06-reference/research

memory supplier inventory days dio anchor

2026-09-07·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
memory-cycleinvesting-anchorsdioinventory-daysedgarphase-markers

DIO is computable — and it is a different metric than the "weeks of supply" the tripwire scoreboard has been reading

The question

Verbatim from the backlog row: "Build a supplier-inventory-days (DIO) time series for Micron/SK Hynix/Samsung from quarterly earnings as a new structured investing anchor — the highest-value leading indicator for the memory-cycle Phase 2→3 turn."

This is open follow-up #1 of [[2026-06-28-chip-memory-cycle-phase2-phase3-indicators]], carried over unbuilt through [[2026-07-07-dram-hbm-phase2-phase3-early-signals]]. Scope here is research, not build: is DIO sourceable per issuer, what did it actually do at the last two turns, and what is the implementable spec. Phase numbering is the capital-cycle ladder (P1 tight-supply → P2 capacity-announce → P3 capacity-online/glut → P4 capex-cuts), not the thesis's own Phase 1/2 model.

What we already know (from the vault)

What the web says

Convergences and contradictions

Synthesis for RDCO

The row's premise survives, with one correction and one downgrade. The correction: DIO is not the same indicator as the weeks-of-supply reading currently sitting in tripwire row #2, and merging them would break the scoreboard. The downgrade: DIO is not cleanly the "earliest, highest-weight" leading indicator the June 28 brief assumed. Measured against Micron's own gross-margin peak — the cleanest available turn marker, and one computable from the same two facts — total DIO troughed two quarters before the 2018 peak (93.0 at FQ ending 2018-03-01, GM peak 61.0% at 2018-08-30) but only one quarter after the 2021 strict peak (93.1 trough at 2021-03-04, then a false-flat 93.6 in the peak quarter itself, with the decisive break to 105.4 arriving in Dec-2021). n=2. The honest characterisation is "turns within ±2 quarters of the margin peak, with real risk of a coincident-to-late read," not "leads by 1-2 quarters." Anything more precise is false precision on two observations.

The genuinely new finding is that the WIP/FG decomposition is where the lead lives, and Micron discloses it quarterly. In both prior cycles the earliest-turning of the three components was work-in-process days, and it turned in the same quarter each time: FQ ending 2018-05-31 (57.7 → 63.5 → 65.2, +13.0% over two quarters, two quarters ahead of the GM peak) and FQ ending 2021-06-03 (69.2 → 72.9 → 72.4 → 79.9, +15.5% over three quarters, one quarter ahead of the strict GM peak). Finished-goods days did the opposite — it troughed at or after the peak (20.3 in Nov-2018, one quarter late; 10.7 in Sep-2021, coincident) and is therefore a confirming, not leading, field. The mechanism is legible and is exactly the shape of a top: wafer starts keep rising while output is sold instantly, so WIP inflates while FG stays on the floor, and the glut arrives when those wafers finish. Today Micron reads FG days 8.7 — the lowest value in the entire series back to 2010, below the 2021 low of 10.7 and far below 2018's 20.3 — while WIP days have risen three quarters running, 92.9 → 96.0 → 97.9, +5.4% off the trailing floor. That is the right shape at roughly one-third the calibrated magnitude, and it is confounded: HBM's TSV-stack and hybrid-bonding flow has structurally longer cycle time, so mix-shift toward HBM raises WIP days independent of the cycle. Note also that the WIP-days baseline has stair-stepped across cycles (≈45 in 2015, ≈58 in 2018, ≈70 in 2021, ≈93 now) as process step-counts grew — absolute levels are worthless as thresholds; only slope off a trailing floor is usable.

The anchor to build. New directory ~/rdco-vault/01-projects/investing/anchors/memory-vendor-inventory/, sibling to the memory-vendor-capex/ anchor proposed in [[2026-09-06-vendor-capex-composition-memory-anchor]] — supply-side, one CSV row per vendor-quarter: vendor, fiscal_period, period_start, period_end, period_days, cal_quarter, cal_offset_days, currency, gaap_basis, form, filed_date, source_url, inventory_total, inventory_fg, inventory_wip, inventory_raw_residual, cogs_qtr, revenue_qtr, gross_margin_pct, dio_total, dio_fg, dio_wip, dio_total_pct_off_8q_min, dio_wip_pct_off_8q_min, dio_fg_pct_off_8q_min, tier. Sourcing. Micron — Tier A, fully automatable today from data.sec.gov/api/xbrl/companyconcept on four us-gaap tags plus revenue, with the FQ4-derivation and residual-raw-materials rules above; backfillable to FY2011 for the components. SK Hynix — Tier B until the 20-F (due ~April 2027) or a results 6-K is parsed; the balance sheet and income statement exist quarterly in KRW under K-IFRS via DART/IR, and because SK Hynix is essentially a pure memory company its consolidated DIO ≈ memory DIO, which makes it the cleanest of the three on contamination and the dirtiest on access. Samsung — Tier C, consolidated-only, contaminated by MX/Display/Harman; carry it as commentary until segment-inventory disclosure is verified, and never let a Samsung consolidated DIO drive a trigger. Cadence quarterly, run T+5 weeks after each vendor's period end. The calendar-offset field earns its place: Micron's quarters end roughly a month before the Korean calendar quarters and it reports ~3-4 weeks after close, so Micron's print lands four to five weeks ahead of SK Hynix's for the nominally-same period — Micron is the anchor's early-warning issuer, not merely one of three. Normalisation across currencies is a non-issue by construction: DIO is a ratio of two same-currency quantities and comes out in days, so no FX conversion ever enters the series. Contamination is the real normalisation problem, and it is unfixable at the disclosure floor: Micron's DIO blends DRAM and NAND, Samsung's blends the whole company. Report DIO per issuer, never as a three-issuer average.

