06-reference/research

dram 2023 trough cash cost floor cxmt position

2026-11-04·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
memory-cycledramcxmtcost-per-bitcapital-cycle

The 2023 DRAM trough stopped at the high-cost producer's firm-level cash breakeven, not at cash cost per bit, and that is what sizes the CXMT modifier

The question

Verbatim: "What was the big three's (Samsung, SK hynix, Micron) DRAM cash-cost floor at the 2023 trough, and would CXMT sit above or below it at a 10% vs a 30% cost-per-bit gap?"

Context: [[2026-09-12-cxmt-ddr5-cost-per-bit-yield-parity-tripwire]] recommends moving ChangXin Memory Technologies' (CXMT's) cost leg out of the Phase-2 tripwire AND-condition and re-filing it as a Phase-3 severity modifier. The 2023-trough cash-cost floor for dynamic random-access memory (DRAM) is the number that sizes that modifier. This brief sizes a framework parameter. It is not a position recommendation and contains no price target or buy/sell call.

Two definitions, kept separate throughout, because conflating them is the main way this question goes wrong.

What we already know (from the vault)

What the web says

Convergences and contradictions

Synthesis for RDCO

The floor, with its derivation shown. Cash cost per bit is not disclosed by anyone, so it has to be backed out. The one derivation that avoids needing bit volumes is: cash cost per bit = realised ASP x (cash COGS / revenue). Micron's cash COGS is its tagged cost of goods sold less the tagged inventory write-down less the cost-of-goods-sold share of tagged D&A. That last term is an assumption, taken here as 85-95%, because Micron does not split D&A between cost of goods sold and operating expense. On that assumption, cash COGS runs 44.0-49.2% of revenue in FQ2'23, 57.6-62.8% in FQ3'23, 65.0-69.8% in FQ4'23 and 62.2-66.3% in FQ1'24. FQ2'23 is biased low because removing a $1.43bn write-down removes a real period charge, even a non-cash one, from a quarter whose output it related to. FQ3'23 and FQ4'23 are biased the other way, because they sold previously written-down inventory at an artificially low carrying value. The least-contaminated single read is FQ4'23, which took no write-down and is the quarter nearest the Q3-2023 price bottom: 65-70% of revenue, blended DRAM and NAND.

Converting that to DRAM only, and then to dollars per gigabit. Micron's DRAM weight is derived by pairing TrendForce's Micron DRAM revenue ($2.95bn in calendar 2Q23, $3.075bn in 3Q23) with Micron's hard-disclosed quarterly totals over the nearest fiscal quarters, giving roughly 77-79%; the periods are misaligned by about a month, so call it 77%. Solving the blend b = w·x + (1-w)·n for the DRAM-only ratio x, with b = 65-70%, w = 0.77, and a NAND cash-COGS ratio n assumed in the range 80-100% of revenue (NAND was the worse business in 2023, and that assumption is this derivation's softest input), gives x = 55-67% of realised DRAM price. Applying that to the chained Q3-2023 DDR4 8Gb low of $1.42-1.58, or $0.177-0.198/Gb, gives Micron's DRAM cash cost of goods sold at the 2023 trough of roughly $0.097-0.132 per Gb, or $0.78-1.06 per GB. Two caveats sit on that number. It is the high-cost producer's floor: Samsung's and SK hynix's are lower, and the TrendForce operating-margin spread of 31-34 points cannot be decomposed into cash cost versus depreciation intensity versus product mix from public data, so this brief does not put a number on theirs. And applying a Micron-wide revenue ratio to a DDR4 commodity contract price biases the floor down, because Micron's realised DRAM ASP was higher than DDR4 8Gb (TrendForce's 2Q23 release notes DDR5 held "a significant proportion" of shipments at a relatively stable ASP).

