SanDisk Is Up 3,585%. The Options Market Has Scheduled the Verdict for 2028

The best-performing large cap in America over the past eighteen months is not an AI lab, a GPU designer, or a datacenter REIT. It is SanDisk, the company that made the memory card in your old camera. It listed at $48.60 on February 24, 2025, spun off from Western Digital to broad indifference. On Friday it closed at $1,790.82.

That is 3,585% in eighteen months, 49% of it in the last five sessions, and at roughly $278 billion of market cap it is now a top-30 US company.

Why is it performing so well? And what can the options market tell us about SanDisk's prospects? Let's start with the thesis.

AI inference is a storage problem

Everyone knows AI runs on GPUs and HBM. That story is three years old and fully priced. The newer, less-priced story is what happens one tier down.

Here is the hierarchy every AI system lives inside. Fast memory is tiny and obscenely expensive. Big storage is cheap and far too slow. Every byte an AI system touches has to find a home somewhere on this ladder:

HBM gets the headlines because it is bolted to the GPU. But HBM is sized for compute, not for state, and modern AI systems generate state in absurd quantities. Count what actually has to live somewhere:

Two of these deserve a closer look, because they are the ones growing fastest.

Weights: the four-orders-of-magnitude decade

Model parameters map almost one-to-one to bytes. When models were 1.5 billion parameters, weights were a rounding error. They are not a rounding error anymore:

The one force pushing the other way arrived in 2024: fp8 and int4 quantization plus sparse mixture-of-experts cut the bytes actually served per parameter by 2 to 4x. The totals exploded anyway. A frontier-class model is still measured in terabytes, and no serving fleet holds one copy. It holds a copy per node, per region, per fine-tune, per experiment. The weights themselves have become a distribution problem that only flash is fast enough and cheap enough to solve.

KV cache: the state that ate the datacenter

The really fun one is the KV cache. Every token of context a transformer holds costs memory for the rest of the conversation, roughly half a megabyte per token for a 405B-class dense model. That sounds harmless until you multiply it by agents running million-token contexts and thousands of concurrent sessions. Try it:

Play with that for thirty seconds and you understand the entire trade. A handful of long-context sessions overwhelms the HBM on a node that costs as much as a house. The industry's answer is not "buy more HBM," because there is no more HBM. The answer is to tier the state: park cold conversations on flash, reload them when the user comes back. SanDisk management now sizes this single workload, KV cache offload, at 75 to 100 exabytes of potential 2027 demand. For scale, that is roughly a third of the entire industry's annual output, from a workload that barely existed two years ago.

Add the embeddings behind every RAG system, the checkpoints behind every training run, and the hundred-petabyte corpora that dataloaders hammer at random, and the conclusion writes itself: the marginal byte of AI infrastructure is increasingly a flash byte. Storage stopped being the boring aisle of the datacenter sometime in 2025. The market took a while to notice.

The shortage

Demand was only half of it. The 2022-23 memory bust was bad enough that every NAND maker cut capex and idled lines, so when AI demand arrived there was nothing spare to sell:

Flash fabs are not light switches. Contract NAND pricing is now forecast up 75 to 100%, and since a wafer costs the same to make at any selling price, most of that increase falls straight to margin. SanDisk says about two thirds of its June-quarter sequential growth came from price, not volume.

Implied vol across the storage complex shows how precisely the market has sorted this out:

Micron makes DRAM and HBM alongside NAND. Western Digital kept the hard drives. Seagate is drives with a NAND garnish. SanDisk is the only large-cap pure play on flash pricing, and it carries a 13 to 20 vol point premium over all of them, at every tenor, out to 2028. The purer the exposure, the wider the distribution.

Datacenter took over in four quarters

That premium is earned. Revenue by end market, every quarter since the spin-off:

Datacenter went from $213 million to $2,977 million in four quarters and is now a third of revenue, growing 103% sequentially. Consumer, the SD cards the company was named after, shrank 5%. Even the boring Edge segment, flash sold into PCs and phones, quadrupled on pricing alone.

Margins are where it stops looking like a memory company at all:

The contracts

Those contracts got a dollar figure on August 13, at SanDisk's Investor Day: $93.9 billion of contracted business under its "New Business Model" agreements. Multi-year hyperscaler deals, quantities detailed by month, fixed price floors, financial guarantees. Ten signed, five since April. The stock rose 49% that week.

Management's fiscal 2028-2030 targets off this base: mid-to-high-teens revenue growth, gross margins around 80%, free cash flow margins near 50%, a buyback authorization now at $15.5 billion. Toll-road economics, if you believe them.

The bull case for the contracts is that they break the destocking spiral that made every previous memory downturn worse. Customers with take-or-pay floors do not run down inventory on a whim, and a hyperscaler signing four years of volume is showing you its own internal demand forecast.

The bear case is polysilicon. After 2011, ten-year take-or-pay contracts got renegotiated, litigated, or simply walked once spot fell far enough below contract. Long-term agreements dampen cycles; they do not repeal them. What $93.9 billion really buys is a change of question, from "will demand hold" to "will these prices hold when supply arrives." A better fight. Still a fight.

