SanDisk (Trans): 90 bn contract, 80% GPM, Adj. FCF margin 50% — all to shareholders!

Below is Dolphin Research’s compiled$Sandisk(SNDK.US) FY26 Investor Day summary.

Key takeaways: FY26 revenue of $20 bn (+175% YoY), a new biz. model with $93.9 bn in total contract value locked, and GPM still ~80% even at floor pricing. CapEx intensity remains mid-to-low single-digit of revenue, non-GAAP OPM reaches 75%, and Adj. FCF margin 50%, with all FCF returned to shareholders. However, FY28–FY30 revenue growth is capped by bit supply at mid-to-high teens. HBF’s first storage die has just taped out, with customer sampling in 2027 modeled only as OpEx/CapEx, not revenue.

I. Core highlights:

1. Shareholder returns: 100% of excess cash to shareholders, executed fully via buybacks

Excess cash is defined as operating cash generated minus cash reinvested back into the biz., with no other deductions. FY26 Q4 generated $5.0 bn in FCF and repurchased $4.5 bn. Of the prior $6.0 bn authorization, $4.5 bn is used and $1.5 bn remains; the BOD added $14.0 bn, bringing total unused authorization to $15.5 bn.

2. Outlook

a. FY27: bit growth mid-teens, with ASPs rising modestly each quarter. b. FY28–FY30 revenue: mid-to-high teens, broadly in line with bit growth.

c. FY28–FY30 profitability: non-GAAP GPM ~80%, OpEx at ~5% of revenue, minimal other income/expense, driving non-GAAP OPM of 75%. d. FY28–FY30 cash: after taxes, working capital and CapEx, Adj. FCF margin at 50%. Model CapEx intensity at mid-to-low single-digit of revenue.

3. FY26 key financials

a. Totals: full-year revenue $20 bn (+175% YoY), double the 2022 record, improving sequentially. b. Margins: full-year GPM 71.6% (vs. 30.3% last year), Q4 exit GPM 84.6%. Full-year non-GAAP EPS $39.25 (vs. $0.29).

c. Cash flow: full-year Adj. FCF $8.7 bn ex-advance payments and deposits under the new biz. model, improving sequentially. Q4 FCF was $5.0 bn, or $20 bn annualized. d. Quarterly vs. base: last week’s print showed Q4 revenue $9.0 bn and Adj. FCF $5.0 bn, vs. FY25 Q3 revenue $1.7 bn and Adj. FCF $220 mn.

4. New Business Model (NBM, i.e., multi-year LTAs) contract scope and financial guarantees

a. Total contract value (TCV) at $93.9 bn and remaining performance obligations (RPO) at $91.1 bn, both measured at floor pricing. Management expects actual realized pricing above floors, implying upside to both. b. $16.5 bn in financial guarantees secured, including $2.9 bn in customer deposits and credit lines (with $2.5 bn received). The rest are LCs or performance bonds from financial institutions.

c. Customers pay on normal terms, so guarantees stay roughly constant over contract life. For a 3-year deal, the ratio of guarantees to remaining obligations about doubles vs. inception after two years.

5. Balance sheet and capital structure

a. No debt; TLB repaid; RCF undrawn with no plans to draw. Cash balances remain within recent ranges. Credit rating improved to BB+ with plans to continue improving. b. This quarter recognized $800 mn gain on an equity stake in a DRAM manufacturer, making it one of the few quarters where GAAP profit exceeds non-GAAP.

6. Strategy: three value-creation priorities

a. Capital allocation stack: reinvest in the biz. first, maintain a net cash balance sheet, and return residual cash to shareholders. b. Priorities: improve profitability, dampen cyclicality, and sustain growth. A year and a half ago, the debate was about achieving profitability; now it is about flattening the cycle.

c. Decisions balance short, mid and long horizons, avoiding single-point answers. d. Two de-risking investments: >$1 bn with Kioxia to extend JV NAND production through 2030–2034, and a 4% stake in a DRAM maker to secure DRAM for DC SSD LTAs (CFO later referenced a DRAM agreement with Nanya).

e. Full-stack control underpins profitability: from NAND IP and front-end fabs with Kioxia, to system and controller engineering, back-end in Malaysia/China, and a direct go-to-market. This eliminates leakage, supporting higher GPM vs. fabless peers.

