LUNAROPS · OPERATIONAL UPLINK 100% UPTIME 1,247d POSTS 893 JEFF.MOON@LUNAROPS.DEV UTC --:--:--

The NAND Commodity Cycle

19 min read The Physics of Computing
nandmemory-marketcommodity-cyclesemiconductorseconomicsstorage
Contents

A NAND flash bit is one of the most sophisticated objects humanity manufactures at scale — a charge trap a few dozen nanometers wide, stacked hundreds of layers deep, read by inferring voltage thresholds at the edge of physical noise — and the market treats it exactly like a barrel of crude or a hog carcass. The price of that bit can fall 70% in eighteen months and then double in two quarters, not because anyone invented anything, but because too many fabs were built at the same time, or too few. This is the central, uncomfortable truth of the storage industry: the engineering is heroic, the economics are agricultural. NAND is a commodity, a bit is a bit, and the thing that actually decides which products ship and which companies survive is not lithography or layer count but where you happen to sit in a violent, multi-year supply-and-demand cycle that nobody controls and everybody loses to eventually. This post is about that cycle — why it exists, why it is structurally unavoidable, why it has ground the industry down to roughly five survivors, and how the 2022-2023 crash gave way to the AI-driven supply crunch that, as of mid-2026, has makers raising prices faster than at any point in living memory.

Why a Bit Is a Commodity

Start with the thing that makes NAND a commodity rather than a differentiated product: fungibility. A 1-terabit TLC die from Samsung and a 1-terabit TLC die from Micron are, to the system integrator buying ten million of them, interchangeable within a price band. They plug into the same controllers, speak the same ONFI and Toggle interfaces, endure roughly the same write cycles, and store the same bits. There is no brand loyalty in a DRAM-less SSD bill of materials. When the product is fungible, buyers optimize on one axis — price per bit — and sellers have no pricing power beyond the marginal cost of the next wafer.

This is the textbook definition of a commodity, and it puts NAND in the same economic family as oil, copper, memory’s sibling DRAM, and the canonical teaching example: pork. The “pork cycle” — described by economists in the 1920s — is the boom-bust pattern that emerges whenever a product is undifferentiated, demand is relatively inelastic, and supply takes a long, fixed lead time to adjust. Farmers see high hog prices, everyone breeds more pigs, the pigs all mature at once, the market floods, prices crash, farmers slaughter their breeding stock, supply collapses, prices spike, and the whole thing repeats. Swap “pigs” for “fabs” and you have the memory industry.

The fungibility also explains why differentiation efforts mostly fail to escape the cycle. Vendors try: power-loss protection, faster interfaces, zoned storage and FDP, specialized endurance grades. These add margin at the SSD and module level, but the underlying NAND die — the thing that consumes the $15-billion fab — remains a commodity whose price is set by aggregate supply and demand, not by any one vendor’s cleverness.

The Capex Lead Time That Breaks Everything

The pork cycle needs one more ingredient to turn vicious: a long, irreversible lead time between deciding to build supply and supply actually arriving. In hogs it is the gestation-plus-maturation period. In NAND it is the time and money it takes to stand up a fab, and the numbers are staggering.

ROUGH NAND CAPEX ECONOMICS (order of magnitude, 2025)

  Cost of a new leading-edge fab        ~ $10-15 billion
  Equipment as share of total capex     ~ 70%
  Time from groundbreaking to volume    ~ 2-3 years
  Industry-wide memory capex (annual)   > $40 billion
  Useful life before node obsolescence  ~ 4-6 years

A NAND fab is one of the most expensive single objects a private company can choose to build, and the bulk of that cost is etch, deposition, and metrology tools — equipment that takes 18 months to order, install, and qualify. (The full ordeal of getting one of these facilities running is its own story; see how a wafer fab works.) The decision to add capacity must therefore be made two to three years before that capacity sells a single bit, based on a demand forecast that far out is essentially a guess.

Here is the trap. When prices are high, every maker’s forecast says “demand is strong, build more.” They all break ground at roughly the same time, because they are all looking at the same high prices. Two to three years later, all that capacity arrives in the same window — and it arrives whether or not demand actually grew to meet it. If demand softened in the interim (a recession, a weak smartphone cycle, a post-pandemic inventory hangover), the new supply lands on a market that cannot absorb it. Prices collapse. The signal to stop building came years too late.

Supply, in other words, is inelastic in the short run and lumpy in the long run. You cannot add 5% more NAND next month; you can only add a whole fab’s worth in three years. Demand, meanwhile, swings continuously. Mismatched cadences guarantee that supply and demand are almost never in balance — they cross each other like two sine waves at different frequencies, and the gap between them is the price crash or the price spike.

