SanDisk: The bet on decommoditization

I've seen gluts not followed by shortages, but I've never seen a shortage not followed by a glut. — Nassim Taleb

That is almost a law — it's rooted in human nature. But applying it blindly to today's storage supercycle — "SSD is a demand induced shortage, therefore a glut is coming, therefore stock price is overvalued" — is far too naive. If the cycle were that mechanical, investing would be easy. What actually matters is how long, and how the ending arrives — and that turns out to depend on things the headlines never mention.

The bull case

The bulls are right at least in face value. The world is genuinely short of NAND flash, and it cannot fix that quickly: a new fab costs more than $10 billion^† and takes 1.5 to 2 years^† to build — so the shortage is locked in for years. And the buyer this time isn't the fickle phone upgrade cycle; it's the bottomless AI data center. One headline called the result a "pricing apocalypse that could last a decade^†." The stock has behaved accordingly — up about 40× in a year. (That's 40x at the time of writing – 06/06/2026 – the later you read it, the funnier the it will be to look at the realtime price chart.)

But let's not stop here. The interesting question isn't whether SanDisk is a good company — it's what this industry actually is.

Behind a 251% yoy revenue spke and 78% gross margin

Below is the most income statements, on Q2 2026, both revenue and gross margin shoot up.

When a company's revenue triples, you expect its costs to rise too. Look at what SanDisk's costs did while revenue went vertical:

In the most recent quarter (ended early April 2026), revenue was $5.95B, up 251% year over year. And the cost of revenue? $1.29B — down 1.9%. They spent slightly less to make the product. The reason isn't a cost breakthrough — it's that they sold fewer units for far more money: on the earnings call the CFO said bit shipments were flat year over year and actually down high-teens^† sequentially. The entire move came from price: NAND contract prices leapt 55–60%^† in a single quarter.

So Q2 2026's 78% gross margin (4.66 / 5.95) is neither moat nor efficiency. It's a price — a commodity, briefly, priced like a patent. Which raises the only question: what kind of industry lets a commodity print a software margin, and can it possibly last? To answer, you have to understand a fab that can't stop.

What this industry actually is

A fab that can't stop

SanDisk makes NAND flash — the chips that remember your photos when the power is off — in giant factories ("fabs") it co-owns with Japan's Kioxia. A fab is the opposite of a lemonade stand. It costs more than $10 billion and takes years to build, and almost every dollar of that is spent before a single bit ships — on the building, the lithography tools, the clean room. Once the fab exists, the cost of making one more bit — a sliver of wafer, some chemicals, power, a touch of labor — is a tiny fraction of the depreciation cost the income statement reports. A NAND fab is an enormous fixed cost bolted to a marginal variable cost.

That single structural fact is the hidden engine of the entire cycle, and turns every glut into a rout.

Here is the process. When demand softens and prices fall, the rational operator does not cut production — because idling a $10 billion machine wastes the fixed cost whether the line runs or not. As long as the price still clears cash cost (that small marginal cost), every extra bit sold throws off a little something toward the enormous fixed bill. So the correct move — for one maker, and therefore for all of them — is to keep the line pinned at 100% and dump the output at whatever the market will pay. Supply does not flex down to meet weak demand; price does. And it falls all the way to cash cost — and in the worst gluts, below it — before anyone finally blinks and idles capacity at a brutal loss. That is why this business doesn't have soft landings; it always crashes. The very fixed-cost fab that mints a 78% margin when bits are scarce is the same machine that drives price through the floor the moment they aren't.

The bit consumers

Where do all those bits go? Three channels: client (~62% — the manufacturer put SSDs inside PCs and phones), cloud / data center (~25% — enterprise SSDs, the AI-booming slice, up 233%^† in a single quarter and has way more lucurative gross margin), and consumer (~14% — retail cards, USB sticks, portable drives).

So the obvious question: if the data center is where the AI money is, why doesn't SanDisk just point the whole fab at enterprise SSDs and skip the cheap stuff? Two reasons, both structural.

First, the data center can't absorb a fab's entire output. Even at full tilt, enterprise is still only about a quarter of revenue — and remember, the line must run at 100%. The client and consumer channels aren't a distraction from the AI business; they're the ballast that keeps the fab from sitting idle when datacenter demand alone wouldn't fill it.

Second — more importantly — you can't pivot a fab to enterprise overnight, because selling into a hyperscaler is a slow, earned privilege. Before Microsoft or Meta will put a drive in a server, it has to pass qualification: many months of testing for sustained throughput, write-endurance over years of duty, power and thermals, firmware stability, and reliability across thousands of drives running at once — followed by an integration to one specific server platform.

