06-reference/research

memory fab subsidy share 2027 28 wave

2026-10-06·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
memory-cyclecapital-cyclechips-actindustrial-policydram

Sizing the subsidy-insulated tail in the 2027-28 memory capacity wave

The question

"Quantify the subsidy share of the 2027-28 wave: what fraction of Micron Idaho/NY, Samsung Taylor, and the Korean clusters is CHIPS/KDB-funded vs maker-funded, since that fraction is the size of the 'capacity insulated from discipline' tail?"

This is open follow-up #2 from [[2026-07-21-memory-maker-fab-financing-supply-side]], which rejected the "external capital marks the end of oligopoly discipline" hypothesis but found a third term the FCF-vs-market binary was blind to: state subsidy. The job here is to size that tail, not re-argue it.

What we already know (from the vault)

What the web says

The question's three named US/Korean anchors mostly fall out of the window on inspection.

The US subsidy stack is far larger than the grant everyone counts, because the tax credit is bigger and nobody counts it.

The money has started moving, but no project-level disbursement figure exists.

The Korean headline package decomposes into almost nothing that functions as a subsidy to the memory makers.

Convergences and contradictions

Synthesis for RDCO

The honest answer is that the question's own framing overstates the tail, and the tail that is real sits somewhere the question didn't name. Three corrections have to land before any ratio means anything. Samsung Taylor is a logic foundry and contributes zero memory bits. Micron Clay NY first produces in ~2030. And Micron ID1, the one US memory fab actually inside the window, has first wafer in 2H CY2027 with Micron itself guiding no meaningful output until 2028. So the CHIPS-funded memory capacity landing as bits inside 2027-28 is a sliver — plausibly 15-50k wpm of nameplate running at partial ramp, against global DRAM additions I bottom-up at roughly 435-600k wpm across Korea, the US, Japan, Taiwan and China. That is ~2-5% of 2027-28 incremental wafer additions, with the KDB component at approximately zero. The mechanism the parent brief named as the tail is, inside the window the parent brief cared about, immaterial.

Two separate ratios were being conflated in the question, and they answer differently. The subsidy share of the capex behind in-window capacity is large: for Micron's US fabs I put the federal share at roughly 30-45% of qualified project cost, and the striking part is that this is dominated by the 48D investment tax credit at 35%, not by the $6.4B of CHIPS grants that gets all the attention. (Derivation, stated so it can be attacked: 35% statutory rate applied to a qualified base of 75-95% of total project cost gives 26-33 points, plus a few points of Idaho-weighted grant allocation. The low end assumes a narrow qualified base and thin grant allocation; the high end assumes buildings and tools almost entirely qualify and Idaho carries a disproportionate share of the $6.1B.) For Korea the same figure is roughly 15-28% — a 20% K-Chips facility credit on a 70-90% qualified base, plus an infrastructure-absorption leg of 2-10 points that is the single largest unquantified number in this brief, plus an interest subsidy of effectively zero. But the share of in-window capacity that is subsidy-dependent is a different question, and on that test Korea does not qualify: a 20% credit does not turn a negative-NPV fab positive, SK Hynix cut debt while raising capex 70% and funded Yongin from cash flow and a $26.5B equity raise, and Samsung demonstrably did pause P5 in the downturn and only restarted it in 2026. That is direct evidence that Korean state support did not prevent discipline from operating.

The genuinely insulated leg at scale is Chinese, not CHIPS and not KDB. CXMT at ~320k wpm heading to ~420k by 2027 and YMTC at ~200k heading to ~300k NAND by early 2027 are each adding ~100k wpm inside the window on something the vault already measures at ~77% state capex intensity. Bottom line, stated as a range: 20-35% of 2027-28 incremental memory wafer capacity is insulated from both cash-flow and market discipline, and 15-30 of those points are Chinese state capital. The low end assumes Chinese ramps at the low end of reported targets and global additions at the high end; the high end assumes CXMT reaches 420k and YMTC 300k while global additions come in at the low end of my bottom-up. In bit terms the Chinese share is probably higher still, because CXMT and YMTC are making commodity DDR5/LPDDR and NAND while the Korean additions are HBM-weighted at ~3x the wafer per bit — so the state-funded leg converts to commodity bits far more efficiently than the privately-funded leg. That is directional, not computed; the year-by-year HBM wafer-to-bit split is not available free, and the one circulating figure set failed verification.

