Semiconductor memory. MU / SNDK. Run 14 September 2026. Built from Micron's FY2023 to FY2025 10-Ks and its Q3 FY2026 10-Q, Sandisk's 10-Ks for the years ended 27 June 2025 and 3 July 2026, both companies' earnings calls, conference appearances and investor days through September 2026, company press releases, SEC 8-K filings, and TrendForce share and pricing data. Events swept through 14 September 2026; most recent event checked: Kioxia and Sandisk announce over $31bn of Japan investment through 2032 and Kioxia begins Kitakami Fab3 site preparation, 27 August 2026. Micron has not yet reported Q4 FY2026, which is due 30 September 2026. Figures in US dollars, both companies' reporting currency, so no FX conversion applies. Not a valuation and not a recommendation.
Three cells carry Micron's next five years and each has a structural driver rather than a price tailwind. Data centre DRAM and HBM is the largest: HBM4 12-high has been in high-volume production for NVIDIA's Vera Rubin since 16 March 2026, with over $1bn already shipped and a ramp management says is running twice as fast as HBM3E. The driver is the trade ratio, not the price. To produce 100 bits of HBM, Micron gives up 300 bits of conventional DDR supply, and that ratio worsens toward 4:1 at HBM4E (Sadana, 10 August 2026), so Micron's fastest-growing product mechanically tightens the market for everything else it sells. Data centre NAND is the second and the one that matters for this duel: Micron's data-centre SSD revenue exceeded $5bn in the third quarter of fiscal 2026 alone and more than doubled sequentially, which makes one product line inside Micron larger than the whole of Sandisk's Datacenter segment. Auto, industrial and embedded is the third, at $9,062m over nine months and a 64.0% operating margin.
Sandisk's counter-cells are genuine. Datacenter went from roughly 12% to 38% of its bits in a single year, TLC enterprise SSD and QLC Stargate are both qualified at major hyperscalers, and High Bandwidth Flash has taped out with first samples due in calendar 2027. But 74.6% of Sandisk's revenue still sits in Edge and Consumer, where its own disclosure shows Consumer bits fell mid-teens and Edge bits grew only high single digits. The revenue growth in both was price, not volume, against a PC and mobile unit base declining mid-teens. Sandisk's growth is one cell deep, and the counterparty in that cell is Micron plus Samsung, SK hynix and Kioxia.
This is the most valuable finding in the run, because the growth winner and the margin winner are different companies. At the aligned cycle peak the two are effectively tied on the income statement: gross margin of 72.6% for Micron against 71.5% for Sandisk. All of Micron's 4.3-point EBIT edge comes from spreading operating expense over 4.5 times the revenue, R&D at 5.3% against 6.6% and SG&A at 1.6% against 3.3%, not from any advantage in what it costs to make a bit.
The divergence is entirely below the operating line. Micron spent 35.0% of revenue on net capital expenditure in fiscal 2025, about 21% in fiscal 2026, and guides fiscal 2027 to "above $45 billion even", more than half of it construction that its own operations chief says will not produce bits before calendar 2028, while its chief executive states that blended DRAM cost per bit will rise from here. Sandisk spends about 6% of revenue, grows bits through nodal transitions on an array base it already owns, and its wafer starts have been falling. Sandisk also quantifies its downside where Micron does not: about 80% gross margin at floor pricing on roughly two-thirds of fiscal 2028 bits, against Micron's unquantified claim that floor margins run "well above" a prior peak that was 59% in fiscal 2018. In one line: Micron is the compounder with the heavy engine, and Sandisk is the lighter engine in the weaker market. The caveat that stops this from flipping the call is that Sandisk's capital-lightness is partly a function of owning 49.9% of its fabs rather than 100%. The fixed costs are shared, not small.
Sandisk's most damaging exposure is that it does not control its own production. Every flash wafer it sells comes from Flash Ventures, which it owns 49.9% of, cannot unilaterally direct, is contractually barred from bypassing, and must fund half the fixed costs of regardless of the output it chooses to purchase. All three joint ventures co-terminate on 31 December 2034. In a demand shock Sandisk eats half of a fixed-cost base it does not control on the day demand falls, and 74.6% of its revenue sits in the most price-elastic cells in memory. It breaks first.
