Industry

Industry — NAND Flash Memory and the AI Storage Super-Cycle

Sandisk does not compete in "storage" broadly. It competes in NAND flash memory — the non-volatile semiconductor that stores data on every phone, SSD, and increasingly every AI server. The industry is a five-player oligopoly with a famously brutal cycle: oversupply crushes prices to below cash cost, demand recovers, supply tightens, and operating margins rocket from negative to over 30% within a few quarters. Three ideas to hold at once: (1) five firms set global ASPs, (2) the cycle is harder than almost any other branch of semiconductors, and (3) for the first time in its history, the dominant marginal customer is AI data centers paying for supply assurance rather than the lowest price.

1. What "NAND" is, and where Sandisk sits in the value chain

NAND flash is the chip that stores data without power. It comes in two physical forms — bare wafers (round silicon disks fabricated in clean rooms) and assembled products (memory cards, USB drives, embedded mobile storage, and solid-state drives or "SSDs"). The value chain runs left to right: a fab etches transistors onto wafers; wafers get cut into chips; chips are packaged with a controller and firmware into a finished device; the device is sold to an OEM, a data center, a retailer, or a consumer.

Sandisk is vertically integrated through wafer manufacture, package, and brand — but it does not own its fabs. Substantially all flash wafers come from Flash Ventures, a 49.9 percent / 50 percent joint venture with Japanese rival Kioxia operating seven fabs in Japan (an eighth begins production in CY2025). Sandisk and Kioxia each take roughly half the output at cost-plus-small-markup. Capacity, capex, and the technology roadmap are all shared with Kioxia — which is also a direct competitor in finished products.

2. Five companies own the world's NAND output

NAND is a tight oligopoly. Five vertically integrated suppliers — Samsung, SK Hynix (including the former Intel NAND unit, now branded Solidigm), Kioxia, Micron, and Sandisk — control roughly 95 percent of global NAND wafer shipments. There is no meaningful new entrant; greenfield NAND fabs cost USD 15-25 billion each and require process know-how built over decades. The Chinese entrant YMTC is subject to U.S. export controls and disconnected from the leading-edge tool ecosystem.

NAND pricing is set by the marginal bit. When one supplier brings a node online with too much capacity, ASPs across all five collapse. When the industry collectively cuts wafer starts — as it did through 2023 and into the first half of 2024 — ASPs reverse violently. There is no fragmented competitive fringe to dampen the move.

3. The cycle — and why memory is the most cyclical place in semis

Memory businesses oscillate harder than logic semis because the bit is undifferentiated. A 1-terabyte 232-layer 3D NAND die from Sandisk is functionally interchangeable with one from Micron. When supply exceeds demand, the only competitive lever is price, and price falls below cash cost until someone shuts down a fab. The chart below uses Micron — the longest-tenured public peer with directly comparable economics — to show what a complete memory cycle looks like.

The rhythm is unmistakable. The 2017-18 super-cycle took Micron's operating margin from 1 percent to 49 percent in two fiscal years, on the back of mobile NAND and the first hyperscale data-center demand wave. The 2023 trough was the worst memory downturn since 2008 — Micron lost USD 5.7 billion at the operating line on USD 15.5 billion of revenue, gross margin went negative, and the industry collectively cut wafer starts by roughly 30 percent. FY2025-26 is the AI rebound: 26 percent operating margins for Micron in FY2025, and Sandisk's most recent quarter shows a sharper-still inflection (next section).

The Sandisk-specific cycle view, even with only three years of separated reporting, shows the same shape:

The Q3 FY2026 quarter produced revenue of USD 5.95 billion and a 78 percent gross margin. Annualized, that quarter alone runs at roughly USD 24 billion of revenue and USD 16.4 billion of operating profit. An operating loss of USD 1.4 billion in fiscal 2025 became an operating profit of USD 4.1 billion in a single quarter twelve months later. No reader should assume the rate is durable — but it is the clearest possible illustration of memory's cycle amplitude.

4. Where Sandisk's revenue actually comes from — and where it is growing

Sandisk discloses revenue in three end-market buckets: Cloud (hyperscalers and private cloud — primarily enterprise SSDs), Client (OEM PCs, mobile, gaming, automotive, embedded), and Consumer (retail SSDs, memory cards, USB drives). Today the mix is heavily weighted to Client and Consumer; the investment case rests on Cloud taking over.

