06-reference/research

vendor capex composition memory anchor

2026-09-06·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
memory-cyclecapital-cycleedgarinvesting-anchorssupply-discipline

Memory-vendor capex composition is not disclosed by purpose — but it IS disclosed by asset class, and that proxy just moved

The question

Verbatim from the backlog row: "Add vendor-side capex composition tracking (greenfield-wafer vs node-migration vs HBM-packaging split) to the hyperscaler-capex EDGAR anchor — dollar level alone misses the disciplined-cycle signal."

This row is the direct execution of open follow-up #2 from [[2026-06-28-chip-memory-cycle-phase2-phase3-indicators]]. The researchable question underneath: can Micron / SK Hynix / Samsung capex composition actually be sourced from public disclosure, at what granularity and lag, and what does the most recent available split say about supply discipline? Phase numbering below 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 — the two are distinct, per the warning in [[2026-08-01-hyperscaler-capex-merchant-vs-inhouse-silicon-decomposition]].

What we already know (from the vault)

What the web says

Convergences and contradictions

Synthesis for RDCO

The row's premise is right and its location is wrong, on two counts. First, anchors/hyperscaler-capex/ is an empty directory behind a v0 scaffold that has never run — bolting a feature onto it is building on sand, and it should not gate this work. Second and more important, the row conflates sides of the equation. Hyperscaler capex (GOOGL/AMZN/MSFT/META) is the demand anchor; memory-vendor capex composition is the supply anchor. Different companies, different filing systems, different signal polarity: hyperscaler capex falling is bearish for the thesis, vendor greenfield capex rising is bearish for the thesis. Merging them into one anchor would produce a series where two components move opposite directions for opposite reasons. The correct build is a new sibling anchor, anchors/memory-vendor-capex/, reusing the EDGAR plumbing that /investing:edgar-watch was specified for but has not yet implemented. Note the dependency honestly: this anchor needs the EDGAR client that the hyperscaler anchor also needs, so the first implementation session should build the shared client once and file both.

The purpose split the row asks for cannot be sourced, but a better-behaved proxy can, and it is machine-readable. Vendors will never publish "X dollars to greenfield wafers, Y to 1-gamma migration, Z to TSV." What they must publish is PP&E by asset class — and for a fab operator those classes map onto the question almost cleanly. Land, Buildings, and Construction-in-Progress are shell: you only spend there to house wafer capacity you do not yet have. Equipment is tools: node migration and HBM/TSV packaging are overwhelmingly equipment purchases dropped into cleanrooms that already exist. So shell_share = Δ(Land + Buildings + CIP) / Δ(total gross PP&E) is a defensible greenfield-intensity gauge, and equipment_not_placed_into_service is the separate, sharper, nearer-dated gauge of capacity about to switch on. Both are disclosed by Micron quarterly. State the proxy's two real weaknesses plainly: Micron's own footnote says CIP "primarily includes building-related construction and tool installation," so CIP is contaminated with equipment; and gross-PP&E deltas are net of retirements and include acquisitions (Tongluo, $1.8B), so they are not capex. Report the metric as a ratio band with the acquisition stripped, never as a dollar decomposition — the same discipline [[2026-08-01-hyperscaler-capex-merchant-vs-inhouse-silicon-decomposition]] imposed on the merchant-silicon stack.

Concretely, the anchor to build. Directory ~/rdco-vault/01-projects/investing/anchors/memory-vendor-capex/, one CSV row per vendor-quarter with columns: vendor, period_end, filing_form, filed_date, source_url, capex_cash, ppe_gross_total, ppe_land, ppe_buildings, ppe_equipment, ppe_equip_not_in_service, ppe_cip, ppe_software, acquisitions_in_ppe, shell_share_pct, shell_share_ex_acq_pct, equip_not_in_service_yoy_pct, guided_bit_supply_growth_pct, capex_per_bit_index, tier. Sourcing per vendor: Micron — 10-Q and 10-K, the PP&E note, which lives in the Financial Report R-files (FilingSummary.xml → the report named Property, Plant, and Equipment (Details)), not in data.sec.gov/api/xbrl/companyfacts, because that API drops dimensional facts and Micron has not tagged ConstructionInProgressGross undimensioned since ~2020 (verified this session: HTTP 404 on the companyconcept endpoint, while the R-file carries the full class table). This is the single most important implementation gotcha and it will silently produce nulls if missed. SK Hynix — CIK 0002120882, 6-K for quarterly results now, 20-F for FY2026 due ~April 2027; because it reports under IFRS, IAS 16.73(e) requires a full PP&E movement schedule showing additions by class rather than only period-end balances, which is strictly better data than Micron's — flag this as a build-time verification against the first 20-F, not as a verified fact. Samsung — not an SEC periodic reporter; requires Korea's DART (OpenDART API) for the quarterly business report, or the English DS-division IR deck, and should be tiered lower and populated last. Cadence: quarterly, run at +5 weeks after each vendor's quarter end for the 10-Q/6-K, with an annual re-key at the 10-K/20-F. Realistic lag is 4-8 weeks for the number, 8-13 weeks for the audited class detail.

Thresholds, and where the signal reads today. Anchor the trigger set on three fields, none of which fires alone. (1) shell_share above 45% on a TTM basis for two consecutive quarters, acquisitions stripped — baseline is Micron's FY2025 31.1%; nine-month FY2026 is 45.7% raw, 38.4-51.8% ex-Tongluo, so this is at the threshold and ambiguous, not through it. (2) equip_not_in_service up >50% YoY at any vendor — tools bought and crated are capacity roughly two-to-three quarters from producing bits; currently +3.7%, clearly not triggered, and this is the field that keeps the near-term read disciplined. (3) capex_per_bit_index — capex divided by management-guided industry bit-supply growth — falling for two quarters means dollars are buying wafers again; this is the metric that most directly operationalises the vault's "capex is rising but bit supply isn't" observation, and it needs a backfill before it has a usable baseline. Composite rule, consistent with the existing Markov spec: raise the Phase-3 prior on (1) alone; declare a composition trigger only on (1) AND (2), or on (3) falling AND (2). Applying that today: Phase 2 holds, and the composition indicator has produced its first genuine flicker — the earliest such flicker the vault has recorded on this indicator, and it is Micron-only, proxy-based, and one quarter old. It should raise the Phase-3 prior modestly and trigger nothing. For a position/capital-cycle horizon the actionable content is not a trade; it is that the indicator the June 28 brief called the disciplined-cycle-specific signal has finally become measurable from primary sources, and it has stopped reading flat.

Why this is in the vault

This closes open follow-up #2 of [[2026-06-28-chip-memory-cycle-phase2-phase3-indicators]] by answering the feasibility question that follow-up left implicit, and it converts indicator #2 of the memory-cycle Phase 2→3 trigger set from a qualitative TrendForce-dependent read into a specified, primary-sourced, buildable anchor with named XBRL/R-file access paths, a field list, a cadence, and numeric thresholds — so the next session implements rather than re-researches. It also corrects the destination: this is a new memory-vendor-capex supply-side anchor, not a field on the demand-side hyperscaler-capex anchor, which is still an empty directory behind an unimplemented v0 scaffold.

Open follow-ups

Related

Sources

Vault

Primary (Tier A — retrieved directly from SEC EDGAR, 2026-09-06)

Web

Unreachable — flagged, not retried