Micron's most damaging exposure is capital timing. More than $45bn of fiscal 2027 capital expenditure, the majority of it construction, is being committed against demand estimated during an acute shortage, into cleanrooms that produce bits from calendar 2028, and the resulting depreciation is fixed whether or not prices hold. The precedent is its own: fiscal 2023 took revenue to $15.5bn, gross margin to minus 9.1% and EBIT to minus 37.0% in four quarters. Micron breaks later and worse, because it carries 100% of its own fixed cost. Note the asymmetry that follows. In the last trough Sandisk's business held a gross margin of plus 16.1% while Micron printed minus 9.1%. The company with the weaker moat has the stronger floor, and that is the single best argument against this verdict.
These are the cells that carry essentially all of the combined end-market economics and about 95% of the geographic revenue. Every score is argued in the tabs below.
| Cell (product x region) | Micron | Sandisk | Why (sourced) |
|---|---|---|---|
| Data centre DRAM and HBM, global | 5 | 0 | Micron: HBM4 12-high in volume production for NVIDIA Vera Rubin since 16 Mar 2026, over $1bn HBM4 already shipped; 23.3% of 2Q26 industry DRAM revenue and the fastest grower of the big three. Sandisk: no DRAM product at all, and bought 3.9% of Nanya in March 2026 to secure DRAM supply rather than make it. |
| Data centre NAND and enterprise SSD, global | 4 | 4 | Micron: third-largest enterprise SSD supplier, share from 5 to 7 percent in 2022 to about 15 percent exiting 2025, first to volume on PCIe Gen6, data-centre SSD revenue above $5bn in Q3 FY2026. Sandisk: TLC eSSD and QLC Stargate both qualified at major hyperscalers, data centre from about 12% to 38% of bits in one year. |
| Mobile and client / Edge, global | 4 | 3 | Micron: 75.5% operating margin on MCBU over nine months, and it is deliberately starving the cell to feed data centre, so bits fell. Sandisk: 60.1% of revenue sits here, but bits grew only high single digits against a mid-teens decline in PC and mobile units. |
| Auto, industrial and embedded, global | 4 | 2 | Micron: $9,062m over nine months at a 64.0% operating margin, reported and managed as its own business unit, with L2+ vehicles carrying over five times average memory content. Sandisk: folded inside Edge, size not disclosed, and sold without a DRAM attach. |
| Consumer retail, global | 0 | 4 | Micron: exited. The Crucial wind-down was announced 3 Dec 2025 and completed in Feb 2026, with supply redirected to data centre and enterprise. Sandisk: the only branded retail franchise in memory, $2,935m and up 29% on price even as its bits fell mid-teens. |
| Greater China, ship-to | 1 | 4 | Micron: barred by China's CAC since May 2023 from selling to critical information infrastructure operators, disclosed as having an adverse impact on its ability to compete; 10.1% of FY2025 revenue by customer headquarters. Sandisk: China plus Hong Kong was $9,629m, or 47.6% of FY2026 revenue. |
| United States, supply origin and demand | 5 | 2 | Micron: 64.5% of FY2025 revenue by customer headquarters, a US investment plan raised to over $250bn through 2035, a goal of making 40% of its DRAM in the US, and a US-supply premium written into its customer agreements. Sandisk: $3,714m ship-to, 18.3% of revenue, and 100% of its wafers made in Japan. |
Scores are per cell and are not summed. The three lenses get their verdicts in the Three Answers above, not here. The two geography rows use different attribution bases, Micron by customer headquarters and Sandisk by ship-to location, and so are scored on presence and control rather than on a like-for-like revenue comparison.
