# NAND Flash Stays Short Through Mid-2027: Slower Price Gains Are Not a Closing Gap > Notes and reflections after listening to the Statementdog podcast episode from 2026-09-06. On the direction of China's new capacity, the split between consumer and server demand, and what changed inside those five-year supply contracts. Educational, not investment advice. Published: 2026-09-08 Locale: en Tags: NAND Flash, Memory, Supply Cycle, Long-term Contracts, Industry Analysis ![A night-time fab shipping corridor, racks in the foreground fully loaded and lit, growing emptier and darker toward the far end](/covers/caibaogou-2026-09-06-562-nand-flash-2027-cover.png) > The exchangeable value of a commodity is regulated not by those who produce it under the most favourable circumstances, but by those who continue to produce it under the most unfavourable. > > — David Ricardo, *On the Principles of Political Economy and Taxation* (1817) ## What This Episode Covers Episode 562 of the Statementdog podcast, released 6 September 2026, focuses on the NAND Flash industry, with an opening section answering listener worries about China's new memory capacity. The headline conclusion: the shortage runs through mid-2027, but the rate of price increases slows from here. What held my attention was not the conclusion but the path to it. Two judgments made me stop and rewind. One: the market treats a slower rate of price increase as evidence the gap is closing, and the host refuses that inference. Two: the market read those five-year contracts as bearish, and he argues it read the headline without reading the clauses. ## Key Points **1. Slower gains are a return to normal, not a turn.** Prices rose 80% to 100% quarter over quarter in the first half. That is not a baseline. In this industry, a 5% quarterly increase used to be a big deal — TSMC negotiating 10% to 15% on advanced nodes for next year shook the sector. So a low-double-digit quarterly increase in the second half, which compounds to roughly 20% over six months, is healthy, not exhausted. As the host put it: nobody notices a 1% gain now, but a few months ago, watching your assets rise 1% in a day would make your week. **2. The threat from China's new capacity depends on which direction it walks.** The framework splits memory into advanced and mature nodes. CXMT sits at roughly 16nm, three to four generations behind Samsung, SK Hynix and Micron. Building the same DDR5 on a trailing node means fewer bits per wafer and worse yield. So however alarming the wafer-start growth looks, the translation into actual bit supply carries a heavy discount. Flip it around: if CXMT turned toward mature nodes instead, the pressure would land on Nanya and Winbond, who sell into the same market. Right now it appears to be pushing on all fronts, tilting toward advanced nodes — which is breathing room for the mature-node players. **3. Five years to catch up; beyond that, nobody knows.** The host noted he said "five years" two or three years ago and still says five years now — not a stock phrase, but a fresh estimate at each point in time. He named three tailwinds for China: domestic substitution creating a protected home market, this cycle filling CXMT's coffers alongside IPO proceeds, and the fact that some of its quoted prices now sit above overseas equivalents rather than undercutting them. He also conceded the counterpoint: the window for buying equipment has closed, but delay is not prevention. **4. Consumer and server demand split right here.** Memory used to be 10% to 15% of a phone or laptop's bill of materials. It is now 40% to 50%. A handset maker at 20% gross margin is doing well — with a cost line moving like that, every unit shipped bleeds. So downstream buyers push back, or simply stop buying. The server side is the mirror image: hyperscalers run gross margins north of 50%, and they treat memory as a four-to-five-year asset with an ROI, not a current-period cost. The same price increase erases one buyer's entire margin and spreads across five years for the other. Nearly all of the second-half price growth now comes from enterprise SSD. **5. Phone sales are falling for the opposite reason than you would guess.** Low-end handset sales across Southeast Asia have collapsed. That looks like consumer weakness. The reading here is a supply shock: manufacturers cannot get the parts, and where they can, there is no margin in it — so they cut low- and mid-tier lines, concentrate memory into premium models, and raise prices to cover cost. Volume falls as a consequence. Meanwhile aggregate retail consumption keeps growing. The money did not disappear; there were simply no phones to buy, or it went somewhere else. **6. This round of five-year contracts contains different clauses than the last.** Historically a "long-term agreement" gave the seller a four-quarter forecast of purchase intentions. The only legally binding piece was NCNR — non-cancellable, non-returnable — on orders already placed for the current quarter. So every time the cycle turned, the agreement evaporated, and the industry stopped taking the phrase seriously. Three things changed. Purchase volumes are binding for every quarter across five years. Breach triggers take-or-pay: leave the goods, pay anyway. And buyers post a deposit up front, roughly 20% of contract value — on Micron's hundred-billion-scale agreement, that is a full year's worth — held in a supervised bank account or paid straight to the supplier. The seller gives up a price ceiling, anchored to one quarter's pricing with no upward revision. It receives five years of visibility. The market is pricing only the first half of that trade. ## Going Deeper ### "This company earns a fortune. Why does it trade at three or four times earnings?" This is the part of cyclicals that confuses people. You open the filings, earnings