Micron’s latest results strengthen the demand case. The harder test is whether Samsung, SK Hynix and Micron can preserve higher earnings and ROE after scarcity pricing fades and today’s investment becomes tomorrow’s supply.
The Question Has Moved Beyond Peak Profits
In August, our Samsung–SK Hynix analysis reached a deliberately uncomfortable conclusion about the memory boom:
“Revenue exploded. Bits did not. Price did most of the work.”
Micron’s fiscal third-quarter results made the distinction visible. DRAM revenue rose 67% sequentially, while bit shipments increased only in the low-single-digit range. NAND revenue rose 99%, with bit shipments increasing in the mid-single digits. Prices explained most of the acceleration.
The next question is harder. A stronger AI demand base may change the industry even if extraordinary prices eventually retreat.
When memory prices normalize, can AI-driven bit-demand growth sustain a structurally higher earnings floor even after the capacity created by today’s shortage reaches the market?
For investors, that means asking whether falling ASPs can coexist with revenue, earnings and ROE troughs materially above those of previous cycles.
The answer is not contained in another record quarter. It requires following memory through price normalization, supply expansion and the conversion of revenue into returns on a growing capital base.
Our present judgment is Partially Supported — Still Unproven. The case for a stronger demand base has advanced. A permanently higher earnings and valuation floor has not yet been demonstrated.
Where the Original GMS Thesis Stands
This analysis extends the original Memory Series rather than replacing it with a Micron earnings story.
The series introduction, AI Is Reshaping the Memory Cycle: How Should Investors Value Korean Semiconductor Stocks?, framed the problem as a distinction between cyclical excess profits and structural earnings power. Its evidence was the coexistence of AI infrastructure demand and an investment response from producers. The unresolved question was how much profitability would survive the next downturn. Capacity, inventory and pricing were the warning signals; the valuation implication was that neither a low P/E nor a historical P/B band could settle the question alone. That framework remains intact.
Part I — AI Is Reshaping the Memory Cycle: Why South Korea Matters argued that HBM qualification, packaging and integration with accelerator roadmaps created a differentiated layer within a cyclical industry. Technology partnerships and successive HBM generations supported the thesis. Whether differentiation would sustain superior returns remained unresolved. Rival qualification and capacity expansion were the warnings. The latest product roadmaps strengthen the technology argument, while the eventual margin premium still requires testing.
Part II — The Memory Cycle Is Not Dead: Why P/E and P/B Send Opposite Signals connected valuation to sustainable ROE. Losses during the 2023 downturn illustrated why P/E can become unusable near a trough; soaring profits illustrated why it can look deceptively cheap near a peak. The unresolved issue was whether normalized ROE had moved higher. Falling prices, new supply and earnings revisions remained the warnings. Today’s gap between low forward P/E and high historical-book P/B makes that argument more relevant.
Part III — AI, China and the New Korean Semiconductor Order argued that AI differentiation and Chinese supply competition could operate simultaneously. HBM development supported the premium tier; domestic substitution and Chinese process progress challenged conventional memory. The unresolved question was how far Chinese suppliers could move beyond protected domestic demand. Yield, qualification and incremental conventional supply were the warning signals. CXMT’s latest announcement strengthens the supply-risk mechanism, without establishing global competitiveness.
Part IV — Samsung vs SK Hynix: Has AI Created a New Valuation Equilibrium? brought those strands together through Volume × Price × Product Mix. Micron’s price-led acceleration and the Korean companies’ extraordinary profits supported the diagnosis. The unresolved question was tomorrow’s bit supply. The monitoring sequence remained bit supply → ASP → pricing → earnings revisions → inventory → operating leverage. The sequence is a monitoring framework, not a claim that every variable must move in that exact order.
The current evidence strengthens AI’s importance to demand and customer relationships. It also strengthens the reason to watch supply. It does not resolve the original valuation question.
Micron Q4: Bits Are Contributing More, but Prices Still Dominate
Micron reported fiscal Q4 2026 revenue of $54.229 billion, up from $41.456 billion in Q3. GAAP gross margin reached 86.8%, and GAAP operating income was $43.751 billion. The quarter ended September 3; the results were released September 30.
https://www.sec.gov/Archives/edgar/data/723125/000072312526000018/a2026q4ex991-pressrelease.htm
The growth decomposition is more important than the headline.
These are company shipments and company-reported pricing changes. The qualitative ranges are not exact percentages and should not be reverse-engineered into precise bit estimates.
https://investors.micron.com/files/doc_financials/2026/q4/Q4-26-Earnings-Deck.pdf
The transition is real but limited. Bits contributed more, particularly in NAND. Prices still did most of the work.