Normalisation first, because the two companies do not use the same words. Micron reports four business units. Sandisk reports three end markets, renamed in fiscal 2026 from Cloud, Client and Consumer to Datacenter, Edge and Consumer. Sandisk's Edge contains automotive and industrial, which Micron reports separately as AEBU, so the like-for-like pairing is Micron's MCBU plus AEBU against Sandisk's Edge. Micron has no consumer cell at all since the Crucial wind-down completed in February 2026. On geography the definitions genuinely do not reconcile: Micron attributes revenue by customer headquarters, Sandisk by ship-to location. Micron's own 10-K bridges it usefully, stating that about 80% of fiscal 2025 revenue was shipped outside the United States, which puts it near Sandisk's 18.3% US ship-to share. Micron does not disclose geography in its 10-Qs, so its latest geographic data is fiscal 2025 against Sandisk's fiscal 2026, and that limit is flagged wherever it is used.
The aligned window throughout is Micron's last twelve months to 28 May 2026 against Sandisk's fiscal 2026 to 3 July 2026, an offset of about five weeks. On that basis Micron earned $90,274m and Sandisk $20,248m. Micron has not yet reported its fourth quarter of fiscal 2026, which is due on 30 September 2026 and guided to $50bn of revenue at about an 86% gross margin, so this comparison is against a Micron that is still accelerating.
What the cell table shows is a single dominant contrast. Micron is a data-centre company that also sells to phones and cars. Sandisk is an edge and consumer company that is becoming a data-centre company quickly. Both statements are true and only one of them describes a durable position, because the data-centre cell contains DRAM and HBM, which Sandisk has no product for at all.
| Cell | Micron revenue (% of LTM) | Growth | Sandisk revenue (% of FY26) | Growth | Margin signal and leader |
|---|---|---|---|---|---|
| Data centre, all memory | $52,512m (58.2%) | +217% | $5,153m (25.4%) | +437% | Micron. CMBU 70.2% and CDBU 72.6% operating margin over nine months; Sandisk reports one segment and no cell margin (ND) |
| of which data centre SSD / NAND | over $5,000m in Q3 FY26 alone | more than doubled q/q | $2,977m in Q4 FY26 | +103% q/q | Micron on size, Sandisk on trajectory. One Micron product line out-earns Sandisk's whole Datacenter segment |
| Mobile and client / Edge | $27,247m (30.2%) | +190%, bits DOWN | $12,160m (60.1%) | +195%, bits +high single digit | Micron. 75.5% operating margin and only 30% revenue dependence, against Sandisk's 60% |
| Auto, industrial and embedded | $10,496m (11.6%) | +173% | ND, inside Edge | ND | Micron. 64.0% operating margin, separately disclosed and managed |
| Consumer retail | $0, exited Feb 2026 | n/a | $2,935m (14.5%) | +29%, bits −mid-teens | Sandisk. The only branded retail franchise in memory; Micron judged the bits better used elsewhere |
| Greater China, ship-to | ND ship-to; 10.1% of FY2025 by customer HQ | ND | $9,629m (47.6%) | China +121%, HK +294% | Sandisk on presence. Micron is barred from China's critical infrastructure operators |
| United States | $24,113m FY2025 by customer HQ (64.5%); about 20% ship-to | ND | $3,714m ship-to (18.3%) | +157% | Micron. US demand and US fabs, with a supply premium written into contracts |
Data centre DRAM and HBM, Micron only. The driver is the trade ratio, which is a supply mechanism rather than a demand forecast: every HBM bit sold removes three conventional DDR bits from supply, worsening toward four to one at HBM4E. The HBM addressable market is now expected to cross $100bn in calendar 2027, a year earlier than Micron previously said. Durability is high because HBM is qualification-gated and moving toward custom single-sourced SKUs. The cost to grow is severe: fiscal 2026 net capex of about $27bn and fiscal 2027 above $45bn. All organic; the $1.8bn PSMC Tongluo fab closed in March 2026 buys capacity, not revenue.