For the capital-cycle thesis the practical implication is a timing one, and it cuts in a usefully unexpected direction. Western subsidy is not cushioning the 2027-28 window — it is cushioning 2029-2031, when Clay NY, Yongin Y2 (₩35.2T, cleanroom June 2029), SK Hynix M17 (₩19.1T, cleanroom December 2028) and Samsung P5/P6 land. The 48D step-up to 35%, with its elective-payment cash treatment and its hard 2026-12-31 construction-start gate, is a strong incentive to break ground now on capacity that produces at the end of the decade. So private discipline should still be the operative force on the Korean makers through the 2027-28 down-leg, which keeps the existing physical-anchor tripwire set in [[2026-09-25-hbm-dram-phase-2-3-transition-q3-2026]] valid for dating the turn; the subsidy-insulation risk should be re-sited as a depth-of-the-down-leg and shape-of-the-2030-recovery variable rather than a Phase-3 trigger. This is the same correction [[2026-09-26-financing-led-capital-cycle-phase3-precedent]] made about financing features: right feature, wrong socket.

Calibration. None of the ratios above deserve more than one significant figure. The denominator is paywalled and its two free fragments disagree by 17%; US per-fab capacity is undisclosed; the Korean infrastructure leg has no won figure attached to Yongin; project-level CHIPS disbursement is not published by anyone. The ratios are built from announced ceilings, statutory rates and a bottom-up capacity stack assembled from secondary trade press. The one thing here that is primary-sourced and high-confidence is the mechanism: the 35% ITC is the dominant US subsidy and it was entirely absent from the vault's prior treatment.

Comparison table

Capex in USD unless marked; ₩ converted at ~1,420 KRW/USD (back-solved from sources' own conversions, ±4%). "Govt share" is of announced program capex, not of in-window spend. Confidence applies to the row's weakest load-bearing figure.