Fixed Costs and the Suicide Pact

The capex lead time creates the gluts. Fixed-cost economics make them catastrophic, because once a fab is built the incentive is to keep it running at full tilt even while losing money on every bit.

Consider the cost structure of running a depreciated fab:

SIMPLIFIED COST-PER-BIT INTUITION

  Total cost  =  Fixed costs  +  Variable costs
                 (depreciation,   (wafers, gases,
                  facility,        chemicals,
                  most labor)      marginal power)

  Fixed costs dominate: a fab represents billions in
  sunk capital that must be depreciated whether the
  tools run or sit idle.

  Once the fab exists, the decision each quarter is NOT
  "is the total cost below the selling price?"
  It is "is the SELLING PRICE above the VARIABLE cost?"

  As long as price > variable cost, every wafer you push
  contributes something toward covering the fixed costs.
  Idling the fab covers nothing — you still pay the
  depreciation and the facility bill, and you lose the
  contribution margin too.

This is the economics of any high-fixed-cost, low-variable-cost business — airlines flying half-empty planes, hotels discounting empty rooms — and it produces the same self-destructive behavior. During a glut, the rational move for each individual maker is to keep producing as long as price exceeds variable cost, because shutting down means eating the fixed costs with zero offsetting revenue. But when every maker reasons this way, the collective result is that nobody cuts supply until the pain is unbearable, and the glut persists far longer and far deeper than any single firm would choose. It is a suicide pact written by accounting.

Worse, the fabs are not idle by default — they are spitting out bits even when those bits are worthless, because the alternative is spitting out nothing while still paying for the building. This is why memory downturns feature the surreal spectacle of companies reporting negative gross margins, selling product below total cost for quarters on end, all of them bleeding, none of them blinking, until someone finally announces wafer-start cuts deep enough to matter. Production discipline is the only lever, and it only gets pulled when the losses become existential.

The Survivors: Five Players and Why Scale Is Life or Death

The cycle is a meat grinder, and it has done what meat grinders do: reduced a field of more than a dozen NAND makers to a handful. Each downturn kills or absorbs the weakest. The economics are Darwinian — the maker with the lowest cost per bit and the deepest balance sheet survives the glut; everyone else gets acquired or exits. After three decades of consolidation, NAND is now an oligopoly of roughly five.

Player Position (2025-2026) Notes
Samsung Market leader, ~30%+ share Scale, deepest pockets, 400+ layer V-NAND with hybrid bonding; sets the tone on capex and price
SK Hynix + Solidigm ~20%+ combined Folded in Intel’s NAND/SSD business (Dalian fab + Solidigm); strong in enterprise QLC SSDs
Micron Mid-teens share Only US-headquartered maker; aggressive on QLC and high-layer-count nodes
Kioxia + SanDisk Combined #2-class bit output Kioxia (ex-Toshiba memory) runs a JV with SanDisk, which spun out of Western Digital in Feb 2025 (NASDAQ: SNDK)
YMTC Emerging, China-domestic Growing on domestic demand; constrained internationally by US export controls

A few things in that table are worth dwelling on. SK Hynix’s position is the product of buying Intel’s flash business outright — first the Dalian fab in 2021, then the IP and remaining assets closing in March 2025 — which is how a DRAM-strong company bought its way to NAND scale and a leadership position in enterprise SSDs. The Kioxia/SanDisk arrangement is the strangest survivor structure in the industry: two nominally separate, publicly traded companies (Kioxia in Tokyo, SanDisk on NASDAQ after its February 2025 spin from Western Digital) that jointly fund and operate the same fabs in Yokkaichi and Kitakami. They share the wafers and compete on the products. That deep coupling exists precisely because neither could afford the capex alone — scale is so existential that former competitors co-own the foundry.

YMTC is the wild card and the geopolitical fault line. China’s national champion has the technology to compete on layer count, but its access to the leading-edge tools and Western markets is throttled by export controls — the same regime that governs why you can’t buy an H100 in Shanghai. Restrict a memory maker’s access to deposition and etch equipment and you cap its bit-cost trajectory, which in a commodity business is a slow-motion death sentence — or, from Beijing’s view, the reason to build a domestic equipment supply chain at any cost.

Why is scale this existential? Because in a commodity, your only durable advantage is cost per bit, and cost per bit is dominated by R&D and capex that must be amortized over volume. The maker shipping twice the bits amortizes the same multi-billion-dollar node-development bill over twice the output, so its cost per bit is structurally lower. In a glut, when price falls to variable cost, the high-cost producer loses money fastest and runs out of cash first. Scale is not a nicety; it is the thing that lets you survive the part of the cycle designed to kill you.

