06-reference/research

sk hynix capex intensity reconciliation

2026-10-06·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
memory-cyclecapital-cyclecapexmeasurement-definitionsk-hynix

Every plausible definition of SK Hynix capex intensity lands within 3 points of 28% — so the 11%-vs-mid-30% gap is not a definitional one, and capex/revenue should be demoted from the discipline threshold

The question

Verbatim: "Resolve the SK Hynix capex-intensity discrepancy (~11% Futurum vs 'mid-30% of revenue' guidance) by pinning the apples-to-apples number before using it as a discipline threshold."

Context: open follow-up #2 from [[2026-07-21-memory-maker-fab-financing-supply-side]], which carried both figures side by side and filed the gap as "different denominators/definitions." That diagnosis was wrong, and the real explanation is worse for the intended use: one figure is an unreconciled forward target and the other is an unsourced forward estimate, and neither is a measured ratio.

What we already know (from the vault)

What the web says

Convergences and contradictions

Synthesis for RDCO

The defensible apples-to-apples series, named definition first. Definition: cash outflow for acquisitions of property, plant and equipment, from the consolidated statement of cash flows, divided by consolidated revenue for the same period, both in reported KRW, IFRS as issued by the IASB, fiscal year ending December 31, Solidigm consolidated, no FX conversion, no segment carve-out. This is the company's own footnoted definition of "capital expenditures," so it is the one that will stay reproducible against future 20-Fs.

Period Capex (KRW m) Revenue (KRW m) Capex / revenue Capex / D&A Capex / OCF Tier
FY2021 12,486,635 42,997,792 29.04% 1.21x 63.1% B
FY2022 19,010,261 44,621,568 42.60% 1.37x 128.6% B
FY2023 8,325,138 32,765,719 25.41% 0.63x 194.6% A
FY2024 15,945,534 66,192,960 24.09% 1.27x 53.5% A
FY2025 27,518,924 97,146,675 28.33% 1.98x 51.6% A
TTM to Jun-2026 35,232,021 189,170,615 18.62% 2.37x 27.8% A/B mix
H1-2026 18,328,836 131,895,000 13.90% 2.36x 20.0% A/B mix

Arithmetic, checkable: 27,518,924 / 97,146,675 = 0.2833. 8,325,138 / 32,765,719 = 0.2541. (7,657,403 + 10,671,433) / (52,576,300 + 79,318,700) = 18,328,836 / 131,895,000 = 0.1390. Three-year window on summed flows: (8,325,138 + 15,945,534 + 27,518,924) / (32,765,719 + 66,192,960 + 97,146,675) = 51,789,596 / 196,105,354 = 0.2641. D&A denominators use the cash-flow-note depreciation + amortisation split (FY2025 13,099,311 + 830,819 = 13,930,130); note that the filing carries three non-tying D&A figures for FY2023 — 13,619,161 in the expenses-by-nature note, 13,282,991 as the cash-flow-note split, and a third in the aggregator series — and does not reconcile them, so capex/D&A should carry a stated numerator source. Confidence in the ratios given the inputs: high. Confidence in the FY2021-2022 and Q2-2026 inputs: moderate (Tier B).

Mapping each conflicting figure to a definition that reproduces it — the mapping is the answer, and one of the two does not map. "Mid-30%" reproduces only as a plan ratio: the later-guided FY2026 capex of ~KRW 48T is 35.0% of KRW 137.1T (48.0 / 0.35 = 137.1), and KRW 137T is ~1.41x FY2025 revenue — a plausible January-2026 forecast for FY2026. A second candidate: Q4-2025 standalone capex intensity was 36.16% (11,870,574 / 32,826,700), the most recently printed quarter when the policy was restated. Both land in the mid-30s; neither can be confirmed, because the filing refuses to define the ratio and the Reuters body is paywalled. "~11%" reproduces only as a forward FY2026E ratio, and only under assumptions no one disclosed: 48.0 / 0.11 implies KRW 436T of FY2026 revenue, which is 4.5x FY2025 and above even Q2-2026 annualised (79.3187 x 4 = KRW 317.3T); on a lower numerator it comes easier — 35.2 / 317.3 = 11.1%, or 40.0 / 363.0 = 11.0%. Verdict on the Futurum figure: not falsified, but not reproducible, and not a measurement. It is 2-3 points below the nearest actually-reported analogue (H1-2026 at 13.90%, Q2-2026 standalone at 13.45%), so it is directionally honest; it was nonetheless published before any FY2026 capex number existed in public disclosure, against an undisclosed revenue estimate, with no source cited for it or for the three vendors it is compared against. It should be carried in the vault as "Futurum estimate, 2026-07-10, inputs undisclosed," never as a figure, and it should not anchor a threshold.