are at a record, and the multiple says three. The first instinct is that the market knows something you do not. What the market knows is narrower than that: it does not believe the earnings survive to 2028. A P/E is not pricing the current profit. It is pricing how long that profit lasts. Visibility of one year buys one year of payment. That has little to do with whether the business is good and everything to do with how far ahead anyone can see. Follow that thread and the contract story changes shape. If the seller trades near-term upside for five years of committed volume backed by cash deposits, the premise behind the discount — visibility of one year — starts to loosen. This does not guarantee a re-rating; whether the market believes it is a separate question. But the ground of the argument shifts, from "how much more can prices rise this year" to "what does this business look like in five years." One line from the episode stayed with me: operators do not run a business by watching the share price. For a memory maker, the boom-and-bust swing is the real pain. Trading short-term upside for stability pencils out on their books. What buyer and seller are negotiating and what the market is calculating were never the same thing. ### "The news says a big contract was signed. Bullish or bearish?" I used to read "long-term agreement" as bullish. Then I read it as meaningless, having heard too many stories of contracts dissolving the following year. This episode reminded me that both reflexes are reactions to a headline. The difference lives in the clauses, and there are three places to look. First: is the volume committed, or is it a forecast? An agreement that only supplies visibility is a memorandum between two parties exchanging information. Second: what does breach cost? Written as take-or-pay, the document moves from gentlemen's agreement to legal instrument. Third: is cash paid up front? The deposit is the hardest signal available — a buyer handing over a fifth of contract value into a supervised account before receiving any goods tells you how real their fear of shortage is. Answer those three and you find that things sharing the name "long-term agreement" range from a commitment to a courtesy. The market usually reacts to the top layer only. The episode offers a corroborating detail: one major chip company's purchase-commitment reserve jumped from around a hundred billion to close to three hundred billion. Nobody could explain the jump at the time. Read against the contract structure, it fits. The test I wrote down: when a contract makes the news, before judging direction, ask what breach would cost if conditions reversed next year. If you can answer that, you have read it. ## "This number is falling — is the cycle ending?" The most counterintuitive stretch of the episode is the explanation for falling phone sales. One downward-sloping line can mean consumers will not buy, or that manufacturers cannot build. The first is demand contraction; the second is a supply shock. The two point in opposite directions — contraction says the cycle is ending, a shock says the gap is still there, perhaps tighter. Separating them takes a piece of outside evidence, not intuition. Here the evidence is aggregate consumption: if consumers were out of money, retail spending would not still be growing. Since it is growing, the drop in phone sales needs an explanation from the supply side. What I admire about that move is that it does not stop at "I think it's supply." It goes and finds a number that would have embarrassed the claim if it were wrong. The same logic applies to pricing. A narrowing rate of increase can mean the gap is filling, or that prices are contractually capped. So the episode steps around price entirely and returns to bits: supply grows roughly 21% to 25% next year, in line with this industry's historical average — no surge. And Samsung, SK Hynix and Micron all make both DRAM and NAND, so any perceived risk in NAND redirects capex toward DRAM, which is tighter and more profitable. On the demand side, something new appeared: DRAM is short enough that vendors are offloading KV cache to NAND, which loads NAND with demand that originally belonged to DRAM. The two are linked for now. While DRAM stays tight, NAND has little room to loosen on its own. ## Where to Look Next - Statementdog podcast, episode 562 (6 September 2026), on the NAND Flash industry and China's new capacity - The preceding episode in the series, covering Phison and the structure of NAND demand - Public filings and earnings-call materials from Taiwan's memory suppliers - Supplier disclosures on long-term supply and purchase agreements ## One Thing Worth Taking Away The idea I am keeping: **the same number falling can be caused by two opposite things.** Phone sales down can mean "nobody wants one" or "nobody can build one." Whichever you pick, everything downstream turns with it, and the two roads end in opposite conclusions. The hard part is that both causes look identical from where you stand — all you have is a line pointing down. Here is something I have tried, and it works outside investing too. Pick something around you that has gotten smaller. How often friends gather. How fast someone replies. The queue at a shop you like. Write down two opposite causes: one where they do not want to, one where they cannot. Then, before deciding, think of a single clue — one that, if you saw it, would tell you which. Maybe the slow replier is still quick in another group chat. Maybe the queue is shorter but delivery orders are stacked at the door. What happens most often when I do this: I discover my original verdict had no evidence behind it at all, only whatever mood I was in. Failing to find a clue is itself a result — it tells you that "they just don't want to talk to me" was a guess.