Nor has price normalization started in these results. Micron’s ASPs continued rising. Independently, TrendForce’s September 30 outlook forecast further calendar-Q4 contract-price increases of 10–15% for conventional DRAM and 15–20% for NAND. Those are forecasts, with a different calendar and price basket from Micron’s fiscal-quarter results.
One detail also complicates the familiar assumption that more HBM must always mean higher margins. Micron’s Cloud Memory gross margin stayed at 83%: higher pricing was offset by higher HBM mix. Under severe conventional-memory scarcity, the technological premium and the immediate margin ranking can diverge.
That does not weaken HBM’s strategic importance. It shows why a durable product advantage must be distinguished from the economics of a particular shortage.
How Much Price Decline Can Bit Growth Absorb?
The central arithmetic is straightforward:
Revenue ≈ Bit Volume × ASP
The original Volume × Price × Product Mix framework still applies. When ASP is a reported blended realization, however, mix is already embedded in that ASP. Adding a separate mix multiplier would double-count part of the effect.
The matrix below holds the comparison period constant and calculates revenue change as (1 + bit growth) × (1 + ASP change) − 1. It is a GMS sensitivity analysis, not a company or industry forecast.
The full matrix is displayed in two panels for readability. Percentages are rounded to one decimal. Source: GMS calculations.
The revenue-flat ASP thresholds are approximately −9.1% at 10% bit growth, −13.0% at 15%, −16.7% at 20%, −18.0% at 22%, −20.0% at 25%, and −23.1% at 30%.
The result sets a useful boundary. Structural bit growth of 20–25% could absorb a moderate price decline. It would not absorb every downturn: at 25% bit growth and a 30% ASP decline, revenue still falls 12.5%.
The matrix also describes changes between two endpoints. It does not imply that 25% annual bit growth permanently offsets a 20% price decline every year, or that all producers share the same growth rate.
A Higher Revenue Floor Comes Before a Higher Earnings Floor
Revenue protection is only the first step.
Suppose a business begins with revenue of 100, cost of sales of 60 and operating expenses of 20. Gross profit is 40 and operating profit is 20.
Now let bits rise 25%, ASP fall 20%, and cost per bit decline 10%. Revenue stays at 100, but cost of sales rises to 67.5: 60 × 1.25 × 0.90. With unchanged operating expenses, operating profit falls to 12.5, a 37.5% decline.
This is an illustrative GMS scenario, not a forecast. It assumes costs scale with shipped bits; actual fixed costs, inventory timing and utilization will change the result.
It nevertheless explains the distinction: structural bit growth can raise the revenue floor before it raises the earnings floor.
Mix, manufacturing cost per bit, utilization and inventory determine the gross-margin outcome. R&D and other operating costs determine how much reaches operating earnings. Taxes, financing and non-operating items affect net income. The equity required to fund expansion then determines ROE.
A producer can earn more absolute profit at its next trough while generating a lower return on a much larger equity base.
The new-equilibrium thesis must therefore pass every stage: revenue resilience, margin resilience, operating earnings resilience and returns on capital. A revenue matrix cannot establish all four.
What Previous Cycles Actually Show
For financial comparability, the table uses Micron’s reported GAAP results. Event time is anchored to each completed cycle’s quarterly revenue peak, not a claimed common DRAM/NAND ASP peak. The two products have different pricing paths, and the public disclosures do not provide a fully harmonized industry ASP series.
Current Q4 is an observation, not an established cycle peak. FY2024 Q3 already includes the emerging AI/HBM recovery and is not a pure pre-AI control. Operating margins are reported or calculated from unrounded dollar figures.
The first downturn offers a direct test of the price–bit mechanism. In Q2 2019, SK Hynix’s DRAM shipments increased 13% sequentially, while ASP fell 24%. NAND shipments increased 40%, while ASP fell 25%. Total company revenue nevertheless fell 5%, and operating profit fell 53%.
More bits did not prevent a severe earnings decline. In NAND, lower prices also stimulated demand—a reminder that rising bits can partly be a response to price weakness rather than independent structural demand.
https://news.skhynix.com/en/sk-hynix-inc-reports-second-quarter-2019-results/
The 2020–2023 cycle supplies an even clearer revenue-floor example. In Micron’s fiscal Q3 2023, NAND shipments increased in the upper-30% range sequentially, while ASP declined in the mid-teens. NAND revenue rose 14%. Total company revenue stabilized, rising about 2%.