Cloud revenue tripled in fiscal 2025, rising from USD 325 million to USD 960 million (+195 percent), driven by a 153 percent increase in exabytes shipped to data-center customers and a 17 percent increase in ASP per gigabyte. Management said in the Q3 FY26 earnings call that subsequent Cloud revenue grew 645 percent year-on-year — a number that is plausible only because the starting base was tiny and AI customers are taking everything available.

Client and Consumer are flat-to-mature. Together they represented roughly 87 percent of revenue in fiscal 2025 but barely grew. These are the legacy SanDisk-brand businesses that defined the company before AI. They generate cash and brand awareness but are not the growth story.

Geography is heavily Asia-anchored. Roughly 80 percent of revenue ships internationally; Asia alone was 61 percent of fiscal 2025 sales. This reflects both manufacturing location (Japan, Malaysia) and end-customer geography (Chinese OEMs, Korean and Taiwanese assemblers, hyperscaler regional spend).

5. The structural shift: AI demand meets the "New Business Model" contract

For thirty years NAND was sold like crude oil — spot pricing, quarterly contract negotiations, almost no forward visibility. Sandisk and its peers are now changing that. The New Business Model (NBM) — Sandisk's term — is a multi-year supply agreement under which a hyperscaler commits to volume in exchange for supply assurance, backed by customer prepayments and third-party financial guarantees that compensate Sandisk if the customer fails to take the agreed volumes.

Why are customers signing? Because the alternative is not having any NAND at all. Hyperscaler data-center capex is on track to exceed USD 1 trillion by 2030 (Sandisk CEO David Goeckeler, Q1 FY26 call). AI workloads — inference, retrieval-augmented generation, KV cache, agentic systems — consume enormous quantities of high-density enterprise SSD storage that sits adjacent to the NVIDIA GPU stack. The customers' own statement, paraphrased by Sandisk and reported by Reuters in January 2026: "customers prefer supply over price."

6. The adjacent markets — DRAM, HDD, and where they overlap with NAND

NAND does not exist in isolation. Investors evaluating Sandisk should understand the two adjacent memory and storage profit pools because they interact with NAND on demand, on capex competition for fabrication tools, and on the customer wallet.

DRAM (volatile memory) — the chip class that holds active data for the CPU/GPU. Made by the same companies (Samsung, SK Hynix, Micron). The flagship sub-product is HBM (High-Bandwidth Memory) — DRAM stacks bonded directly to AI accelerator packages. HBM is the single largest beneficiary of AI capex and SK Hynix is the share leader (HBM revenue more than doubled in CY2025; SK Hynix's group operating margin reached 49 percent for the full year). Sandisk does not play in DRAM, but is developing High-Bandwidth Flash (HBF) — a NAND analog — to attack the same accelerator-package socket.

HDD (mechanical disk drives) — Western Digital (post-spin) and Seagate. Lower cost per terabyte, higher latency. Hyperscalers still use HDD for the bulk of "warm" and "cold" storage tiers; SSD/NAND wins where speed matters. The crossover price point keeps shifting in NAND's favor as 3D NAND layer counts increase (Sandisk is on BiCS8, with BiCS9 in development).

The peer-level scoreboard below shows how those three pools price differently in the public markets today:

Pure Storage demonstrates the alternative path: it does not make NAND at all. It buys NAND from Sandisk's peers, wraps it in software, sells it as enterprise arrays with a subscription called Evergreen, and earns a 70 percent gross margin because software pricing dominates the bill of materials. That is the profit pool downstream of Sandisk — more stable but smaller in absolute dollars.

7. Underlying market size and growth

The broader Hardware Storage market — finished storage devices of all kinds — is forecast at USD 58.8 billion in 2025 rising to USD 100.3 billion in 2035, a 5.5 percent CAGR (Market Research Future). That is not the right denominator for Sandisk. The right denominator is NAND flash semiconductor revenue, which sits inside the broader storage pool and tracks much closer to the AI capex curve.

The 5.5 percent CAGR understates the NAND-specific opportunity. AI training and inference are massively storage-intensive on a per-workload basis — frontier model training runs require petabyte-scale checkpointing, inference clusters need low-latency NAND adjacent to GPU memory, and retrieval-augmented generation pushes whole reference corpora onto SSDs. The CEO's framing of "USD 1 trillion of data-center investment by 2030" is the right anchor for the NAND-into-AI thesis; the 5.5 percent storage CAGR is the right anchor for a cycle-neutral baseline against which the AI upside is measured.

8. Regulation, geopolitics, and the cost of being a memory maker

NAND is among the most geopolitically exposed industries on the public market. Three forces are worth holding in mind.