Data centre NAND and enterprise SSD, contested. Micron has moved from 5 to 7 percent enterprise SSD share in 2022 to about 15% exiting 2025, is first to volume on PCIe Gen6, is lead-qualified on NVIDIA's STX platform and leads QLC at 245TB. Sandisk has both TLC enterprise SSD and QLC Stargate qualified at major hyperscalers and has demonstrated a 256TB drive in E3 form factor. Durability for Sandisk is medium: this is a five-supplier market with no HBM-style qualification lock, where calendar 2Q26 NAND revenue share ran Samsung 29.3%, SK hynix about 18.2%, Micron about 15.1%, Kioxia 13.6% and SanDisk about 11.4%. Micron moved up to third, passing Kioxia, on the fastest sequential growth of the top five.
Edge and mobile, both. Micron's MCBU revenue rose 190% over nine months while its bit shipments fell, because supply was deliberately redirected to higher-value segments. Sandisk's Edge rose 195% with bits up only high single digits, against a PC and mobile unit base declining mid-teens, all of it at the low end. Both cells grew on price alone. As a growth engine, durability is low for either.
Consumer retail, Sandisk only. Micron announced the exit from Crucial on 3 December 2025 and completed the wind-down in February 2026, redirecting the supply to data centre and enterprise. Sandisk kept its franchise, growing revenue 29% on price while its bits fell mid-teens, and describes its global consumer presence as a meaningful differentiator. Its chief executive also concedes the mechanism that limits it: in markets moving this fast, consumer pricing cannot be raised as quickly as the transactional ones.
Micron sells at fixed, negotiated, generally short-term prices and offers price protection to its distributors, with the amount undisclosed. Over that it has now layered sixteen strategic customer agreements: take-or-pay, typically five years from calendar 2026 to 2030, with a ceiling at the calendar Q2 2026 market price and a floor through the term, covering about 20% of DRAM volume and a third of NAND volume, or close to 25% of revenue. Price is renegotiated quarterly inside the band and can move neither above the ceiling nor below the floor. Crucially, there is no walk-away provision. Micron holds $22bn of commitments, about $18bn of it unrestricted cash deposits.
Sandisk publishes the thing Micron does not: its sales incentive and price-protection programmes ran at 11% of gross revenue in fiscal 2026, down from 19% in each of fiscal 2025 and fiscal 2024. That is a direct, disclosed measure of how much list price the channel claws back, and it halved. Over that sit eight New Business Model agreements with a weighted average duration above four years, covering more than 50% of fiscal 2027 bits and about two-thirds of fiscal 2028 bits, with fixed and variable elements subject to floors and ceilings. But they are exit-able. The chief executive describes the mechanic plainly: a customer forgoes billions held by a third party and the contract ends. The pool is $16.5bn, of which only $2.9bn is deposits and credits and only $2.5bn is actually in Sandisk's bank.
This is where the two businesses stop resembling each other. Micron owns its fabs in the United States, Japan, Taiwan and Singapore, carrying more than $46.7bn of long-lived assets. Sandisk owns $674m of long-lived assets in total and no wafer fab at all: 100% of its flash wafers come from Flash Ventures, three joint ventures with Kioxia across eight Japanese fabs, in which Sandisk holds 49.9%. It cannot unilaterally direct most of the venture's activities, is contractually restricted from having flash fabricated anywhere else, must pay half of the venture's fixed costs regardless of the output it takes, guarantees half of the venture's sale-leaseback obligations, and buys its wafers in yen while selling substantially in dollars. All three ventures co-terminate on 31 December 2034. Sandisk's answer to its DRAM gap was to buy about 3.9% of Nanya for roughly $1.0bn in March 2026 with a multi-year supply arrangement attached, which secures access rather than capability.
In calendar 2Q26 NAND revenue share ran Samsung 29.3%, SK hynix group about 18.2%, Micron about 15.1%, Kioxia 13.6% and SanDisk about 11.4%, on a market of roughly $78.8bn. Only Samsung's and Kioxia's shares were published explicitly; the others are implied from the same revenue base. Micron grew 99.2% sequentially, the fastest of the top five, and SanDisk 50.7%, the slowest. In DRAM, Samsung held 39.4%, SK hynix 24.9% and Micron 23.3%, with Micron growing fastest of the three and sitting 1.6 points behind second place. Sandisk has no DRAM presence. CXMT grew 99.3% sequentially but TrendForce published no share for it; its December 2025 listing prospectus put it at roughly 4% of global DRAM.