Yet Micron still posted a $1.761 billion GAAP operating loss. Revenue recovery preceded earnings recovery.
Inventory and utilization explain part of that gap. Micron’s fiscal 2023 accounts recorded $1.83 billion of inventory write-downs and $382 million of period costs from facility underutilization. Subsequent sales of written-down inventory could also improve reported margins without demonstrating an equivalent improvement in underlying manufacturing economics.
https://www.sec.gov/Archives/edgar/data/723125/000072312523000054/mu-20230831.htm
ROE followed the earnings collapse. Using annual GAAP net income divided by average opening and closing shareholders’ equity, GMS calculates Micron’s ROE at approximately 18.5% in fiscal 2022 and −12.4% in fiscal 2023. These are annual calculations, not annualized quarterly figures.
https://www.sec.gov/Archives/edgar/data/723125/000072312522000048/mu-20220901.htm
The present boom has not yet supplied comparable post-price-peak observations. Its exceptional starting margins may provide a considerable cushion. How much survives the next supply response remains an open question.
Supply: The Test Is Larger Than the Fab Count
Future supply combines new wafer capacity, process migration, yield improvement and product allocation. Packaging can constrain HBM output, while node transitions can increase conventional-memory bits before a new fab contributes materially.
https://www.samsung.com/global/ir/reports-disclosures/public-disclosure-view.84599/
https://news.skhynix.com/en/fab-facility-investment-2026/
https://investors.micron.com/files/doc_financials/2026/q4/Q4-26-Earnings-Deck.pdf
CXMT says its G5 platform increases gross die per wafer by at least 50% versus G4, normalized to an 8Gb density baseline. That is a company claim about gross die density—not a 50% increase in yielded, qualified industry supply.
The economic risk does not require CXMT to lead HBM4. Additional mainstream DRAM can displace other suppliers’ domestic sales, change allocation and affect marginal pricing. Chinese market share alone does not prove global competitiveness, but a competitive pressure channel exists before technology leadership is reached.
Micron’s fiscal 2026 net capex reached $27.37 billion. The company says most of its newly increased fiscal 2027 construction spending supports cleanroom availability in late 2028 and beyond. Spending is already responding to today’s returns; much of the physical response arrives later.
That supports 2027 as an early-warning year and 2028–29 as a stronger test window. It does not establish a scheduled oversupply event. Process-driven growth can arrive earlier; equipment deferrals and yield problems can push effective supply later.
The decisive question is whether AI bit demand can absorb the combined output of the Big Three, process improvements and an expanding fourth large-scale DRAM supplier.
Demand: Count Units, Content and Efficiency Separately
The demand case is broader than HBM, but it cannot be reduced to an AI capex headline.
Independent evidence supports the infrastructure expansion. NVIDIA’s quarter ended July 26 produced $89.0 billion of Data Center revenue, up 117% year over year. That is observed supplier revenue, not a count of memory bits.
TrendForce’s August forecast put 2026 AI-server shipment growth near 31%. NVIDIA’s Rubin architecture supports up to 288GB of HBM4 per GPU. Together, units and content establish credible demand mechanisms; they do not establish a precise global bit-demand result.
https://www.trendforce.com/presscenter/news/20260803-13161.html
Micron forecasts high-teens server-unit growth in calendar 2026 and 2027. Crucially, it also says expected content growth is modestly lower than previously anticipated amid tight memory supply. The shortage is already affecting configurations.
Consumer demand supplies another constraint. IDC’s August 26 smartphone forecast projected a 16.7% shipment decline in 2026; its June PC outlook projected an 11.3% decline. These are dated independent forecasts, not completed full-year outcomes. They show why premium-device content growth cannot simply be extrapolated across all units.
AI itself is not price-inelastic. Google’s TurboQuant research demonstrates a mechanism for compressing KV-cache memory. It is evidence of optimization potential, not evidence that aggregate memory demand will fall. Lower cost per workload can encourage more workloads, leaving the net effect unresolved.
Customers can also select lower memory configurations, shift architectures or delay procurement. Conversely, falling memory prices may unlock configurations and deployments suppressed by the shortage. Both responses belong in the demand analysis.
Industry Shipments Are Not Unconstrained Demand
Micron’s latest forecasts put calendar-2026 industry DRAM bit-shipment growth in the mid-20s and NAND in the low-20s. For 2027 and 2028, it expects DRAM growth in the low-20s and NAND in the mid-20s, with both markets supply-constrained.
These are Micron forecasts of industry shipments. They are neither actual future outcomes nor a direct measurement of underlying demand without supply constraints.
