Price, not bits: one memory maker's revenue rose 251% while the exabytes it shipped stayed flat

Every figure below was read from the filed document on sec.gov and is cited with its accession. Word counts — "the word appears zero times" — come from a case-insensitive search over the raw filing HTML, so a hit buried in markup would still be caught. Phrase counts across all filers come from EDGAR full-text search, which indexes filing documents from 2001 onward; it counts documents, not filings, and covers neither earnings-call audio nor image-based exhibits. 13F figures were read out of each filing's information table XML and reconciled to the cover-page total. Nothing is summed across filers. Arithmetic we did on filed values is labelled as ours. Research, not investment advice.

Memory led every trend scan we ran this week — Micron ranked first on September 18 and again on September 23, with SK Hynix, Sandisk and DRAM-as-a-topic filling out most of the rest of the top ten. The drivers our scan recorded were "HBM ASP hikes projected for 2027", "memory long-term agreements out to 2031", and Intel's CEO quoted on memory prices up five to seven times.

Three of those four cannot be checked against a filing. That is not a complaint about the feed — it is the reason this post exists. Because the thing the filings do say, and say in a specific place and a specific format, is the thing the headlines flatten. A revenue number is a product of two factors. The SEC asks companies to disclose both. Almost nobody quotes both.

A revenue number is two numbers

Here is the decomposition that started the week, from Micron's 10-Q, accession 0000723125-26-000015, for the quarter ended May 28, 2026:

Sales of DRAM products increased 343%, primarily due to a low-260% range increase in average selling prices and a low-20% range increase in bit shipments.

Sales of NAND products increased 361%, primarily due to a mid-310% increase in average selling prices and a low-double-digit increase in bit shipments.

Read those twice. The DRAM business roughly quadrupled, and about nine-tenths of that came from price. Micron shipped somewhere in the region of twenty per cent more bits than a year earlier — more, certainly, but not a different order of magnitude. (Our arithmetic, on the midpoints of Micron's own ranges: 3.6 × 1.2 ≈ 4.3, against the 4.43 the filed 343% implies. The ranges are wide by design; the point is the shape, not the decimal.)

The revenue lines underneath, from Note 14 of the same filing, in millions: DRAM $31,328 against $7,071, NAND $9,943 against $2,155, total revenue $41,456 against $9,301. Gross margin 85% against 38%. Income before taxes $33,212 million against $2,113 million.

Those are extraordinary numbers, and they are not the interesting part. The interesting part is that Micron files why, in two components, and so does everybody else in this trade — and the components do not point the same way.

The cleanest version of the split is not Micron's

It is Sandisk's, and it is a single sentence. From the 10-Q, accession 0001628280-26-029401, for the quarter ended April 3, 2026:

Net revenue increased 251% in the three months ended April 3, 2026 from the comparable period in the prior year, primarily due to a 248% increase in average selling prices ("ASP") per gigabyte. The exabytes sold remained flat from the comparable period in the prior year.

Revenue up 251%. Volume: flat. There is no clearer statement of what this cycle is in any filing we read. Revenue $5,950 million against $1,695 million; gross margin 78.4% against 22.5%; income before taxes $4,107 million against a $1,901 million loss.

The end-market splits in the same section show where the bits actually moved, and it is not where the narrative puts them:

Datacenter is the only line where volume did real work. Edge and consumer shipped fewer bits at much higher prices, which is what an allocation-constrained market looks like from the inside — and why "demand is exploding" is an imprecise description of a business whose total exabytes did not move.

The annual filing, 10-K 0001628280-26-057406 for the fiscal year ended July 3, 2026, says the same in Sandisk's more cautious annual vocabulary: revenue $20,248 million against $7,355 million, "Total products sold increased by mid-teens percent on an exabyte basis", Datacenter "Revenue per gigabyte increased by almost 150%". (Caveat the filing supplies itself: fiscal 2026 was a 53-week year.)

And one number that says more about supply than any commentary: Sandisk's capital expenditure that year was $177 million, against $20,248 million of revenue. The filing says this was deliberate — "we maintained what we believe to be a conservative capital expenditure strategy for fiscal years 2026 and 2025" — with increased investment flagged only for fiscal 2027. The industry did not answer this price move with capacity.

The same period end, the opposite decomposition

Sandisk and Western Digital split in February 2025 and kept the same fiscal calendar, so their quarters end on the same day. That makes the comparison unusually clean, and it goes the other way.

Western Digital's 10-Q, accession 0001628280-26-029054, quarter ended April 3, 2026 — the same day as Sandisk's:

Net revenue increased by 45% for the three months ended April 3, 2026 from the comparable period in the prior year, driven by a 34% increase in exabytes sold and a 9% increase in average selling price per exabyte, both of which were driven by strong demand across all of our end markets.