In data centre DRAM and HBM the risk is Micron's alone, and it is that a missed generation forfeits a socket for years in a market moving toward single-sourced custom SKUs: the same mechanism that is the moat is also the concentration. In data centre NAND the more exposed party is Sandisk, because that cell carries its entire growth case while its competitor there can cross-subsidise from DRAM. In Edge and mobile the more exposed party is Sandisk again, at 60.1% of revenue against Micron's 30.2%, in a cell whose units are falling. In consumer the exposure is Sandisk's by definition, since Micron exited. In Greater China the asymmetry cuts both ways: Micron has been barred since May 2023 from selling to China's critical information infrastructure operators, disclosed as adversely affecting its ability to compete, while Sandisk ships 47.6% of revenue to China and Hong Kong, which is both its largest geographic dependency and the market most exposed to domestic substitution by CXMT and YMTC.
HBM sockets. Durability: high. HBM4 12-high in high-volume production for NVIDIA Vera Rubin since 16 March 2026, over $1bn already shipped, ramping twice as fast as HBM3E. Management expects HBM4E to bring custom SKUs and a market that is "a two-supplier market, and oftentimes, just a single-sourced opportunity". A well-funded rival needs years of co-design and qualification, not money.
Process and capacity. High on know-how, medium on capacity. First EUV DRAM node and G9 NAND both ramping, described as the highest-volume nodes in Micron's history, on fabs in four countries carrying over $46.7bn of long-lived assets.
Supply origin. Medium-high. Over $250bn of US investment through 2035, a goal of 40% of DRAM made in the US, up to $6.4bn of CHIPS grants and a 35% investment tax credit, converted into price through a US-supply premium written into the customer agreements. Policy-dependent, but policy is moving toward it.
Not a moat: Micron takes prices. Its returns are set by aggregate industry supply behaviour, of which it controls about 23%.
Shared fab access. High as an asset, but not controlled. Three joint ventures, eight Japanese fabs, roughly 80% of Kioxia's owned capacity; together the two produce about a third of the world's NAND. Extended to 31 December 2034 in January 2026 for $1,165m payable through 2029. The barrier is real, as its chief executive puts it: duplicate Yokkaichi and "find a lot of money and I'll see you in about 10 years". But Sandisk owns 49.9% of it. See the supplier section.
Capital efficiency. Medium. The claim is that the industry spent 2.66 times more capital per unit of output in 2025, and that over five years the pair produced 29% of industry bits on 13% of the capital. Note that this is measured for Sandisk and Kioxia combined, not Sandisk standalone (inferred: the standalone figure is less flattering, since the 6% capital intensity excludes the venture's leased and subsidy-funded equipment). It is a roadmap advantage, replicable in principle by any supplier willing to trade growth for capital.
Consumer brand. Medium. The only branded retail franchise in memory; it held price on a declining volume base, up 29% on revenue with bits down mid-teens. Real, but it sits in a 14.5% cell.
High Bandwidth Flash. Unproven. First die taped out, samples to inference-device customers in calendar 2027, first OCP specification published with SK hynix on 3 August 2026 with Google and Tenstorrent in the consortium. Sandisk's own model excludes HBF revenue entirely.
Sandisk has the better book. No single customer exceeded 10% of net revenue in any of fiscal 2024, 2025 or 2026, and its top ten were 44% of revenue in fiscal 2026. Micron discloses that approximately one half of its total revenue came from its top ten customers in each of the last three years, that approximately one half of revenue is concentrated in the data-centre end market, and that one customer was 17% of fiscal 2025 revenue. Micron's receivables also rose to $31.0bn at 28 May 2026 from $9.3bn at fiscal 2025 year end, a large credit concentration by any standard, against Sandisk's $4.7bn with three customers at 19%, 12% and 10% of the balance.