Where CXMT sits, on the band rather than a midpoint. On the literal question CXMT sits above the floor at both gaps by construction; the magnitudes are small in absolute terms, roughly $0.010-0.013/Gb at a 10% gap and $0.029-0.040/Gb at a 30% gap. The decision-relevant comparison is CXMT's cash COGS against the price the 2023 trough actually cleared at. Carrying the full band through: at a 10% gap CXMT's cash COGS is $0.107-0.146/Gb against a $0.177-0.198/Gb trough price, a cash-COGS margin of 18-46%. At a 30% gap it is $0.127-0.172/Gb, a margin of 3-36%. Because required-price-fall-to-breakeven and cash margin are the same quantity, those are also the further declines below the 2023 low that would put CXMT under water on a per-bit basis: 18-46% at a 10% gap, 3-36% at a 30% gap. So the per-bit answer is above the floor and below the trough price across the whole band in both scenarios, but the 30% case comes within about 3 points of breakeven at the pessimistic end while the 10% case keeps at least 18 points everywhere. That is where the 10-versus-30 distinction actually lives, and it is a statement about the band's tail, not its middle.

The firm-level answer, which is the one that predicts behaviour. Micron cut wafer starts twice, and the first cut of about 20% was announced at FQ1'23 when its gross margin was a positive 21.9%, long before any cash test. By FQ3'23 it was at operating-cash breakeven ($24M) and, with hard-disclosed fiscal-2023 capital expenditure of $7,676M against $1,559M of operating cash flow, deeply free-cash-flow negative. So the 2023 price floor was behavioural: it was set by discretionary supply cuts taken at negative gross margin and firm-level cash breakeven, roughly 30 points of revenue above cash cost per bit. Carrying Micron's own overhead load of 23-26% of revenue across to CXMT as a scenario assumption, CXMT's firm-level cash cost at 2023 trough prices would be 87-123% of that price at a 30% gap (cash-negative over most of the band, positive only at its optimistic end) and 78-108% at a 10% gap (cash-positive over most of the band, negative only at its pessimistic end). That is scenario arithmetic on an assumption CXMT's public filings cannot yet support, since its only reported quarters are boom quarters at 79-88% gross margin ([[2026-09-12-cxmt-ddr5-cost-per-bit-yield-parity-tripwire]], third-party, TrendForce), and it is labelled as such.

How to re-file the cost leg, concretely. Three changes. First, the modifier should bite in both scenarios, because at neither gap does a 2023-style trough reliably force CXMT off a per-bit cash basis. Second, it should bite harder at 10% than at 30%, but the difference is in the tail rather than the central case, so the ~10% threshold should be retired as a binary trigger: the number was proposed without derivation, and the derivation now available shows no step change there. Third, and most useful: key the modifier to CXMT's bit share at the time of rollover multiplied by (1 minus the probability CXMT cuts), and let the cost gap enter only as a prior on that probability. Those priors are unanchored Ray judgment, not derived: roughly 0.1 at a 10% gap, roughly 0.25-0.35 at a 30% gap, on the reasoning that a state-funded producer at 77% capex intensity cuts only when overhead-inclusive cash losses become politically expensive. That keys the modifier to two things the vault can already observe, the bit share from [[2026-09-08-cxmt-ymtc-bit-share-phase3-threshold]] and a cut-or-no-cut event now that CXMT is listed, rather than to a single-sourced cost estimate not refreshed since 1Q26. The cost gap still matters, but for a Phase-2 question rather than a Phase-3 one: at either gap CXMT can take share on a positive per-bit cash margin at the big three's trough prices, which is a statement about share-gain velocity, not trough depth.

One precedent worth recovering before anyone leans on this. 2023 is a trough that stopped at a firm's cash breakeven. The cycle that plainly went below cash cost on a per-bit basis is 2008-09, and what it produced was exits rather than a price floor (Qimonda's insolvency, Elpida's path to it). That is background knowledge in this brief, uncited, and item four of the follow-ups below turns it into a second and much harsher anchor for the modifier.

Why this is in the vault

It supplies the missing parameter for the specific change [[2026-09-12-cxmt-ddr5-cost-per-bit-yield-parity-tripwire]] proposed to tripwire row 5a in [[2026-07-07-dram-hbm-phase2-phase3-early-signals]], and it fills the 2023 hole in the Micron gross-margin and cash-cost series that [[2026-09-07-memory-supplier-inventory-days-dio-anchor]] left open, with primary filing data the anchor files can ingest directly.

Open follow-ups

Related

Sources

Vault

Hard disclosure

Analyst estimate

Paywalled or unreachable, flagged and not retried