Thirty-four points of volume, nine points of price. Sandisk, same day: zero points of volume, 248 points of price. Both are storage. Both are selling into AI data centres. One is in a shortage and one is in a boom, and the only way to tell them apart is to read the second number.

(One trap worth naming: Western Digital's FY2026 pre-tax income of $9,905 million (0001628280-26-057139) includes $6,498 million of non-cash mark on its retained Sandisk stake. Operating income was $4,453 million. Do not read the pre-tax line as an operating result.)

Committed, and the sentence printed next to it

The second thing filings separate is a promise from a projection, and Micron's 10-Q does it in a way that is easy to quote wrongly. The filing discloses remaining performance obligations of "approximately $5 billion, of which $422 million has been recognized as contract liabilities", against "not material" a year earlier — under take-or-pay agreements whose largest contracts carry "a ceiling price for existing products that approximates the market price in the second calendar quarter of 2026, and a floor price through the term of the agreement."

But the sentence directly before the $5 billion, in the same note, is the one nobody quotes:

Our remaining performance obligations disclosure is based on minimum committed volumes and minimum pricing and is not expected to be indicative of future revenue under these contracts. Agreements without fixed pricing or price bands are not included in the remaining performance obligations disclosure.

So the $5 billion is a floor, deliberately constructed as one: minimum volumes at minimum prices, with the market-priced agreements excluded entirely. It is not a backlog. Separately, in Liquidity and Capital Resources rather than in a note, Micron says it expects to receive "cash deposits and related financial commitments of $22 billion for agreements concluded to date", of which "approximately $18 billion … will be in the form of cash deposits." Expects. The amount that has actually landed as a contract liability is $422 million. We went through the take-or-pay structure in July and are not repeating it.

Sandisk files the same species of disclosure and it is an order of magnitude larger. Remaining performance obligations were $41.6 billion at April 3, 2026 and $59.8 billion at July 3, 2026 — a $18.2 billion increase in one quarter — "entirely related to NBMs", Sandisk's name for its long-term agreements. Contract liabilities $1,242 million. And a line Micron does not have: refund liabilities of $1,500 million, which are refundable customer security deposits, up from $126 million a year earlier.

Western Digital, same page of the same calendar, files this:

The transaction prices allocated to the Company's remaining performance obligations as of July 3, 2026 and June 27, 2025, were not material.

Western Digital has long-term agreements too. It discloses them only in risk factors, as a downside: they commit it to "predetermined or formula-based prices for the duration of the contract terms", so "we may be unable to take full advantage of periods of rising prices, increased demand, or supply shortages." Same instrument, opposite framing, because one company contracted at today's prices and one contracted earlier.

Who pays, in the buyers' own MD&A

This is the part that does not require inference, because the cost shows up in other people's income statements with their own explanations attached.

HPE's Q3 FY2026 10-Q (0001645590-26-000080, filed September 3, 2026) is the most direct statement in the set:

Server net revenue increased by $1.8 billion, or 35.3%, predominantly due to an increase in the average selling price. The increase in average selling price was primarily driven by commodity price increases, especially memory and SSDs.

Server revenue $6,766 million against $5,000 million, and the entire increase attributed to price, and the price attributed to memory. HPE also names the mechanism, which no other filer does:

We are affected by the worldwide shortage in memory components that began to impact the semiconductor industry throughout fiscal year 2026 … and expect such dynamics to continue in the medium term as memory supply constraints may persist until memory vendors transition greater production allocations towards high performance memory components required by AI workloads.

(A structural caveat: HPE merged Server, Hybrid Cloud and Financial Services into one Cloud & AI segment effective November 1, 2025, so "Server" is a revenue line now and there is no filed server-level operating margin. The segment margin is 17.0% against 7.0%.)

Dell's Q2 FY2027 10-Q (0001571996-26-000046, quarter ended July 31, 2026) uses the word "memory" exactly twice, both in one bullet: "current limitations in capacity from memory manufacturers, has resulted in global supply constraints and substantial inflation in memory component costs." Its Infrastructure Solutions Group reports AI-optimized servers $16,401 million against $8,208 million (+100%) and traditional servers and networking $10,531 million against $4,736 million (+122%) — and files a different driver for each. AI servers: "primarily driven by an increase in units sold". Traditional servers: "primarily driven by an increase in the average selling price". Units on one line, price on the other, in the same segment, in the same quarter.