Micron has the harder paper and the bigger backstop. Its agreements cannot be cancelled and it holds about $18bn of unrestricted customer cash. Sandisk's can be exited for a fee, and only $2.5bn of its $16.5bn guarantee pool is actually on its balance sheet, the rest sitting with third-party institutions. Against that, Sandisk is the only one of the two to have quantified its floor economics, at about 80% gross margin on roughly two-thirds of fiscal 2028 bits, and confirmed the arithmetic publicly. Micron has twice declined to quantify its equivalent, saying only that floor margins run well above a prior peak that was 59% in fiscal 2018.
Micron, decisively, on control; Sandisk, on the evidence of the last downturn, on outcome. Micron makes its own wafers and can idle them, which is exactly what it did in fiscal 2023, printing a minus 9.1% gross margin and a minus 37.0% EBIT margin. Sandisk gets every wafer from a venture it owns 49.9% of, cannot direct, cannot bypass, must half-fund regardless of offtake, and which expires in 2034; its predecessor's equivalent down year produced a plus 16.1% gross margin, with underutilisation charges of $75m and inventory write-downs of $24m in fiscal 2025 rather than a negative margin. The 10-K also names the instability plainly: Kioxia's stakeholders "may include, or have included in the past, competitors, customers, a private equity firm, government entities or public stockholders", and Western Digital and Kioxia merger talks were reported on 15 July 2026 without primary-source confirmation.
| Cell | Route control | Pocket price | Continuity | Outcome | Confirming KPI |
|---|---|---|---|---|---|
| Data centre DRAM and HBM | Micron direct to hyperscalers and accelerator vendors. Sandisk absent | Micron above, on single and dual-sourced sockets with custom SKUs coming at HBM4E. Sandisk ND | Micron constrained by its own greenfield timing: ID.1 mid-CY2027, ID.2 late CY2028. Sandisk n/a | Micron share up, margin up. Uncontested, because Sandisk cannot enter | HBM4E ramping on schedule in CY2027; Micron closing the 1.6pp DRAM gap to SK hynix |
| Data centre NAND and enterprise SSD | Both direct. Sandisk has three US hyperscalers inside its eight NBMs; Micron is lead-qualified on NVIDIA's STX platform | Parity to Micron-above. Micron is first to volume on PCIe Gen6 and leads QLC at 245TB; Sandisk has TLC and QLC qualified and demonstrated 256TB in E3 | Micron can flex wafers between DRAM and NAND and is deliberately growing NAND slower than the industry. Sandisk cannot flex: its fab plan is fixed years ahead | Micron share up, margin flat to up. Sandisk share up, margin exposed. Sandisk grows into a cell where a larger rival can subsidise from DRAM | Micron quarterly data-centre SSD revenue against Sandisk Datacenter revenue |
| Mobile and client / Edge | Micron direct plus distributors with price protection, amount undisclosed. Sandisk OEM plus channel, give-back disclosed and halved to 11% of gross revenue | Micron above on disclosed margin, 75.5% operating margin over nine months. Sandisk ND at cell level | Micron is starving the cell on purpose and its bits fell. Sandisk depends on it for 60.1% of revenue | Micron share down by choice, margin up. Sandisk share up, margin exposed. Sandisk inherits volume Micron does not want | Sandisk Edge exabytes turning negative year on year |
The tension this exposes is worth stating plainly, because it usually runs the other way. Sandisk's two largest revenue cells, Edge at 60.1% and Consumer at 14.5%, are the two where it has the least pricing leverage and where unit demand is shrinking, while its strongest cell by trajectory is the one where Micron has both the deeper attach and the larger absolute business. Growth and power point the same direction here, which makes the call firmer than a four-and-a-half-times size gap alone would.
Product scope. Major, and not closable. DRAM was 77% of Micron's third-quarter fiscal 2026 revenue, with HBM in volume for Vera Rubin and an addressable market crossing $100bn in calendar 2027. Sandisk has no DRAM product and bought 3.9% of Nanya in March 2026 specifically to secure DRAM supply rather than to make it. To close the gap Sandisk would have to build or buy a DRAM franchise, which takes a decade, and its Flash Ventures agreements restrict what it can build.