A raw NAND bit is a pure commodity — and it stays one. What qualification adds is a thin skin on top of the bit: a controller, firmware, and an earned slot in a customer's server that takes years to win and can't be swapped overnight. That skin earns the fattest margins in the mix (the "higher-value mix" the CFO credited for the jump) and is why SanDisk can't conjure more enterprise drives on demand. But the skin is thin — NAND is the majority of the drive's cost — so a qualified enterprise SSD is still a commodity riding on a commodity bit: its price falls when NAND falls, just with a lag and a smaller swing. (In the 2022–23 glut, datacenter drives fell too; they just fell less, and later, than retail cards.) The qualification moat buys margin stability and stickiness, not immunity from the cycle. It is emphatically not HBM — the stacked memory that sits beside an AI GPU — where the engineering is most of the value and a single buyer makes scarcity real. Enterprise SSD is a commodity in a nicer suit. This is exactly why NAND has to manufacture durability with multi-year contracts, as we will see later.

A commodity facing a deflationary gravity

Here is the first thing the supercycle headlines bury: NAND is, structurally, a deflationary business. A gigabyte that cost roughly $2 in 2010 costs a few cents today. Its demand booms have almost always been volume stories at falling prices — when smartphones detonated NAND demand in 2011, average selling prices were set to decline 35 percent^† even as the market grew — bit shipments simply exploded. More bits, cheaper: that is the normal state of this business. Hold that against today's 78%.

On top of that price secular decline runs the demand cycle — the loop that has governed this business for thirty years.

A shortage lifts prices; fat margins lure everyone to add capacity; the new supply lands all at once; prices crash, often below the cash cost of making the chip; everyone cuts; the shortage rebuilds. The down-legs are violent precisely because of the fab that can't stop — output keeps pouring out into collapsing demand. As recently as 2022–23 the entire industry sold NAND below cash cost (more on that later). Three years on, we sit at the highest peak the loop has ever produced.

The oligopoly that can't act like one

Now the puzzle: NAND is a textbook oligopoly in headcount. Four suppliers control ~95%^† of output — Samsung, SK Hynix, Micron, and the Kioxia–SanDisk pair — behind a $10B+ fab and decades of process IP that no swarm of newcomers can clear. Concentrated industries are supposed to earn fat, stable margins. NAND doesn't: this same company earned a 22.5% gross margin as recently as early 2025 (shown in the above income statement table), and the whole industry was losing money in 2023^†. The SSD industry is an oligopoly in headcount, perfect competition in price.

To recap, the SSD has the following signature:

• A bit is a bit — no brand, no switching cost; everyone sells at the same spot price per gigabyte.
• A technology treadmill — density rises every node, so even flat wafer capacity grows the bit supply. You cannot hold barrels off the market the way OPEC can.
• The fab is almost all fixed cost — so in a glut the rational move is to run flat out and sell anything above cash cost, racing price to the floor.
• Capacity is lumpy and lagged — added in $10B chunks on a two-year delay, by everyone, at the top of the cycle.
• Significant entry barrier – huge upfront captital investment and engineering capacity fend off new competitors.

The competition level reflects in recent gross margin, to which the valuation is hyper sensitive.

The latest quarter printed a 78% gross margin; averaged across the trailing year it is 57% (the card below), and a year ago that trailing figure was about 30%. At the single-quarter level the jump is starker still: 78% today against 22.5% in the year-ago quarter, and outright losses in 2023.

Only memory got rich on price

One interesting perspective is to compare memory to other players in the AI buildout supply chain. Almost every winner grew by doing more — more GPUs, more wafers, more lithography systems — at margins that barely moved. Only memory grew by charging more:

The line between the two clusters is physical. Logic could pour more wafers and add a shift; memory's bits are gated by fabs that take years, so when AI demand hit, logic answered with volume and memory could only answer with price. That is why memory — and memory alone — is the cyclical joint of the AI trade: the one link whose boom is price, and price is the input that mean-reverts.

So how different is this time? To judge that, you need the two things the headlines skip — what this industry's history actually says, and what its survivors, scarred by the last bust, are doing differently now.

The last bust, and the scar it left

We keep pointing forward to 2022–23 glute, a textbook example of how brutal things are. You cannot read what SanDisk is doing now without that context.

