 19.0 trillion won in 2022; Micron spent $12.0 billion in fiscal 2022. Combined capex peaked in 2022 at 84 trillion won. The market collapsed the following year, 2023 (-35%).
This time, the response diverged. SK Hynix cut 2023 capex 65% to 6.6 trillion won and moved to reduce output; Micron cut fiscal 2023 capex 42% to $7.0 billion. Only Samsung held investment steady at 48.3 trillion won, defending market share. That production cut became the starting point for the 2024-25 supply shortage and the current cycle. On the return side, 2023 was the year SK Hynix maintained its fixed dividend of 1,200 won even while posting a loss, and the year Micron kept its dividend but effectively suspended buybacks -- the first real-world test of the fixed-dividend floor's credibility.
In 2024, HBM (high bandwidth memory) demand restored the market to $97 billion, and 2025-26 has become a broad shortage extending even to conventional server DRAM. The capex response has been steeper than in either prior cycle. SK Hynix told its July earnings call it expects 2026 investment in the high-40 trillion won range (up in the high 50% area from 30.2 trillion won a year earlier; internal estimates put the figure near 47 trillion won). Micron's CFO confirmed fiscal 2027 capex only as "mid-$40 billion range or higher" -- nearly double fiscal 2026's $27 billion. Samsung does not issue separate guidance but is trending upward with the new Pyeongtaek fab and the second Taylor phase. Combined capex across the three has more than doubled from the 2023 trough of 65 trillion won in three years.
Capex intensity (capex divided by revenue, %) -- revenue is growing faster, so intensity is actually falling. Both prior downturns had intensity above 40% beforehand
In absolute terms this expansion is the largest ever, but capex intensity relative to revenue tells a different story. In the year before each of the two prior downturns, Micron's and SK Hynix's capex-to-revenue ratio exceeded 40% (Micron 39% in fiscal 2019, 45% in fiscal 2023; SK Hynix 47% in 2019, 43% in 2022). Currently, revenue is growing faster than capex, so intensity has actually fallen into the low-30% range. Micron's fiscal 2026 capex of $27 billion against roughly $120 billion of revenue (reflecting fourth-quarter guidance) puts intensity at 22-23%; even placing fiscal 2027 capex of $45 billion-plus against $200 billion of revenue yields 22-25%. This cycle's capex is not yet outrunning revenue in the pattern of past overinvestment.
The limits of this metric matter, though. The denominator -- revenue -- reflects today's peak prices; once prices normalize, intensity will spike automatically. The 2019 spike in intensity, too, was produced by a revenue collapse, not a capex increase. Capex intensity is a gauge of whether investment looks excessive right now, not a guarantee of future safety.
Mechanically applying the historical rule -- market collapse the year after a capex peak -- points to 2027-28 as the test window. But this cycle has two structural features absent from the past. First, LTAs (long-term agreements) and SCAs (strategic capacity agreements) have locked in volume and floor prices by , changing the speed at which supply responses reach price. Second, the source of demand is not the consumer but a capital-allocation decision made inside Big Tech (M7) capex. A different structure does not mean no cycle at all. Contracts only create a lag; combined capex above 135 trillion won will eventually convert into physical production capacity.
Sign in with Google — your tier will be checked automatically and access granted if eligible.
Sign in with GoogleThis site runs on ads — the tier system rewards community contributions.
This report is provided for informational purposes only and does not constitute a recommendation to buy or sell any financial instrument. Investment decisions should be made based on your own judgment and responsibility. The analysis and opinions contained herein are based on information available at the time of writing and are subject to change.
