06-reference/research

capacity announce asp lag 2021 2022 rekey

2026-11-03·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
investingmemory-cyclephase-historycapexdramnand

The cross-cycle DRAM lag is 11-16 months at n=2, but the 2021-22 rollover was demand-driven, so this is not n=2 on a supply lag

The question

Verbatim: "Re-key the capacity-announcement to ASP-rollover lag on the 2021-2022 memory cycle, where the vault holds monthly TrendForce data from 2021-02, to take the lag calibration from n=1 to n=2 and state it as a cross-cycle range."

Context: the lag between a capacity announcement and the subsequent ASP rollover is what the memory-cycle thesis uses to time the Phase 2 to Phase 3 transition. The 2018 re-key landed at n=1 and had to bracket to a quarter because the vault's ASP anchor holds no pre-2021 data. 2021-2022 is the one prior cycle where the anchor data is already on hand.

What we already know (from the vault)

What the web says

Convergences and contradictions

Synthesis for RDCO

The measurement, with every keying choice named. The 2021-22 DRAM rollover dates cleanly from the anchor: contract prices rose 1Q21 through 3Q21, the first forward decline call published 2021-08-10 (PC DRAM 0-5% for 4Q21), the first actual QoQ decline was 4Q21 at -3 to -8% overall ASP, and the vault's own phase-history independently marks 2021-12 and 2022-03. Keyed to Samsung Pyeongtaek P2, 2020-08-30, that is 11 months to the first published decline call, 13 months to the last up-quarter ending, and ~15 months to mid-decline-quarter: DRAM 11-15 months. NAND, keyed to the same event, is 13-15 months. Keying instead to SK Hynix M16's 2021-02-01 completion gives 6-10 months for the same rollover, and keying to Samsung's 2021-05-13 capex raise gives ~6 months while being the wrong product line entirely. Announcement selection moves the answer by about five months, which is as wide as the range itself. I keyed to Pyeongtaek P2 because it is the only event in the 2020-21 window structurally matched to the 2018 keying event: a Samsung Pyeongtaek line announcing mass production on an already-standing shell. That is a defensible choice and it is still a choice.

The cross-cycle range at n=2: DRAM 11-16 months, NAND 6-15 months. The DRAM union is narrow and the brackets overlap at 12-15 months. The NAND union does not overlap at all, so there is no NAND range worth quoting. Two observations are two observations. Do not read 11-16 as a distribution, do not take a midpoint, and do not let it set a date. What it can bear: nudging a prior, and sanity-checking a lag someone else proposes. What it cannot bear: triggering a position change, setting the shell_share threshold in [[2026-09-06-vendor-capex-composition-memory-anchor]], or being quoted as "the" lag. An honest band on an n=2 estimate where keying discretion is ±5 months is at least ±3 quarters.

The re-key does not get the thesis to a clean n=2 on supply lag, and that is the finding. The 2018 interval was already contaminated: the parent brief flagged a simultaneous iPhone-pricing and trade-war demand air pocket. The 2021-22 interval is worse - the proximate cause per every contemporaneous primary source is an inventory unwind and a post-COVID consumer-electronics demand normalization, against a supply surplus of 1.6pp that the forecasting analyst described as arising from conservative expansion. So the supply-only calibration stands at n=1 contaminated, not n=2, and arguably n=0 clean. The more parsimonious reading of the matching 11-16 month numbers is uncomfortable: memory demand wobbles every four to six quarters, so "roughly a year after a splashy capacity press release" will land on a demand wobble in almost any cycle regardless of whether the capacity mattered. The threefold difference in up-leg length across the two cycles, with no corresponding difference in lag, is the tell. The charitable counter-reading is that announcement-to-incremental-wafers really is ~12-15 months for tool-fill on a standing shell, and the demand wobble is what converts latent supply into a visible price decline - a supply-readiness plus demand-trigger composite. Both readings fit the data; neither supports using the lag as a clock.

What this does to the Phase 2 to Phase 3 call. Applying 11-16 months to the current window's three candidate keying events gives: the 2025-11-13 TrendForce capex print to Oct 2026 - Mar 2027; Micron's 10-Q period ended 2026-05-28 to Apr 2027 - Sep 2027; the Jul 2026 Korea commitment and SK Hynix's 2026-07-28 capex raise to Jun 2027 - Nov 2027. The union is Q4 2026 through Q4 2027, essentially unchanged from the parent brief - the re-key tightens the input bracket by one month and changes no date. The shell-first correction still applies and now looks larger, not smaller: both cycles I keyed were tool-fill on standing shells, whereas the current window is buildings-first with the first greenfield step dated H2 2027 and the mega-commitment's bits dated 2029-2030. So the raw range remains a floor. The actionable change is a demotion, not a date. Both measured rollovers were triggered by an inventory and demand event, and the live tripwire set already has supplier inventory days as its highest-value signal, currently at single-digit weeks against a ~5-6 week trigger. This brief says: keep weight on inventory days, strip the announcement-to-ASP lag back from a timing instrument to a prior, and retire the NAND-leads-DRAM ordering that the 2018 brief proposed as a tripwire - it failed its first out-of-sample test.

Why this is in the vault

This closes open follow-up #2 of [[2026-09-09-2018-capacity-announce-asp-rollover-lag]] ("repeat this re-keying on the 2021-2022 cycle... take the lag calibration from n=1 to n=2") and its follow-up #4 (the supply-versus-demand decomposition), and it returns a negative result on both: the DRAM number replicates while the mechanism does not, so the calibration that [[2026-09-06-vendor-capex-composition-memory-anchor]] wanted for its 45% shell_share threshold still does not exist. It also supplies the out-of-sample test that demotes the "NAND leads DRAM by ~3 quarters" rule the 2018 brief recommended adding to the Phase-3 tripwire set in [[2026-07-07-dram-hbm-phase2-phase3-early-signals]], and it documents that phase-history.csv carries no capacity-announcement row for the 2022 cycle, which /investing:label-historical-phases should either fill with Samsung P2 (2020-08-30) or leave empty deliberately.

Open follow-ups

Related

Sources

Vault

Web (fetched and quoted)

Web (cited from search results, body not fetched)

Not retrieved (flagged, not retried)