HBF: flash on the GPU package

There is one more leg, further out, and it needs a paragraph of packaging to make sense.

HBM is fast for a packaging reason, not a silicon one. Take ordinary DRAM dies, thin them, stack eight or twelve high, drill thousands of vertical connections through the stack, and mount the whole thing millimeters from the GPU on a shared slab of silicon. Width does the work: the stack talks to the GPU over thousands of wires at once, which is how you get terabytes per second out of memory that is not individually fast. The cost is capacity. DRAM cells are large, stacks have height and heat limits, and you end up with 36 to 48 GB per stack of the most expensive memory ever mass-produced.

High Bandwidth Flash is the same packaging trick with NAND dies in the stack instead of DRAM. NAND stores an order of magnitude more bits per die, so the published spec lands at 512 GB per stack at 1.6 TB/s: HBM-class bandwidth, more than ten times HBM capacity.

Flash has two real handicaps as memory. Reads take microseconds instead of nanoseconds, and cells wear out if you write them constantly. Both handicaps miss the inference workload almost entirely. Weights are written once and read billions of times. KV cache reloads are big sequential streams, the one access pattern where flash latency hides behind bandwidth. HBF would be useless as general-purpose memory, and it is not aimed at general-purpose memory. It is aimed at the workload from the first section of this piece.

If it ships, eight stacks put about 4 TB next to one accelerator, room for an entire frontier model's weights on the package, where today's HBM holds a fraction of them. SanDisk co-authored the spec with SK hynix and published it through the Open Compute Project in early August. First samples are due in the second half of 2026, first inference devices in early 2027.

None of it is in current earnings, and none of it should be modeled as revenue. It is a call option stapled to the stock: right partner, open standard, right workload. Paying something for it is reasonable.

Whose news moves whom

SanDisk, Micron, Western Digital, and Seagate now trade as one macro bet. SNDK and MU daily returns have correlated at 0.85 over the past 60 days. The options market prices them together too, but not identically, and the differences are where the information is.

The level difference is the purity premium from earlier: SNDK at 91%, the rest 13 to 20 points below, at every tenor. Same question, different leverage to the answer.

The more interesting difference is what each surface says about whose news matters. Micron reports on roughly September 22. None of the other three companies has any event that week. Their forward curves know anyway:

Micron's own report week prices 12 points over its neighbors, which is what an earnings week looks like. But Seagate, with nothing on its calendar, carries the same 12-point bump, and SanDisk carries five and a half. Only Western Digital, the one name with no NAND on the income statement, sleeps through it. In a shortage, the first supplier to disclose contract pricing moves everyone who sells flash.

Run the test backwards and it fails. SanDisk reports November 5, and its own forward for that window runs 96.7%, eleven points over base. Micron's forward through the same window: 71.2%, between a 70.0% stretch before and 69.0% after. SanDisk's biggest day of the year does not exist on Micron's surface. The market has decided information flows one way down this supply chain, from the diversified bellwether to the leveraged pure play, never back.

The leverage is quantifiable. Total variance across SNDK's expiries fits base-plus-events with about one vol point of error: 85.8% base, plus a 17.7% move for each of the next seven earnings reports. The same fit on Micron gives 70.6% and 4.2%. A Micron print is a data point. A SanDisk print is a referendum, priced at roughly double anything SanDisk has actually delivered:

We do not think that repricing is crazy. Under the NBM model, earnings day became the day the contracted book gets marked in public: new signings, new floors, new customers. November 5 is the first print against both the $10.3 to $10.8 billion guide and the Investor Day targets.

The far end of the curve makes the same point on a longer clock:

The top of the entire curve, 99.4%, sits in the first half of 2028, and stays at 91.3% even after stripping out both earnings reports inside that window. That is not an earnings hump. It lines up with the industry's own supply calendar:

Until late 2027 the shortage is unfalsifiable: no quarter can prove the bears right while the fabs that could oversupply the market do not exist. The first real test of the $93.9 billion book against new capacity comes in 2028, and the surface has parked its maximum uncertainty exactly there.

In dollars, here is the whole piece on one time axis: the realized path in, the implied distribution out:

A range of $509 to $6,361 by mid-2028 is the market holding both endings live: the commodity cycle in a party hat, and the structurally-80%-margin infrastructure company. The 25-delta risk reversal is inverted, calls over puts by 3 to 4 points at every expiry, so of the two tails, the market pays up for the melt-up. Meanwhile the equity trades near 10 times annualized guided earnings, a multiple that says "temporary." The stock and its own options are having an argument. We would rather own the argument than either side of it.

The bottom line

AI made storage a first-order input. That demand hit the one commodity whose supply had spent three years shrinking, at the purest-play supplier, which converted the squeeze into four years of contracts and margins no memory company has printed before. Nobody finds out whether it lasts until supply returns in 2028. Until then the path itself is priced: seven referendums at 17.7% each, a checkpoint every time Micron speaks, and a final exam in the first half of 2028.