Bit Growth, Layers, and QLC as a Supply Lever

There is a second supply lever beyond building fabs, and it is unique to memory: you can make the bits you already have capacity for cheaper and denser, flooding the market without pouring a single new slab of concrete. This is the technology-transition lever, and it is why NAND supply can grow even when capex is flat.

Two transitions do the work. The first is vertical: 3D NAND architecture stacks cells in layers, and each generation adds layers — 128, 176, 232, 321, and now 400-plus — so the same wafer area yields more bits. By 2025, 3D NAND was effectively the entire market, and layer counts past 300 were shipping in volume. The second is horizontal in the value sense: cramming more bits into each cell. Moving from three bits per cell (TLC) to four (QLC) is a one-third capacity gain on the same silicon, the cleanest density win in the business — and, as covered in detail in QLC, PLC, and the density endgame, it comes with exponential costs in endurance and read margin that constrain where it can be deployed.

The supply-side point is that both levers convert engineering into bits faster than fabs can. “Bit growth” — the year-over-year increase in total bits the industry can produce — routinely runs ahead of “demand growth” measured in bits, and that gap is the structural source of price decline. Historically, NAND bit demand grew ~30% a year and makers built supply to match. The problem is that the levers don’t turn off cleanly: a node transition that was supposed to deliver 20% more bits to a growing market still delivers those bits if the market stops growing, and now you have a glut you can’t un-manufacture.

THE STRUCTURAL TENSION

  If  bit_supply_growth  >  bit_demand_growth   ->  prices fall
  If  bit_supply_growth  <  bit_demand_growth   ->  prices rise

  Layer adds and QLC make bit_supply_growth easy and cheap.
  Demand growth depends on phones, PCs, datacenters, AI --
  things makers do not control and cannot forecast 3 years out.

  Recently the industry has GUIDED DOWN long-run demand-growth
  assumptions from ~30%/yr toward ~10-15%/yr, which means even
  modest supply additions can tip the market into oversupply.

The downward revision of demand-growth assumptions is the quiet, important shift. When the industry believed demand would grow 30% a year forever, building supply aggressively looked prudent. Now that makers plan around 10-15%, the same aggressive bit growth from layer adds and QLC overshoots far more easily — which is exactly why the post-2022 discipline around wafer cuts has been more deliberate than in prior cycles. The makers learned, expensively, that the technology lever cuts both ways.

(Underneath all of this sits yield engineering: a node only delivers its promised bit-cost reduction once yields mature, so the same transition that floods the market in year three was a margin-killer in year one. The cost curve and the supply curve are the same curve viewed from different ends.)

The 2022-2023 Crash

The most recent full cycle is a clean case study, and it ran almost exactly to the textbook. The COVID era had inflated demand: locked-down consumers bought PCs and phones, datacenters expanded, and — critically — buyers abandoned just-in-time inventory in favor of stockpiling against supply-chain chaos. Makers read that demand as durable and kept supply growing. Then in 2022 the music stopped. Consumer electronics demand cratered, the inventory that buyers had hoarded turned into a glut they now had to work down before ordering more, and the over-built supply landed on a market that had vanished.

The collapse was brutal. From mid-2022, NAND contract prices fell more than 70%. Micron’s CEO described it as the worst supply/demand imbalance in DRAM and NAND in thirteen years. Hyperscalers — usually the steadiest buyers — bought roughly 54% less flash in 2023 than in 2022 as they drew down inventory. The NAND units at Samsung, Micron, and Kioxia all swung to operating losses; gross margins went single-digit and then negative. Micron reported a fiscal Q1 2023 with revenue down 47% year over year and a net loss.

The fixed-cost suicide pact played out exactly as theory predicts: for months the makers kept running fabs and selling below cost, because each one’s variable-cost math said “keep shipping,” and none wanted to cede share. Only when losses became existential did real discipline arrive. Through 2023 the cuts stacked up — Micron reduced wafer starts ~20% and guided to roughly 40% lower bit supply growth, Kioxia cut up to 30%, Samsung made its own large reductions. Production discipline, not a demand miracle, put a floor under prices. The cycle bottomed because the makers finally, collectively, stopped digging.

The 2024-2025 Recovery and the AI Crunch

The recovery began in 2024 and accelerated through 2025, and its character is what makes this cycle genuinely unusual. The early-stage recovery was the normal script: production cuts tightened supply, smartphone and PC inventories normalized, and prices stabilized. By Q2 2025, NAND wafer contract prices were rising 10-15% as the supply-demand balance flipped. Then a second, much larger demand vector arrived and changed the shape of the whole cycle: AI.