Through the second half of 2022, the bottom fell out. In a single quarter NAND wafer prices fell an estimated 20–25%^†, and they kept sliding until contract prices were approaching the factory cost^† — the bare expense of pushing a wafer through the line. Then they went through it: by early 2023, makers were selling large volumes below what it cost to produce them. This is the fab-that-can't-stop, live: demand had air-pocketed, but the lines kept pouring out bits, and with output fixed, price did all the adjusting — straight into the ground.

SanDisk felt it directly; back then it was the flash half of Western Digital. The damage was stark: WD's flash business helped drive a $321 million^† operating loss in the December-2022 quarter, and the bleeding deepened to a $472 million^† operating loss the quarter after. There was no clever way out. The only lever that works in a glut is to make fewer bits — so WD did the thing the industry hates most: it cut flash wafer starts by 30%^†, idling a third of its own production to stop feeding the collapse. Surgical supply removal, not a new fab in sight.

And it wasn't just them. Every major NAND maker crossed into losses before 2022 was out — Micron, SK Hynix, Kioxia all bled red, capex plans torn up across the board. A concentrated, supposedly-rational oligopoly spent the better part of a year selling its product for less than it cost to make, because each player, individually, was still better off running the fab than idling it. The prisoner's dilemma from a few sections back, settled in billions of dollars of real losses.

That is the scar. The survivors of 2022–23 walked out of it determined never to let the fab run blind into a wall again. And this time, they think they have a way.

SanDisk's bet against the cycle: the New Business Models

If the cycle is the disease, SanDisk thinks it has a treatment. The company calls them New Business Models, or NBMs: multi-year supply contracts, backed by financial guarantees, that lock in volume, price, and supply with hand-picked data-center customers. The stated goal, in the company's own glossary, is to reduce the cyclicality of the business.

The scale is already serious. SanDisk has signed five such agreements; the three from last quarter alone carry $42 billion^† of minimum revenue, over $11 billion^† of financial guarantees, and $400 million^† of cash prepaid up front. Together they cover more than a third^† of the bits SanDisk expects to ship in fiscal 2027 — and management says that share is climbing.

Here is the clever part, and the trade-off. What kills a fab in a downturn isn't only low prices — it's that the line keeps running into collapsing demand. NBMs attack exactly that: the customer commits to consume a set volume every quarter for up to five years^†, and if they don't, financial commitments come to SanDisk immediately. In return, SanDisk gives up some price: the contracts have fixed and variable elements^†, so the locked bits won't catch the full upside if spot prices keep ripping. They are trading price variability for volume certainty — selling away some of what helps most in a boom (peak spot pricing) to insure against what hurts most in a glut (idle output).

So how protective is it, really? Today the locks cover roughly a third of fiscal-2027 bits, and management wants to drive that above 50%^† over the next several quarters — but even at that eventual target, half the output still rides the spot cycle, fully. And the guarantees (~$11B) back only a fraction of the contracted revenue (~$42B), so the floor is only as good as the customers' credit and their willingness to keep paying through a downturn. NBM doesn't repeal the cycle. In a historical lens, it may not even trade a crash for a soft landing — more likely for a softer crash.

The demand under all of it — and the second derivative

Everything so far is the supply side: a brutal cost structure, a fresh scar, a new tool to tame it. But the whole edifice rests on the other half of the ledger — the AI demand that switched the shortage on. How solid is that?

The scale is genuinely staggering. The four biggest spenders — Amazon, Microsoft, Alphabet, Meta — are on track for roughly $725 billion^† of capital spending in 2026, up 77%^† in a single year, most of it aimed at AI data centers, with the trajectory pointing past $1 trillion in 2027^†. SanDisk's enterprise bits ride directly on that river of money.

But AI spending staying high is not enough for NAND. It needs it to keep accelerating. The shortage is a gap between demand growth and supply growth — so the moment capex merely plateaus, even at a trillion dollars a year, bit-demand growth flattens, the supply the industry is busily adding catches up, and the glut clock starts ticking. Memory is leveraged not to the level of the AI boom but to its second derivative. For NAND, a boom that simply stays enormous is the beginning of the bust.