Here the obvious story is not quite the filed one. Dell's gross margin expanded 260 basis points to 20.9%, and Dell attributes that to "disciplined pricing", not to component costs. Its memory language sits in business trends and guidance. Dell is out-earning the inflation, not absorbing it.

IBM's Q2 10-Q (0000051143-26-000078) uses "memory" exactly once, about what buyers did:

Many clients redirected spending toward servers, storage, and memory purchases to secure supply-constrained infrastructure ahead of expected price increases. We saw this dynamic firsthand, with Distributed Infrastructure revenue increasing 37 percent, our strongest quarter on record.

That sentence appears in IBM's explanation of a miss — deals that failed to close because budget moved elsewhere. Total Infrastructure revenue was $3,835 million, down 7.4%, with IBM Z down 42.0%.

And the counterexample, which belongs here for the same reason. Super Micro's FY2026 10-K (0001375365-26-000022) does report memory as a constraint "which affected the timing of certain of our product deliveries, as well as the pricing of these items" — but its gross margin fell, to 10.8% from 11.1%, and the filed cause is not memory: "primarily due to our strategy to offer competitive pricing to gain market share, change in product and customer mix, and higher manufacturing related expenses." One server maker's margin went up and blamed its own pricing; another's went down and blamed its own pricing. Neither blamed memory for the margin.

What all four do show, unambiguously, is a buying panic in the balance sheet:

Dell's production materials alone went from $6,696 million to $15,894 million in six months. Super Micro says it may "place non-cancellable inventory orders for certain product components in advance of our historical lead times, pay premiums and provide deposits to secure future supply and capacity." HPE says its "efforts to secure memory components and SSDs in advance … may further increase our inventory levels." Three companies, one behaviour. We traced the first version of this squeeze in June; what has changed is that it is now in the numbers rather than the risk factors.

The three largest buyers never say the word

Now the negative result, which is the one we did not expect.

Occurrences of "memory", case-insensitive, in each company's most recent 10-Q:

Zero is a real zero. We checked the same documents for "DRAM", "component cost" and "semiconductor" in case the idea was being expressed in other words, and found nothing there either. Oracle's is the most recent 10-Q of the five, filed September 11, so staleness does not explain it.

Microsoft's single mention is not about its own cost base at all — it is about its customers: "growth in Windows OEM as OEM partners continue to build inventory due to increasing memory pricing." And Microsoft's FY2026 10-K (0001193125-26-323660, filed July 29, 2026), which is more recent than that 10-Q, contains zero occurrences.

Amazon's two are the same clause in cautionary-language lists — "resource and supply volatility, including for memory chips" — and the finding there is the date. Amazon's 10-Q for the quarter ended June 30, 2025 (0001018724-25-000086) contains zero occurrences of "memory". The phrase first appears in the filing for the quarter ended March 31, 2026 and has been carried since. That is a datable change in a hyperscaler's own risk framing, and it is the only one of the five.

The shape is odd. Memory makers describe a historic price move. Server builders describe it landing on them and are pre-buying inventory at double and triple prior levels. The companies buying those servers — the reason the demand exists at all — do not mention it. We are not going to tell you what that means. We will say that if you hold a view about hyperscaler capex in 2027, the filed disclosure supporting the memory-cost half of it is currently one sentence about Windows OEM licences.

What the last downcycle looked like, from the same filer

No forecast is needed here, because the symmetric case is already written down, by the same company, in the same section of the same form. Micron's FY2023 10-K, accession 0000723125-23-000054, for the fiscal year ended August 31, 2023:

Sales of DRAM products decreased 51% primarily due to a high-40s percent range decline in average selling prices and decreases in bit shipments in the high-single-digit percent range.

Same sentence structure, same MD&A location, opposite sign. What it produced:

Cost of goods sold at 109% of revenue is the whole cycle in one ratio. Around it, the same filing records $1.83 billion of inventory write-downs, "$382 million" of "period costs from fabrication facility underutilization … due to wafer start reductions", a headcount reduction expected "to approach 15% by the end of calendar 2023", the "suspension of our 2023 bonus company-wide", and "cuts to 2023 executive salaries across the company". Capital expenditure fell from $12.07 billion to $7.68 billion. Micron's FY2026 guidance, in the current 10-Q, is "approximately $27 billion".

Two more things from that document belong in any file on this topic. First, the range disclosure, which is the single most useful sentence a memory investor can read:

Since 2017, annual percentage changes in DRAM average selling prices have ranged from approximately plus 35% to a minus high-40s percent range. Since 2017, annual percentage changes in NAND average selling prices have ranged from nearly flat to a minus low-50s percent range.