Trade and tariffs. Sandisk discloses in its fiscal 2025 10-K that the U.S. announced trade-policy changes in 2025 including new tariffs on imported goods. The majority of Sandisk's U.S.-sold products are currently exempt, but any loss of exemption would raise cost of goods sold for the U.S. revenue mix. Sandisk manufactures wafers in Japan and assembles in Malaysia — meaning Sandisk-branded products are not domestically produced even though the company is U.S.-headquartered. China sales also face U.S. export-control risk on advanced NAND nodes.

Government industrial policy. The U.S. CHIPS Act, EU Chips Act, Japan's METI, Korea's K-Chips Act, and equivalent Chinese subsidies are reshaping where future memory fabs are built. Sandisk specifically flags in its risk factors that "we may also have difficulty effectively competing with manufacturers benefiting from governmental investments." A state-backed competitor with a lower cost of capital is the most credible long-term threat to industry margin discipline.

Tax regime changes. The U.S. "Big Beautiful Bill Act" signed July 4, 2025 reversed mandatory capitalization of U.S. research-and-development expenditures but kept the foreign research-and-development capitalization rule. Sandisk also faces the OECD Pillar Two 15 percent global minimum tax — material in fiscal 2026 onward as more jurisdictions enact local versions. Malaysian tax holidays on Sandisk's assembly operations expire in stages 2028-2031.

The shape is the point: today's biggest risk is the cycle itself (Sandisk is at the peak by any reasonable historical reading). On a three-year view, structural threats — China capacity, the Kioxia JV expiry windows of 2029 and 2034, and export controls — become the dominant concerns.

9. Why the JV with Kioxia is the most important fact about Sandisk's economics

No other peer has this structure. Flash Ventures — Flash Partners (expires 2029), Flash Alliance (expires 2029), and Flash Forward (expires 2034) — operates the seven (soon eight) Japan fabs that produce roughly 80 percent of total capacity in facilities owned by Kioxia. Sandisk holds a 49.9 percent equity interest; Kioxia, 50.0 percent. Each side takes half the output at cost plus a small markup. Each side is obligated to fund 49.9-50 percent of capital investment.

This means three things investors must internalize:

1. Sandisk's reported property, plant, and equipment understates its real capacity asset base — most of the fab assets sit on Kioxia's books. "Fixed asset turnover" ratios are not comparable to Micron's or SK Hynix's.

2. Sandisk cannot unilaterally expand capacity beyond what the JV agreement specifies. Both parties must agree, both must fund, and the term must be extended. Flash Partners and Flash Alliance both expire on December 31, 2029 unless extended — a renewal that has happened before but is not automatic.

3. Sandisk's earnings from Flash Ventures are recognized one quarter in arrears on the equity-method line ("Other expense, net"), not in revenue or gross profit. That timing mismatch means Sandisk's reported gross margin and the JV's underlying margin trajectory can diverge for a quarter at cycle inflection points.

The post-spin Sandisk is essentially a marketing, brand, controller-design, and assembly company with a 50 percent claim on the world's #3 NAND fab footprint. That is more capital-light than Micron and more capital-heavy than Pure Storage. There is no exact public-market analog.

10. Outlook — where the industry goes from here

Three scenarios bound the next twenty-four months.

Upside (continuation): Hyperscaler capex stays at the current run-rate, NBM contracts cover a growing share of Sandisk's output, BiCS8 yields the cost-per-bit advantage the company expects, and Cloud revenue compounds at triple-digit rates for another four to six quarters before normalizing into a 25-35 percent operating-margin steady state — well above any prior cyclical peak.

Base case (gradual normalization): AI capex remains strong but ASPs grow more modestly as supply re-enters. NBM contracts smooth the trajectory but do not eliminate the cycle. Operating margins fade from the Q3 FY26 peak toward the high-teens to low-20s range — better than any prior up-cycle plateau because of contracted volume, worse than the most recent quarter's spike.

Downside (memory regression to the mean): A hyperscaler capex pause, a recession that pressures Client/Consumer, or a China-driven NAND oversupply re-introduces the historical cycle. ASPs fall, NBM contracts get tested, the industry collectively cuts wafer starts, and the cycle bottoms in 2027-28 before recovering. Margins compress to fiscal-2024 levels or below.

The investment task in the rest of this report is to weight those scenarios. The industry's lesson is that all three have happened in the last seven years — the memory cycle is not gone, it has been temporarily masked by a demand wave large enough that no historical comparison fully applies.