And there are real cracks in the acceleration. The spending is outrunning the cash it throws off: Amazon's free cash flow collapsed from $26 billion to $1.2 billion^† in a year and is set to turn negative in 2026 as a $200 billion^† capex budget laps its operating cash flow. Across big tech, the hyperscalers are each spending 21–35% of revenue^† on capex — more than the average utility today, and more than AT&T at the height of the telecom bubble^†. And the payoff is, so far, mixed: some formal studies found established companies, albeit eager to adopt agentic AI to boost productivity, struggled to get measurable output (an MIT-affiliated study found 95%^† of enterprises investing in generative AI booked zero measurable return); some large enterprises are now reining in token usage rather than maximizing it, a turn from treating token-maximization as a KPI. On the other hand, OpenAI and Anthropic are seeing monstrous revenue growth, and anyone who has tried their paid plans understands why: the productivity at the personal level is very real.

None of this says the boom ends tomorrow. It says the boom is being funded, at the margin, by faith and debt — and faith and debt are the first things to evaporate when growth merely slows. The eventual state of AI revolution almost certainly a huge net positive, but we can't foresee how curvy the path can be.

Never cross a river 4-feet deep “on average”. -- common sense

Luckily, "in pure investment terms, there’s no intrinsic reason for long-term investors to be concerned with volatility (as distinguished from the risk of permanent loss)." So let's focus on the end game: what does the current valuation price-in?

Cash flow simulation

With the above background of the industry, we can model the future cashflow the business can generate, based on explicit assumption on every moving parts, then discount those cashflow to today's value. We can then compare it with today's market cap.

DCF is extremely sensitive to input. Therefore, more often than not, it can't give us a conclusive result unless the "safety margin" is too high. It's still very useful just too rule out very obvious situations, though, like when a company is obviously ove-rvalued or under-valued.

The model follows.

The three acts. SanDisk's future isn't one regime; it's three, in sequence — and each gets its own revenue growth and average gross margin:

1. The big shortage (now) — demand outruns locked supply, price spikes, margins sit at the all-time peak. Short, violent, lucrative; the only real question is how long it lasts.
2. AI growth with stable supply — new fabs finally land (~2027–28); real AI demand keeps volume climbing, but price resumes its deflation, so the margin falls back toward normal even as the business grows.
3. Post-buildout — the steady state on the far side, valued in perpetuity: whatever SanDisk settles into once the AI buildout matures.

Turning revenue and margin into cash. A gross margin is not free cash flow. Here is the bridge, and it is deliberately simple:

The gross margin already carries the fab's depreciation — the single largest cost — because for a chipmaker that depreciation sits inside cost of goods. From the gross profit we subtract operating costs (R&D and SG&A) as a percent of revenue, then tax (15%). Two assumptions make the depreciation and capital-spending terms vanish:

• Capex ≈ depreciation. We assume SanDisk spends just enough to maintain its capacity, so the capital going out roughly equals the depreciation already inside the margin, and the two cancel. This is not a technicality — it is the supply-discipline bet itself. The moment SanDisk or its rivals start building new fabs, capex jumps above depreciation, this cash flow is overstated, and the fresh supply ends the shortage. You cannot buy the expansion without buying the glut. In the long run, I belive it's a good first-order approximation.
• Operating-cost margin moves across the acts. It has fierce operating leverage — it falls toward 7–8% of revenue at the price peak (the engineers and sales staff cost about the same whether a gigabyte sells for 3 cents or 30) and climbs back toward the mid-teens once the spike unwinds. So you set it per act — low in the shortage, higher in the steady state — rather than pretending one number fits all three. Tax and the discount rate are dials too (defaults: 15% and 10%).

(Working capital — the cash a revenue ramp ties up in inventory and receivables, which is exactly why SanDisk's reported free cash flow today sits below this formula — is treated as neutral; it washes out over a multi-year horizon.)

If you play with the simulator long enough, you probably have found a input combo that justified your valuation, which ever it is. For example, the following input outputs the fair market cap of $269B, pretty close to today's real price.

• Starting annual revenue: 30B
• Act 1: 2 years, revenue growth 60%, gross margin 70%, opex margin 8%
• Act 2: 3 years, revenue growth 40%, gross margin 40%, opex margin 20%
• Act 3: growth 2.5%, gross margin 30%, opex margin 20%
• Discount rate 10%, tax rate 15%

The conclusion is not that SNDK is fairly valued. It is that it's possible that the valuation is right, even some combo of "strong" opinions that are qualitative.

Can enterprise SSD decommoditize?

This is the most critical question because no matter how you tune the model, majority of the value is from Act 3. Many criticize DCF for this reason. But I think this is a feature: what matters most for long term investors is the survival of the business, or the moat. That's why if SSD stay being a commodity, SNDK can't be valued by P/E. The consensus is valuing such companies using P/B ratio.