Thresholds, and where the anchor reads today. All three are computed per-issuer against that issuer's own trailing-8-quarter minimum, never as absolute levels. (A) Lead — dio_wip ≥ +10% off trailing-8q min for 2 consecutive quarters. Calibration: 2018 hit +13.0% two quarters before the GM peak; 2021-22 hit +15.5% one quarter before. Today Micron is +5.4% and rising three quarters running — moving, roughly a third of the way, not triggered. (B) Confirm — dio_total ≥ +5% off trailing-8q min for 2 consecutive quarters. Calibration: fires in the 2018 GM-peak quarter itself (+6.0%, +7.0%) and in Mar-2022 (+13.2%, +26.6%), one quarter ahead of the 2022 margin collapse. Today Micron is at its own trailing minimum, 0% — not triggered, and still making new lows. (C) Late confirm — dio_fg ≥ +25% off trailing-8q min for 2 consecutive quarters. Calibration: 2018-19 fires at FQ ending 2019-05-30 (+25.6%, +27.6%), three quarters after the peak; 2021-22 fires hard (+117.8%). Today Micron is at an all-series low — not triggered, maximum distance to travel. Composite, consistent with the Markov spec: (A) alone raises the Phase-3 prior; declare a DIO-anchor confirmation only on (A) AND (B). (C) is a post-hoc audit field, useful for scoring the anchor after the fact, not for acting.

What this means for position posture: nothing changes. Zero DIO triggers have fired at the one issuer measurable from primary sources, gross margin is at an all-time high and accelerating, and finished-goods days are at their lowest in the fifteen-year record. The one honest flicker is the WIP-days slope, and it is a third of calibrated magnitude with a legitimate HBM-mix explanation. Set against the third-party read that the metric actually deteriorating is customer-side inventory (2-4 → 7-9 weeks against an 8-week warning line), the useful strategic conclusion is uncomfortable for the anchor being proposed: the supplier balance sheet may be the wrong balance sheet. In a rationed market suppliers hold nothing by construction, and the glut assembles in customers' warehouses first — which is exactly the 15-31-week hoarding pattern that preceded prior peaks. Build the supplier DIO anchor because it is cheap, primary, backfillable to 2010, and it discriminates WIP from FG in a way no weeks-of-supply commentary does. But do not promote it to the top of the trigger stack on the strength of the June 28 assumption, and open a separate question on whether channel inventory is observable from OEM/distributor filings (Arrow, Avnet, and the module makers all publish inventory) rather than from paid market-research commentary.

Why this is in the vault

This closes open follow-up #1 of [[2026-06-28-chip-memory-cycle-phase2-phase3-indicators]] — carried unbuilt through [[2026-07-07-dram-hbm-phase2-phase3-early-signals]] — by proving the series is computable from primary SEC XBRL for Micron, producing the actual FY2011→FY2026 readings, and converting indicator #2 from a TrendForce-commentary row into a specified anchor with named tags, two verified parser gotchas, per-issuer tiering, and thresholds calibrated against the two prior turns. It also corrects a live defect in the tripwire scoreboard: row #2's threshold is in weeks-of-supply units and cannot be fed by DIO, so the anchor must sit beside that row rather than replace it — and it downgrades DIO from "earliest, highest-weight leading indicator" to "turns within ±2 quarters of the margin peak," which changes how much weight the Markov Layer-1 should put on it.

Open follow-ups

Related

Sources

Vault

Primary (Tier A — SEC XBRL companyconcept, CIK 0000723125, all retrieved 2026-09-07)

Tier B (aggregator — must be re-keyed to DART/6-K before entering the anchor)

Web

Not verified this session — flagged