AI demand hits NAND through several doors at once. Training and inference clusters need enormous fast storage, and high-capacity enterprise SSDs — increasingly QLC, because the density and power-per-terabyte advantage over hard drives is decisive at AI scale — became the fastest-growing segment. NVIDIA’s Blackwell ramp pulled enterprise SSD volume hard. A nearline HDD shortage pushed even more capacity demand toward high-capacity QLC SSDs. The hyperscalers that had bought 54% less flash in 2023 came back as voracious buyers, and this time they signed long-term agreements to lock in supply rather than buying on the spot market.

By late 2025 and into 2026, this had tipped from recovery into something the trade press began calling a memory “crisis” or “supercycle.” Makers resumed output cuts in the second half of 2025 even as demand surged — supply discipline plus an AI demand wall is a recipe for a price spike, and that is what happened. Reported figures from this period are eye-watering: enterprise SSD prices guided up 53-58% quarter-over-quarter in Q1 2026, overall NAND contract prices guided up sharply into Q2 2026, and Samsung reportedly negotiating 2026 contract hikes in the 20-30%-and-up range. Kioxia stated its entire 2026 production was effectively sold out, with even long-term customers absorbing roughly 30% year-over-year price increases.

Two cautions are mandatory here. First, every one of those forward percentages is a guidance figure or a negotiation report, not a settled fact — memory price forecasts are notoriously volatile and have a long history of being revised hard in both directions. Treat any specific number about future quarters as a snapshot of expectations, not a promise. Second, the open question is whether AI demand represents a structural break from the pork cycle or merely a very large, very steep upswing that will eventually invite the same over-building and the same crash. The optimistic case is that profitability-based wafer allocation and long-term AI contracts have changed the industry’s behavior. The historical case is that every memory cycle has ended the same way — with too many fabs — and that the cure for high prices is, as always, high prices, which fund the next round of over-building. Bet on the structural break at your peril; the cycle has humbled smarter forecasters than any of us.

How the Cycle Decides Who Lives

Step back and the through-line is this: in NAND, the cycle outranks the engineering. A maker can have the best node, the highest layer count, the cleverest QLC, and still be destroyed if it is sub-scale when the glut hits — because the glut doesn’t care how elegant your charge trap is, only how low your cost per bit is and how much cash you have to lose. Conversely, a maker can survive a generation of mediocre technology if it has the scale and balance sheet to outlast the downturn that kills its rivals. Consolidation to five players is not an accident of business strategy; it is the cycle selecting for the only trait that matters at the bottom — survivability.

This is why the structures look the way they do. It is why Kioxia and SanDisk co-own fabs they could never each afford. It is why SK Hynix bought Intel’s flash business to reach scale rather than growing it organically. It is why Western Digital finally split storage’s two commodities — HDD and NAND — so each could face its own cycle on its own balance sheet. And it is why YMTC’s fate rides less on its engineers than on export-control policy. The product decisions follow too: QLC enterprise SSDs won the AI storage tier not purely on merit but because the cycle made HDD capacity scarce and made cheap, dense NAND bits the answer. The technology is the input; the cycle is the function that turns it into winners and losers.

For everyone downstream — the SSD vendor, the datacenter architect, the homelabber pricing out a ZFS pool that doesn’t lie — the practical lesson is to respect the cycle as a first-class planning variable. The right time to buy capacity is during a glut, when makers are selling below cost and begging you to take their bits; the wrong time is during a crunch, when the same bits cost double and “sold out” is the standing answer. The bit you store data on is a triumph of physics. The price you pay for it is a hog futures contract.

Verdict

NAND flash is a commodity, and that single fact dominates everything else about the industry. The combination of perfect fungibility, multi-year and multi-billion-dollar capex lead times, and fixed-cost economics that punish idling fabs produces an unavoidable boom-bust cycle — the same pork cycle that governs hogs and oil — in which supply and demand are almost never in balance and the price oscillates with a violence wildly disproportionate to the underlying technology. That cycle has consolidated the field to roughly five survivors, each of which exists in its current form because of the cycle: scale is the only durable defense, so the weak got absorbed and the strong co-own each other’s fabs. The 2022-2023 crash followed the script to the letter — over-building met a demand collapse, makers ran fabs at a loss until the pain forced discipline, prices bottomed only when wafer starts were cut. The 2024-2026 recovery, supercharged by AI and enterprise SSD demand into something resembling a shortage, is the same cycle running the other way, and the honest analyst’s position is humility: the forward pricing numbers are guidance, not gospel, and whether AI has genuinely broken the pork cycle or merely set up its largest crash yet is unknowable from here. What is knowable is the durable lesson — in storage, the engineering tells you what is possible, but the commodity cycle tells you what survives.

Sources

Comments