Three series must remain separate: Micron’s own quarterly shipments, total industry shipments and underlying demand at a given price and configuration. Market-share changes, inventory movements and supply shortages can drive wedges between them.
The forecast strengthens the possibility of sustained bit growth. It does not independently establish the growth rate demand will maintain after prices fall and new capacity arrives.
DRAM and NAND Face Different Tests
DRAM has the clearer differentiated-demand mechanism: HBM requirements, server DRAM content, customer qualification and allocation of scarce front-end capacity.
NAND benefits from enterprise SSDs and inference storage, but has a different manufacturing response. Higher layer counts, cell-level changes and renewed wafer capacity can materially increase bits. Consumer exposure also remains important.
The forecasts are not unanimous. TrendForce’s July 30 outlook expected DRAM to remain tight in 2027, but NAND to move toward looser conditions in the second half as supply expanded. Micron’s September 30 outlook instead projected supply constraints in both NAND and DRAM through 2028.
The observations have different dates and assumptions. We should not erase the disagreement or treat the older forecast as equally current. The public evidence nevertheless does not establish a single agreed NAND balance for 2027–28.
Enterprise SSD demand may tighten the premium segment while other NAND segments weaken. The DRAM/HBM thesis is partially supported; NAND’s aggregate earnings floor is more uncertain.
SCAs Can Dampen the Cycle. Their Downturn Performance Is Untested.
Micron’s Strategic Customer Agreements are meaningful evidence of a changing commercial structure.
At Q4, the company had signed 26 SCAs, representing more than 35% of its expected revenue through 2030. About three-quarters of that expected SCA revenue had defined pricing frameworks, mostly with floor-and-ceiling bands. The remaining quarter was subject to periodic market-based pricing negotiations.
The denominator: approximately 25% of expected SCA revenue, not 25% of all Micron revenue, is periodically repriced to market.
These agreements are not one uniform fixed-price product. Take-or-pay commitments and deposits can increase planning visibility and the cost of walking away. They do not disclose the actual price floors, termination rights, renegotiation clauses, penalties, default provisions or treatment of extreme floor/spot divergence.
Micron also disclosed approximately $150 billion of remaining performance obligations associated with SCAs having determined pricing frameworks. RPO is contracted revenue-related visibility—not recognized revenue, cash profit or a guarantee of earnings.
https://investors.micron.com/files/doc_financials/2026/q4/Q4-FY26-Prepared-Remarks.pdf
The strongest bullish interpretation is that contractual minimum prices preserve substantially better margins through the next cycle. Micron explicitly expects floor-pricing margins above previous cycle peaks. That is management’s expectation, not an observed downturn result.
Its Q3 filing also acknowledges that customer failures to honor commitments could require enforcement and create disputes affecting the business and customer relationships.
The economic stress test is straightforward. If market prices collapse far below a contractual floor, customers have an incentive to seek some adjustment: renegotiation, reduced volumes, termination where permitted, a penalty settlement or another commercial arrangement. Those are possible outcomes, not claims about undisclosed contractual rights.
Suppliers have incentives to enforce commitments, but also to preserve major customer relationships and future platform access. The eventual outcome depends on enforceability, switching costs, supply assurance and bargaining power.
SCAs are therefore a potentially important cycle-dampening mechanism whose effectiveness has not yet been tested through a severe memory downturn.
Long-term agreements may dampen the next memory downturn, but they cannot be assumed to repeal commodity economics.
Samsung and SK Hynix: Updating the Original Valuation Thesis
The original comparison was about sustainable company returns, not assigning the same AI multiple to two memory producers.
Samsung’s Q2 2026 revenue was KRW 171.5 trillion and operating profit KRW 89.5 trillion. Its Device Solutions division generated KRW 89.2 trillion of operating profit. The presentation reported quarterly ROE of 56%; this is a period-based measure, not proof of through-cycle ROE.
The same presentation showed a KRW 0.7 trillion operating loss in Mobile eXperience and Networks. High component costs can benefit Samsung’s semiconductor operations while hurting its device businesses. Diversification changes the exposure; it does not automatically insulate consolidated returns.
SK Hynix’s Q2 release showed KRW 79.32 trillion of revenue and KRW 60.54 trillion of operating profit, alongside the HBM4 ramp and advanced-process transitions. Its more concentrated memory exposure transmits AI demand more directly into operating performance—and remains more directly exposed to a memory downturn.