Second, a zero. "Take-or-pay" appears zero times in the FY2023 10-K. The filing says the opposite:

Due to volatile industry conditions, our customers are generally reluctant to enter into long-term, fixed-price purchase contracts.

Three years later the same filer discloses take-or-pay agreements with floor prices and expected deposits of $22 billion — and the current 10-Q dates it: remaining performance obligations were "not material" as of August 28, 2025. The entire contract structure appeared inside one fiscal year. Whether it survives the next turn is an open question the filers themselves flag: Micron's risk factors say "Our industry goes through cycles" and that AI demand "can change significantly"; Sandisk's say "delays in data center build-outs could result in excess inventory, underutilization of capacity, or other costs".

The claim we could not find anywhere, and the one we found in the wrong place

Back to the feed. The most-shared Intel item of the week was the CEO on memory prices rising five to sevenfold. We went looking for it in Intel's filed documents and it is not there.

What is there is more useful. Intel has escalated memory-shortage language through three consecutive periodic reports, and the escalation is datable. The FY2025 10-K (0000050863-26-000011, January 23, 2026): shortages "may further limit our ability to meet CCG and DCAI customer demand in 2026." The Q1 10-Q (0000050863-26-000079): the same construction, plus a new disclosure that Intel expects supply agreements that "may involve prepayments". The Q2 10-Q (0000050863-26-000157, July 24, 2026), where the hedge drops out:

In Q2 2026 and YTD 2026, market demand exceeded our available product supply due to capacity constraints at our factories and industry-wide supply constraints. We expect these industry-wide shortages of substrates, memory and other critical components to persist into next year, which may limit our ability to fully meet customer demand.

The risk bullet naming "current industry-wide substrate and memory shortages" is new in the Q2 filings — the Q1 10-Q and the FY2025 10-K carry the same bullet ending at "or shortages" with no memory clause. So the company has said, in writing, that it is short of memory and expects to stay short into 2027. The CEO has not been quoted on memory in any filed Intel document we could find, and Intel has never used the exact phrase "memory shortage" in a filing in its history.

That phrase turns out to be rare everywhere. Eighteen documents filed in 2026 contain it, and the filers are not the memory makers — they are the buyers: AMD in four separate filings, plus CDW, PC Connection, Diodes, Alpha & Omega Semiconductor, Newegg and a handful of small caps. And a phrase that used to be the standard vocabulary for exactly this situation, "DRAM pricing", appears in zero documents filed in 2026 against 209 since 2001. The industry stopped writing about DRAM pricing and started writing about memory as a supply constraint, which is a different claim.

One correction we had to make on ourselves. We expected SK Hynix to be unreadable, because Korean and Chinese memory makers do not report to the SEC. That is still true of Samsung Electronics (251 EDGAR filings, none periodic, most recent 2015), Kioxia (an ADR registered on Form F-6, no 20-F, no 6-K), Nanya (two filings, last 2017), Winbond (five filings, last 1998), and YMTC and CXMT (no EDGAR presence at all). It is no longer true of SK hynix, which listed ADSs on Nasdaq as SKHY in July 2026 — F-1 0001193125-26-280172, final prospectus 424B4 0001193125-26-299963 at US$149.00 per ADS for US$26,507,100,000 — and has filed twenty-one 6-Ks since. The largest memory maker outside Micron became a readable filer nine weeks ago.

Who is positioned, as disclosed

And that prospectus is where the last thread ties off. From the cover page of the SK hynix 424B4:

Baillie Gifford Overseas Limited, acting on behalf of a number of its and its affiliates' clients, investment funds managed by Coatue Management, L.L.C., and Situational Awareness Partners LP (in alphabetical order) (collectively, the "Cornerstone Investors") have, severally and not jointly, indicated an interest in purchasing up to an aggregate of US$7 billion of the ADSs offered in this offering at the initial public offering price…

An indication of interest, the prospectus says, and "not binding agreements or commitments to purchase." Now read the 13Fs for the quarter ended June 30, 2026, filed August 14, each reconciled to its own cover-page total.

Situational Awareness LP (0000935836-26-000418) reports 26 positions totalling $20,242,292,228, and two memory names are the majority of it: SanDisk 2,495,344 shares / $5,673,738,513 (28.03%) and Micron 4,828,786 shares / $5,573,819,392 (27.54%) — together $11,247,557,905, or 55.57%. Coatue Management (0000919574-26-005478) reports 211 positions worth $48,629,053,706 with Micron at 3,142,269 shares / $3,627,089,685 (7.46%) and no Sandisk line at all.