Control of the production base. Major, and structural in both directions. Micron owns its fabs and therefore its capital bill: $27bn in fiscal 2026, above $45bn in fiscal 2027. Sandisk owns $674m of long-lived assets and spends about 6% of revenue. Neither can close this, and neither wants to: Micron's capex is the price of the HBM moat, and Sandisk's asset-lightness is the price of not having one.
Customer diversification. Moderate, and closable. Micron has one customer at 17% of fiscal 2025 revenue and about half of revenue in the data centre; Sandisk has none above 10% and 44% in the top ten. Micron is already closing it, expecting approximately half or more of company revenue under strategic customer agreements once all planned agreements execute, which is twelve to twenty-four months of work. Because it can be closed with money and time, it is a current condition for Sandisk, not a moat.
No reconciliation problem arises here. Both companies report by function under US GAAP with directly comparable captions, so all five ratios are clean for both and none is ND. Three mapping notes apply consistently across every year. First, neither company's exceptional operating lines are reclassified and EBIT is as-reported operating income for both; Sandisk's fiscal 2025 EBIT margin of minus 18.7% includes an $1,830m goodwill impairment, and excluding it the figure was plus 6.2%, so both are shown. Second, Sandisk separated from Western Digital in February 2025, so fiscal 2024 and part of fiscal 2025 are carve-out financials; fiscal 2023 was not obtained and is ND. Third, and most important, the two fiscal windows capture different parts of the same cycle.
That third point governs how this tab should be read. Micron's last three completed fiscal years end on 28 August 2025, before its price spike. Sandisk's last three end on 3 July 2026, fully inside it. A three-year average across those windows would say Sandisk earned a 51.8% aggregate gross margin against Micron's 24.4%, which is true and meaningless. The table below therefore uses calendar-aligned pairs instead, and the headline pair is Micron's last twelve months to 28 May 2026 against Sandisk's fiscal 2026 to 3 July 2026, an offset of about five weeks.
| % of sales | Micron, LTM to 28 May 26 | Sandisk, FY26 to 3 Jul 26 | Gap | What drives it |
|---|---|---|---|---|
| COGS | 27.4% | 28.5% | Micron 1.1pp lower | Near parity. Micron's DRAM mix is offset by Sandisk's cost-plus wafer pricing from the joint venture |
| Gross margin | 72.6% | 71.5% | Micron +1.1pp | Effectively a tie at the peak. Neither has a durable advantage in what it costs to make a bit |
| R&D | 5.3% | 6.6% | Micron 1.3pp lower | Scale. Micron spreads a larger absolute bill ($4,784m vs $1,328m) over 4.5 times the revenue |
| SG&A | 1.6% | 3.3% | Micron 1.7pp lower | Sandisk carries a retail and channel selling organisation that Micron exited in February 2026 |
| EBIT margin | 65.6% | 61.3% | Micron +4.3pp | Entirely operating leverage. The gross-margin edge is only 1.1pp of it |
| Window | Micron | Sandisk | Reading |
|---|---|---|---|
| Cycle recovery: Micron FY2025 against Sandisk FY2025, about two months offset | Gross margin 39.8%, EBIT 26.1% | Gross margin 30.1%, EBIT minus 18.7% (plus 6.2% excluding goodwill) | Micron converted the recovery and Sandisk did not. Micron's DRAM and HBM mix recovered first and further. |
| Cycle trough: Micron FY2023 against Sandisk FY2024, the nearest overlapping down years, not a clean pair because Micron's FY2023 is a pure trough while Sandisk's FY2024 already contains recovery months | Gross margin minus 9.1%, EBIT minus 37.0% | Gross margin 16.1%, EBIT minus 7.0% | Sandisk far better. Micron carries 100% of its own fixed cost and idled owned fabs; Sandisk shares half of Flash Ventures' fixed cost and takes underutilisation charges rather than a negative margin. |
| Micron | Sandisk | |
|---|---|---|
| FY2025 capital expenditure | About $13.1bn net on $37,378m revenue, or 35.0% | Not stated on a comparable basis |
| FY2026 capital expenditure | About $27bn net on roughly $129bn of revenue, about 20.9% (inferred: nine months actual plus the company's Q4 guidance midpoint) | Gross capex 6.3% of revenue in Q4, about 6% for the full year |