There is an additional earnings-quality issue. SK Hynix’s filed preliminary Q2 figures showed pretax profit of KRW 122.71 trillion against operating profit of KRW 60.54 trillion. Net profit reached KRW 93.92 trillion. The large non-operating contribution means reported net earnings cannot simply be annualized as recurring memory earnings or normalized ROE.
For Samsung, the new evidence strengthens the execution case and the possibility of improved memory returns. For SK Hynix, it strengthens the concentrated AI-memory earnings case. Neither company has yet shown what its next operating-profit and ROE trough will be after the supply response.
The original warnings have advanced unevenly. Supply investment and process migration are visible. Price collapse has not appeared in Micron’s latest results. Consumer configuration pressure is visible; industry-wide earnings deterioration is not established. Micron’s inventory days rose to 129, but management attributed the increase partly to build-ahead and manufacturing compensation absorbed into inventory. One higher DIO observation is not proof of excess market supply.
The thesis remains about the next trough, not the size of the current peak.
How Much New Equilibrium Is Already in the Price?
Rapid profit accumulation makes the old book denominator stale. Forecast revisions and non-operating items complicate the earnings denominator. Dividing these mismatched multiples to infer current ROE would therefore be misleading.
The snapshot shows that investors are paying substantial premiums to the historical equity base while anticipating much larger earnings. It does not establish exactly how much permanent ROE improvement is priced in.
The original justified-P/B framework remains useful as a sensitivity:
P/B = (ROE − g) / (r − g)
Using illustrative assumptions of a 10% required return and 3% sustainable growth, normalized ROE of 15%, 20%, 25% and 30% corresponds to justified P/B of approximately 1.71×, 2.43×, 3.14× and 3.86×.
Conversely, on a consistent normalized book base, paying 3×, 4× or 5× book requires approximately 24%, 31% or 38% sustainable ROE under those assumptions. These are model sensitivities, not fair-value estimates for either stock; the framework assumes a stable relationship among growth, reinvestment, distributions and returns.
The critical distinction is between a higher trough ROE and a higher sustainable through-cycle ROE. A better trough can improve the latter, but does not by itself justify treating the trough or today’s peak as permanent.
Samsung’s broader asset base and SK Hynix’s concentrated exposure require separate normalized-return assumptions. Buybacks, retained earnings, investment gains and new capacity also change the equity denominator.
A higher future earnings floor can justify a higher valuation floor. If the price already assumes an improvement beyond what the next cycle delivers, the business can improve while the stock still disappoints.
The Strongest Counterarguments
The strongest objection to the new-equilibrium thesis is that current demand visibility was established during a shortage. Advance ordering and supply assurance may overstate the pace of end-use consumption once availability improves.
Second, process migration, better yields and product reallocation can add bits before the headline fab openings. CXMT’s conventional supply increases the amount that AI must absorb even without leadership in premium HBM.
Third, technology differentiation does not guarantee a permanent margin premium. Competitors can qualify, customers can redesign systems, and a scarcity-priced conventional product can temporarily generate higher margins than HBM.
Fourth, more inference does not automatically produce proportionally more installed memory. Compression, lower configurations and shifts between storage tiers can offset part of the workload growth.
Fifth, even a successfully defended revenue floor can coexist with lower profits and ROE as depreciation, operating costs and invested capital rise. Long-term contracts can also induce supply investment, making future stability depend more heavily on the promised demand being realized.
There is a serious counterargument to excessive pessimism as well. If lower memory prices unlock enough suppressed demand, stronger utilization, falling cost per bit and differentiated products could prevent a repeat of the old earnings trough. The starting margins and stronger balance sheets may provide substantial resilience. That is the possibility the research must continue to test.
What Would Establish—or Weaken—the New Equilibrium?
The thesis strengthens if, during a sustained ASP decline, industry bits continue growing without inventory accumulation; cost per bit and differentiated mix protect margins; contracts perform close to their disclosed economic intent; and the expanded asset base still earns higher normalized returns after new supply ramps.
It weakens if shipment growth reflects inventory loading, ASP falls faster than bits and costs can compensate, utilization declines, or the next earnings and ROE trough approaches previous-cycle levels despite much greater AI exposure.
For now, the demand mechanism is stronger than it was when this series began. The evidence for a permanent earnings floor remains incomplete, and NAND carries greater uncertainty than DRAM/HBM.
Bits are beginning to do more of the work, but not yet enough to prove a new equilibrium.
AI does not have to eliminate the memory cycle to change valuation. It must change what producers can earn after pricing normalizes and today’s capacity commitments become actual supply.
That is the next test for Samsung, SK Hynix and Micron—and the test that determines how far the original GMS thesis has actually progressed.