The March report from the same filer (0002045724-26-000008) is a different animal, and this is the part worth sitting with. It held 42 positions worth $13,676,657,577, of which 11 lines totalling $8,459,056,999 — 61.85% of reported value — were puts, on NVIDIA, Oracle, Broadcom, AMD and a semiconductor ETF among others. By June the book was 26 lines and 99.66% plain common. (Precision: March held Micron common of just 17,362 shares, but also $583,686,168 of Micron puts and $422,300,000 of Micron calls. The move is in the common line, not the whole exposure.)

Across the concentrated managers we checked, the two names are almost mutually exclusive. Appaloosa (0001656456-26-000003) holds Micron at $1,125,432,750 — 14.57% of a 27-line book, and no Sandisk. Whale Rock (0001172661-26-003733) holds Sandisk at $1,356,382,263 — 10.89%, plus Western Digital and Seagate, and no Micron. Tiger Global holds Seagate and nothing else from the complex; Third Point, Soroban and Baupost hold none of the four. Situational Awareness is the only manager in the set holding both.

And Lone Pine (0000919574-26-005485) plays it without touching a memory maker at all. Its 34 positions worth $16,357,760,275 include ASML, Applied Materials and Teradyne — plus a new Seagate line, 1,000,109 shares / $965,105,185 (5.90%), absent from every row of its own March table (0000919574-26-003439). Equipment, materials and hard drives; no DRAM, no NAND.

(Caveats that come with the form: a 13F is US-listed long equity and options only, filed 45 days after quarter end. It carries no shorts, no cash, no bonds, no non-US listings, and says nothing about what has happened since August 14. The name on the SK hynix prospectus, "Situational Awareness Partners LP", is not the same filed string as the 13F filer "Situational Awareness LP"; they are plainly related but we did not verify the legal relationship from a filed document.)

What the filings separate, and what the headlines merge

Price versus bits. Micron files DRAM +343% as a low-260% range price rise on a low-20% range bit rise. Sandisk files +251% on +248% price with exabytes "flat". Western Digital, the same day, files +45% on +34% exabytes and +9% price. Three storage businesses, three completely different engines, and a headline that says "memory is booming" captures none of it. MD&A gives the decomposition because the SEC asks for it.

Committed versus indicative. Micron's ~$5 billion of remaining performance obligations is filed with a sentence saying it is "not expected to be indicative of future revenue" and is built from minimum volumes at minimum prices. Its $22 billion of deposits is money it "expects to receive"; the amount booked is $422 million. Sandisk's $59.8 billion is a bigger number with $1,242 million of contract liabilities and $1,500 million of refundable deposits behind it. Western Digital's is "not material". Four different objects, all correctly called "contracts".

Who pays, and whether they say so. HPE attributes a 35.3% server revenue rise entirely to price and the price entirely to memory. Dell files the cost pressure in trends and guidance while attributing its margin expansion to its own pricing. Super Micro's margin fell and it blamed its own pricing too. IBM files clients pre-buying ahead of price increases. Alphabet, Meta and Oracle do not use the word.

What the same filer said last time. Cost of goods sold at 109% of revenue, $1.83 billion of write-downs, $382 million of idle-fab cost, a 15% headcount reduction and a suspended company bonus — all in one 10-K, from the company now reporting an 85% gross margin. We are not predicting the next turn. We are pointing out that the filer has already described one, in its own words, and still publishes the range: DRAM prices have moved between roughly +35% and −high-40s in a year since 2017.

Run it yourself

Start with the decomposition, not the headline. Micron's financials give you the standardized statements and /filings/MU takes you to the documents; the sentence you want is in MD&A under "Results of Operations" — search the page for "average selling prices" and read what sits next to it. Do the same for SNDK and WDC and you have the three engines side by side in ten minutes.

Then check who is on the other side. /holders/MU shows who reports owning Micron and /holders/MU/owns turns it around. Every concentration figure above reconciles to its 13F cover page to the dollar.

For the language questions, EDGAR full-text search takes an exact phrase, a form type and a date range, and a zero is information. Run "DRAM pricing" restricted to 2026 and see for yourself.

If you would rather have the filings inside an agent, the same search and statement tools are on the EvidInvest MCP server. It is free to start, credit packs begin at $10, and there is no subscription.

What to watch next is a wording change, not a number. Micron's FY2026 10-K is not filed yet — the last three landed in the first week of October — and it will be the first annual document written entirely inside this price regime. If the bit-shipment side of the decomposition starts doing the work, or if the take-or-pay language picks up the qualifiers the 2023 filing used about customer reluctance, the story changes before the revenue does. Our Thesis Monitor reads each new filing against what you already believed, which a quarterly headline cannot do.

Research, not investment advice.