Project Total announced capex Announced / awarded government share Maker-funded 2027-28-allocated capacity Tier / conf
Micron Boise ID1 (DRAM) Not disclosed per fab Inside the $6.1B definitive CHIPS award (2024-12-09, amended 2025-06-11) covering ID1+ID2+2 NY fabs — no per-fab split published; plus 35% 48D ITC on qualified capex Remainder; Micron in $24.4B net cash, no raise disclosed, capex quoted net of anticipated incentives First wafer 2H CY2027, "no meaningful output until 2028" → partial ramp only; nameplate undisclosed PRIMARY (award, timing) / high; capacity not disclosed
Micron Boise ID2 (DRAM) Not disclosed Added to the same $6.1B ceiling 2025-06-11 (reallocation, ceiling unchanged) + 48D Remainder Guided late CY2028 → ~0 effective bits in window PRIMARY / high
Micron Clay NY (DRAM, 4 fabs) up to $100B over 20+ yrs; first phase $20B by ~2030 Inside the same $6.1B + up to $5.5B NY state, explicitly NON-BINDING term sheet over 20+ yrs + 48D Remainder First production ~late 2030 → ZERO in window 10-K PRIMARY for program shape; $100B + 2030 slip SECONDARY / med-high
Samsung Taylor TX "over $37B" Central Texas PMT ~$6.4B (Apr 2024) cut to ~$4.745B final (~2024-12-20) Remainder LOGIC FOUNDRY — zero memory bits. Excluded. Customer production 2027 SECONDARY (commerce.gov 403) / med
Samsung Pyeongtaek P4 Ph2/Ph4 (DRAM, brownfield) Ph4 construction contract ₩4.2T (~$3.0B) 20% K-Chips facility credit; no project-specific grant or loan identified Overwhelmingly maker; 2026 plan >₩110T (~$73B) Completion pulled to Apr 2027, tools Q3 2027; >200k wpm of 1c by YE2026 → in-window adds ~60-120k wpm SECONDARY / med
Samsung Pyeongtaek P5 "minimum ₩50T" (weak) 20% K-Chips credit Maker Mass production late 2028 → ~0 effective by end-2028 SECONDARY / low-med
SK Hynix Yongin Y1 Cluster ₩600T current plan (was ₩120T/4 fabs in 2019); ~₩120T per fab revised 20% K-Chips credit; share of ₩6.5T combined drawn policy-loan balance at ~100bp discount; infrastructure 50-100% publicly fundable by statute but no Yongin won figure Maker: debt fell to ₩18.59T while capex +70%; ~$44.8B net cash; $26.5B ADR raise Cleanroom Feb-Mar 2027, +60k wpm per cleanroom, +60k every 6 mo, 360k by H1 2030 → ~120-180k wpm effective by end-2028 (my derivation from the sourced cadence) SECONDARY / med
SK Hynix Yongin Y2 ₩35.2T (board-approved 2026-08-07) as above Maker Construction July 2027, cleanroom June 2029 → ZERO in window PRIMARY (SK newsroom) / high
SK Hynix M15X Cheongju >₩20T (~$14.6B) 20% credit; ₩500B KDB low-interest approval Dec 2025 (~$340M) Maker ~40k wpm 2026 → ~80-90k wpm 2027 → in-window adds ~40-50k wpm SECONDARY / med; sources disagree ~2x on the 2026 slope
SK Hynix M17 Cheongju ₩19.1T as above Maker Cleanroom Dec 2028 → ~0 in window PRIMARY (SK newsroom) / high
CXMT (China, DRAM) Not disclosed ~77% capex intensity on state capital (vault) ~23% ~320k wpm now → ~420k by 2027 → in-window adds ~100k wpm SECONDARY / low-med
YMTC (China, NAND) Not disclosed State-backed — ~200k wpm → ~300k by early 2027 → adds ~100k wpm SECONDARY / low-med
Context: Micron Hiroshima +25k wpm late 2027 (HBM); Nanya 5A ~60k; Winbond 15k included in the denominator SECONDARY / med

Bottom line. Insulated fraction of 2027-28 incremental memory wafer capacity: 20-35%, of which 15-30 points are Chinese state capital and only ~2-5 points are CHIPS-funded, with the KDB/Korean policy-loan contribution ≈0. Low end: Chinese ramps at the low end of reported targets, global additions at the high end of my 435-600k wpm bottom-up, Micron ID1 contributing 1-2 points. High end: CXMT to 420k and YMTC to 300k, global additions at 435k, ID1 at the top of its ramp. Subsidy share of capex behind in-window capacity is a different and larger number: ~30-45% for Micron's US fabs (ITC-dominated) and ~15-28% for the Korean clusters (credit plus an unpriced infrastructure leg). The band cannot honestly be narrowed without the paid capacity series and a won figure for Yongin infrastructure.

Why this is in the vault

This closes open follow-up #2 of [[2026-07-21-memory-maker-fab-financing-supply-side]] and amends the memory-cycle risk register: the "subsidy-insulated capacity" tail should be re-labelled as predominantly Chinese and re-dated to 2029-31 rather than 2027-28, which means the existing physical-anchor tripwire set in [[2026-09-25-hbm-dram-phase-2-3-transition-q3-2026]] stays valid for dating the Phase 2→3 turn and does not need a subsidy overlay. It also puts the 35% Section 48D credit into the vault for the first time — the largest single subsidy mechanism in the stack, previously absent from every memory-cycle brief.

Open follow-ups

Related

Sources

Vault:

Primary (US):

Primary (Korea):

Secondary (labelled as such throughout):

Explicitly NOT publicly available without a paid subscription, and therefore the reason the bottom-line band is wide: reconciled global installed DRAM/NAND capacity in kwspm (two free fragments disagree ~17%); per-fab nameplate kwspm for Micron ID1/ID2/Clay (Micron discloses cleanroom square footage and dollars only); year-by-year incremental DRAM additions in kwspm for 2026/2027/2028; regional share of 2027-28 additions; and HBM share of wafer input vs share of bits by year from a named publisher.