If we follow history literally, decommoditization seems long shot. But two companies are up against it.

A decade ago NVIDIA sold graphics cards — fast ones, but fundamentally merchant silicon, the kind of part that competes on price and gets commoditized on schedule. It is not that today. CUDA, the software lock-in, the full-stack systems, a cadence no rival can match turned a chip into a franchise earning margins competitors have spent years failing to crack. TSMC was a contract manufacturer — the textbook commodity business, building what others designed for whoever paid. At the leading edge today it is a near-monopoly with real pricing power, because three nanometers turned out to be something almost no one else on Earth can do. Both sold commodities once. Neither does now. Forever is not automatically a fantasy — sometimes a company climbs out of the commodity and the high margin really does become the floor. What Apple's mobile revoluation gave TSMC, AI revolution can bestow on the SSD industry. That future is not a impossibility.

The company's own answer

What's the management's reading of the situation? CEO David Goeckeler names the structural flaw directly: SanDisk is making decisions on investing on a 10-year time horizon^† while the market has "traditionally" priced "the economics … on a quarterly basis" — a supplier with decade-long horizons paid at quarterly spot rates, which he calls the root cause of every NAND downturn. The stated mission now is to convert that into durable, structurally higher earnings^† and a less cyclical business — and management is pursuing it on three distinct fronts, in rising order of ambition:

• Manufacture the durability by contract — the NBMs. Prong one is the long-term agreements above: lock volume, price floors, and prepayments so the fab never again runs blind into a wall. This attacks the cyclicality, not the commodity.
• Shift the mix to where the bit is least commoditized — enterprise. Enterprise SSDs now make up 25% of revenue^†, up sharply and at structurally fatter margins, and SanDisk unveiled a 256TB UltraQLC SSD^†, among the largest in the world, aimed at AI "data lakes." This is climbing the value ladder within the existing product.
• Escape the bit entirely — HBF. The genuinely NVIDIA-shaped move: High-Bandwidth Flash, a new memory tier sitting between expensive HBM and the ordinary SSD, which SanDisk is co-standardizing with SK hynix^† for AI inference, with first samples due in the second half of 2026^†. The pitch: HBF offers up to 8–16× the capacity^† of HBM at similar cost — a large new market the fungible bit never could reach. A standard you co-own is, by definition, not a commodity — this is the bid to leave the fungible bit behind.

So the question is not "will the cycle turn" — it will. It is: can enterprise storage decommoditize the way compute and foundry did — and can SanDisk, from last place, be the one that captures it? The tools are on the table: a qualification moat that takes years to clear, the NBM contracts that try to manufacture the durability the bit itself won't supply, AI-specific co-design that could make a drive something more than a pile of fungible gigabytes. Maybe that bends the curve. Or maybe the field is too crowded and the bit too fungible, and storage stays the thing that always, eventually, gets cheaper — and SanDisk, the smallest of five, is the wrong horse to bet on it anyway.

We are not going to answer that for you, and you should distrust anyone who claims to. What this piece gives you is the frame: a knowable floor — a real shortage, real volume growth, a contracted core that reaches years out — and a wide unknowable above it: the terminal margin, the share, whether the bit ever escapes its own nature. The simulator hands you the dials. Put in your read of how durable the AI build is, how far SanDisk can climb, what margin survives the crowd — and read off the company the price is asking you to underwrite. Then decide whether SNDK – or SSD more broadly – is the right investment at this point.

What to watch

1. The first big greenfield capacity commitment. Watch Samsung, SK Hynix, Micron, and the Kioxia–SanDisk JV. The move that ends the supercycle won't be a price print; it'll be a capex announcement. A large new-fab commitment starts the ~3-year clock toward the next glut. Continued restraint says the discipline — the first thing that has ever let this industry act like the oligopoly it is — is holding.

2. How fast the NBM book grows. This is decommoditization-by-contract, measured directly. Push the locked share past a third — half, two-thirds — and the durable core is becoming real. If the signings stall, the spot cycle still owns the story.

3. The NAND contract price index. Because SanDisk's earnings are now almost pure price-leverage on a fixed cost base, the contract price is the earnings. Whether it keeps climbing or resumes its historical slide is the single cleanest tell on whether the bit is escaping its nature — in real time.

4. The AI buildout progress. Everything here is built on AI's revolution. If that's falsified, SNDK will lose 99% of its value. But so are many others.

This research is done on 06/06/2026, the market value of SNDK is $231B. When you read this, it's fun to see how the reality developed since then.