| FY2027 plan | "above $45 billion even", more than half of the increase from construction | About 6% of revenue, rising in dollars as BiCS8 and BiCS10 ramp |
| Long-lived assets | Over $46.7bn across four countries | $674m |
| Direction of unit cost | Rising. Blended DRAM cost per bit "to rise from current levels" | Falling. 27% bits-per-wafer compound growth; wafer starts have come down |
| When new capacity produces bits | Greenfield capacity "really starts to contribute to bits in calendar 2028" | Continuously, through nodal transitions on existing arrays |
Four caveats on Sandisk's figure, stated because they materially soften it. The 6% is gross capital expenditure on Sandisk's own books; Flash Ventures' equipment is separately funded through depreciation recovered in cost of revenue, sale-leaseback financing Sandisk guarantees half of, and government subsidies, a decomposition the company itself gave on its November 2025 call. The 2.66 times capital-efficiency claim is Sandisk and Kioxia combined, not Sandisk standalone. The $1,165m joint-venture extension payment is expensed through cost of revenue over nine years rather than capitalised, so it sits outside the 6% entirely. And management contradicted itself at its own investor day, with the chief financial officer saying "mid-single digits capital intensity" in the model and "low single digits" minutes later in question and answer.
The most persistent structural difference is not in any of the five income-statement ratios; it is in capital intensity, and it is present in every year of both records. At the cycle peak the two companies are within 1.1 percentage points on gross margin and within 4.3 points on EBIT margin, and the whole of Micron's EBIT edge comes from spreading operating expense over four and a half times the revenue rather than from any advantage in what it costs to make a bit. Below that line the businesses diverge completely: Micron spent 35.0% of revenue on net capital expenditure in fiscal 2025 and guides above $45bn for fiscal 2027, more than half of it construction that its own operations chief says will not produce bits before calendar 2028, while its chief executive states that blended DRAM cost per bit will rise from here; Sandisk spends about 6% of revenue and grows bits through nodal transitions on an array base it already owns, with wafer starts falling. Strategically this says Micron is buying a position and Sandisk is harvesting one. It confirms the power map in one direction, because Micron's capital bill is the price of the HBM moat, so the spending is a consequence of the advantage rather than a substitute for it. It contradicts the power map in the other, because Sandisk absorbed the last trough far better, at a plus 16.1% gross margin against Micron's minus 9.1%, which means the company with the weaker moat has the stronger floor. That lowers the confidence in any verdict resting on Micron outperforming through a full cycle rather than through this part of one.
| Metric | Threshold | By when | If it hits, it favours | Where published |
|---|---|---|---|---|
| Micron FY2027 net capex guidance | Above $50bn | 30 Sep 2026, Q4 FY2026 call | Sandisk | Micron earnings call, then FY2026 10-K |
| Micron DRAM average selling price, change q/q | First negative quarter | Any quarter from 30 Sep 2026 | Sandisk | Micron 10-Q, MD&A |
| Sandisk NBM coverage of FY2028 bits, and the guarantee pool | Reaches about two thirds as guided, with the $16.5bn pool growing alongside RPO | Q1 FY2027, about Nov 2026, and each quarter after | Micron if it falls short | Sandisk earnings call and 10-Q |
| Micron quarterly data-centre SSD revenue versus Sandisk quarterly Datacenter revenue | Micron above Sandisk for four consecutive quarters | Through Q4 FY2027 | Micron | Both companies' earnings calls |
| Sandisk full-year gross capital intensity | Above 9% of revenue | FY2027 results, about Aug 2027 | Micron | Sandisk 10-K |
| CXMT or YMTC qualified in leading-edge DRAM or HBM at a top-three non-Chinese customer | Any confirmed qualification | Any date | Neither. Damages both, Micron more | TrendForce, Counterpoint, customer disclosure |