Why capex/revenue has to be demoted, independent of which number is right. From FY2023 to the trailing twelve months, SK Hynix's absolute capex rose 4.23x (8.33T to 35.23T) while capex intensity fell from 25.41% to 18.62%. The metric moved the wrong way through the largest capex escalation in the company's history, because revenue moved 5.77x against capex's 4.23x. In a price-driven memory upcycle capex/revenue is not merely noisy, it is perversely signed: it falls fastest exactly when spending accelerates hardest, because ASP rather than volume drives the denominator. At quarterly frequency it is worse — 36.16% in Q4-2025 to 13.45% in Q2-2026, a 22-point swing in two quarters with no behavioural change that could justify it. A tripwire keyed to a capex-intensity threshold would have stayed green through 2026 and would stay green through a KRW 48T year. Report capex intensity as a ratio with its vintage attached; do not make it the threshold, and never compare two vintages of it.

The replacement, recommended explicitly, with its cost stated. Primary: absolute capex indexed to the prior-cycle peak, per maker. SK Hynix's prior peak was FY2022 at KRW 19.01T; FY2025 is 1.45x it, TTM 1.85x, and the guided FY2026 2.52x (48.0 / 19.01). That series says plainly what intensity conceals — this is the most aggressive absolute capex posture in company history. Companion, for normalisation: capex / depreciation and amortisation, 1.21x / 1.37x / 0.63x / 1.27x / 1.98x for FY2021-2025 and 2.37x at TTM, with a clean economic reading at 1.0x (replacing the asset base exactly); FY2023's 0.63x correctly marks the capex-cut phase that intensity's 25.41% completely missed. The reason to prefer it is denominator stability, not smoothness: over FY2021 to TTM, D&A moved 1.45x while revenue moved 5.77x, so movement in capex/D&A is attributable to the numerator — the behaviour being measured. Third, as a forward observable rather than a backward one, and the single best find in this brief: Note 29 contracted-but-unrecognised capital commitments, disclosed KRW 8,837,748m at 2024-12-31, 6,667,863m at 2025-12-31, and 30,921,877m at 2026-03-31 — a 4.64x jump, +KRW 24.3T, in one quarter, which is larger than any full year of cash capex before FY2024 and which leads cash capex by construction, since it is signed purchase orders not yet in PP&E. That is the discipline observable the intensity ratio was standing in for. The cost of switching, paid knowingly: (1) cross-vendor comparability on one scale is lost — the "SK Hynix 11% vs Micron 21% vs Samsung 25-30% vs CXMT 77%" line cannot be restated in capex/D&A without building a D&A series for all four, and for CXMT and Samsung's semiconductor arm that data may not exist; (2) capex/D&A levels are sensitive to depreciation policy and asset vintage, so cross-issuer levels are no more comparable than DIO levels were — same constraint, same resolution, each issuer against its own history; (3) prior-peak indexing requires a declared cycle-peak date per vendor, which is a judgement call where hindsight can leak in; (4) the commitments line needs a verified Micron and Samsung analogue before it can be a three-maker gauge. Net: these are per-issuer time series, not a cross-vendor leaderboard — a downgrade in what we can say and an upgrade in what we can trust.

One line for the risk register, separate from the measurement point. The guided KRW 48T requires H2-2026 capex of KRW 29.67T, up 62% on H1's KRW 18.33T — a steep back-load, and the KRW 30.9T of contracted commitments at 2026-03-31 is consistent with it landing. If it does, FY2026 capex/D&A sits near 2.9x (48.0 / ~16.5T, annualising Q2-2026 D&A of 4.03T) and absolute capex at 2.52x the prior-cycle peak, with "CapEx Discipline" still in the press-release headline and intensity printing in the low teens. That conjunction — record absolute spend, a falling intensity ratio, and discipline language — is the pattern worth watching, and the metric that was about to become the threshold is the one that would have hidden it. Separately: Futurum's Micron figure ("about 21%", year unstated) was not reproduced here and should not be relied on.

Why this is in the vault

It retires a specific number from the memory-cycle tripwire set. The Phase-2→Phase-3 indicator work in [[2026-06-28-chip-memory-cycle-phase2-phase3-indicators]] and the per-maker discipline gauge proposed in [[2026-07-21-memory-maker-fab-financing-supply-side]] were both heading toward a capex-intensity threshold, and this brief shows that threshold would have stayed green through a 4.2x capex escalation. It also specifies three concrete fields for the memory-vendor-capex/ anchor schema drafted in [[2026-09-06-vendor-capex-composition-memory-anchor]] — capex_vs_prior_cycle_peak, capex_over_da, contracted_capital_commitments — confirms that brief's unverified IAS 16.73(e) expectation (additions by asset class are disclosed, and they run 2.0-2.7T above cash capex), and downgrades the Futurum ~11% from fact to attributed estimate wherever the vault carries it.

Open follow-ups

Related

Sources