06-reference/research

micron cip contamination unwind

2026-10-14·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
memory-cyclecapital-cycleedgarinvesting-anchorsmicron

CIP contamination cannot be unwound, and it was never the binding constraint — government-incentive netting is a bigger contaminant, and the honest band is 29-57%

Not a recommendation. No buy, sell, short, or allocation is expressed or implied anywhere below. This is disclosure forensics against a threshold the vault pre-registered on 2026-09-06.

The question

Verbatim from the backlog row: "Can Micron's construction-in-progress contamination be unwound — does any disclosure (capitalized interest, assets-not-yet-in-service, segment detail) separate building construction from tool installation, or is the shell-share threshold permanently a band rather than a number?"

This is open follow-up #5 of [[2026-09-06-vendor-capex-composition-memory-anchor]], which proposed shell_share = Δ(Land + Buildings + CIP) / Δ(total gross PP&E) as the supply-discipline gauge and flagged CIP contamination as its main precision limit. The item was gated on the FY2026 10-K existing.

Gate status: PASSED. Micron Technology (CIK 0000723125) filed its FY2026 Form 10-K on 2026-10-09, accession 0000723125-26-000023, period ended 2026-09-03 (SEC submissions API, retrieved 2026-10-14 01:02 ET). It is the primary source for everything below that is not explicitly labelled otherwise.

What we already know (from the vault)

What the web says

Everything in this section is primary-filing evidence except the final bullet. "$M" throughout.

Convergences and contradictions

Synthesis for RDCO

The direct answer is a band, and the defensible bounds are wider than the vault has been carrying. Start with the raw full-year number: Δ(Land + Buildings + CIP) = 0 + 4,320 + 5,906 = 10,226 over Δ gross PP&E of 24,480, or 41.8%. Then apply the two disclosed contaminants in the directions the disclosure permits. Tongluo ($1.80B, a wafer-fab acquisition whose own shell/tool mix is undisclosed) moves it to 37.2-45.1%. Government-incentive netting ($6.43B, allocation undisclosed) moves it to 33.1-53.9%. Taking both at their extremes gives the honest outer band: 28.9% to 57.2%. The CIP shell/tool question — the one this brief was commissioned to answer — sits inside those bounds and does not widen them further, because CIP enters the numerator whole either way; resolving it would only reallocate between Buildings and Equipment within an already-wider interval. That is the finding the queue item did not anticipate: unwinding CIP, even perfectly, would not deliver a number. A pre-registered 45% threshold cannot be adjudicated against a 28-point band, and the FY2026 reading straddles it on every variant.

So the indicator should be re-specified rather than re-measured, and the shape of the fix is already visible in the data. Three changes, in order of value. First, gross up for government incentives explicitly and report shell_share as a labelled interval with the netting assumption named, never as a point. Micron discloses the in-year incentive offset annually in Note 15, so the gross-up is mechanical; it belongs in the anchor's field list as gov_incentive_ppe_offset_fy alongside the existing columns. Second, demote shell_share from a trigger to context and promote ENIS to the primary. ENIS is a clean, single-line, uncontaminated disclosure of tools bought and not yet producing; it is not netted against incentives, not blended with buildings, and not an acquisition artifact. It is also the nearer-dated signal — crated tools are two to three quarters from bits, where a shell is three to six years from them. This run is the proof: ENIS gave an unambiguous +51.1% while shell_share gave three different answers depending on which contaminant you corrected. Third, add U.S. long-lived assets as the geographic shell proxy. Clay, New York has zero tools installed and no wafer output until calendar 2030; Idaho ID2 has none until late 2028. U.S. long-lived assets rising 71% to $14,463M is as close to a pure construction signal as Micron's disclosure gets, and it is annual, audited, and already in Note 22. This mirrors the demotion [[2026-10-06-sk-hynix-capex-intensity-reconciliation]] performed on capex/revenue: when a ratio's definitional band swamps its threshold, replace the ratio, do not keep refining it.

What would have to change for a number to exist, stated precisely, so this is never re-researched. Exactly one of four things. (a) Micron voluntarily disaggregates footnote (2) into building-related construction and tool installation — no rule compels it and the competitive incentive runs against it. (b) Micron files a PP&E movement schedule by class; IAS 16.73(e) requires one and ASC 360 does not, which is why SK Hynix under IFRS will be structurally better data than Micron the moment its first 20-F lands (~April 2027) — the vault should stop treating Micron as the reference vendor for this metric. (c) The CHIPS direct funding agreements become public with their milestone schedules; Note 15 confirms construction and tool-installation milestones are contractually distinct and separately funded, so the split exists on paper, though the agreements are not EDGAR exhibits and would have to come through a Commerce Department release or FOIA. (d) Micron begins allocating assets by segment, explicitly refused in two places in this filing, and not plausible while the CODM does not review it. Until one of those happens, the honest ceiling is a band, and the band is 29-57%, not the 38-52% the vault has been carrying.

The read on the thesis, kept narrow. Applying the pre-registered composite rule to the FY2026 print: tripwire (2) fires (+51.1%, through the >50% line); tripwire (1) does not (41.8% raw, band straddling 45%, and decelerating to 35.3% in Q4). The rule requires (1) AND (2), so no composition trigger, and no phase change. But the texture has inverted from the September read, and that is worth recording: nine months ago the finding was "dollars into concrete, tools not stacking." The full year says concrete paused in Q4 while tools began stacking, against FY2027 capex guided at ~$11.5B in Q1 and ~$25B in the first half with the second half "higher" — implying well over $50B net of incentives, roughly double FY2026's $30.71B gross. Buildings-then-tools is the normal ordering for capacity that switches on in 2027-2028, and ID1 is guided to first wafer output mid-CY2027. The composition indicator has not triggered; it has started behaving like a fab that is about to turn on.

Why this is in the vault

This closes open follow-up #5 of [[2026-09-06-vendor-capex-composition-memory-anchor]] with a negative answer at primary source, and in doing so corrects that brief's diagnosis: CIP contamination is real but is not the binding precision limit on shell_share — undisclosed allocation of government-incentive PP&E offsets is, and it is larger. It therefore changes what gets built in anchors/memory-vendor-capex/: add a gov_incentive_ppe_offset_fy field and a gross-up interval, demote shell_share from trigger to context, promote equipment_not_placed_into_service to the primary supply-discipline gauge, and add U.S. long-lived assets as the geographic shell proxy. It also records that tripwire (2) crossed its pre-registered >50% line on the FY2026 print while tripwire (1) did not, which the composite rule resolves to no trigger.

Open follow-ups

Related

Sources

Vault

Primary (Tier A — retrieved directly from SEC EDGAR, 2026-10-14)

Web (Tier C — benchmark and negative-result context only; no load-bearing number depends on these)

Searched and not found (a finding, not a gap in effort)