7. Biz. model shift: from quarterly pricing to multi-year LTAs

a. Within two quarters, visibility extended from 3 months to 4+ years, with 50% to two-thirds of supply under agreements. About two-thirds (66%) of FY28 shipments will be under NBM. b. Eight customers signed, including three US hyperscalers. The earliest deal was only signed in Jan., and two customers already sought to extend term or add bits within the same contract period.

c. Contracts are bespoke yet share a common framework: volume certainty over price, multi-year tenure, quarterly/monthly volumes, pricing with fixed and floating bands (with floors and ceilings), and financial guarantees. Each requires BOD approval and engagement with customer CFOs, treasurers and CEOs. d. Avg. term exceeds 4 years, max 5 years, with discussions underway for longer terms. Goal is for NBM to become the primary operating model for decades, while staying selective and focusing on winners willing to pay and commit long term.

e. The non-NBM book remains market-priced to serve sub-scale or LTA-unsuited customers and emerging NAND use cases. Following last week’s results, customers have proactively proposed 3- or 5-year deals, including self-declared pricing and guarantee structures.

8. NAND technology: BiCS plus CBA hybrid bonding

a. Strategy prioritizes horizontal scaling, logic and architecture for higher capital efficiency over simply stacking more layers. Since 2001, 19 NAND generations have shipped, with BiCS9 and BiCS10 added in the past 18 months. b. Capital efficiency: in 2025, the industry’s avg. capital per incremental PB is 2.66x the company’s. From 2021–2025, with Kioxia, 13% of industry capital delivered 29% of industry bits, about 2x the next-best competitor.

c. CBA hybrid bonding enables a ‘derived roadmap’: pair existing array wafers with next-gen CMOS to rapidly boost performance with minimal incremental capital, addressing compressed PCIe Gen4→Gen5→Gen6 cycles in data centers. d. BiCS9 combines BiCS8’s mature array with next-gen CMOS, built to hyperscaler specs (performance, scale, speed) and included in NBM. BiCS8 remains the capacity workhorse and industry benchmark, outperforming peers’ 2xx-gen on performance and power, while peers’ 2yy/3xx show performance gains but limited power gains.

e. BiCS10: 1Tb TLC die sampling to customers this month, ahead of schedule. The 2Tb QLC die is the world’s densest memory chip, >60% higher density vs. BiCS8, 2x+ read/write bandwidth, and 75% better power efficiency. A BiCS10 2Tb wafer carries 65% more bits than a BiCS8 2Tb wafer.

f. Bit-per-wafer CAGR from BiCS5 to BiCS11 is 27% annualized (avg. 54% per node across five nodes in CY20–CY30), about 50% above long-term demand forecasts in the high teens, hence wafer starts are trending lower. BiCS11+ are in development, with BiCS13 teams formed, with HVM well past 2030.

9. Consumer and edge: cycle shock absorbers

a. A global footprint of 351k points of sale and billions of units sold over the past decade, with high brand recognition. Since independence, global share rose by 2 ppt, and the portfolio has been fully rebranded. b. Consumer GPM tracks the industry over time but is notably resilient in downturns, remaining profitable even through the 2023 trough, smoothing the cycle. Management does not expect a repeat of 2023 conditions.

c. This segment uses wafers more intensively than other markets, supporting better profitability. The current strategy is brand-first to scale, as the business is not yet large enough to fully offset company-wide cyclicality.

10. Industry and market outlook

a. Global flash shipments expected to reach 1.2 ZB in 2026. Data centers’ share of bit TAM rose from 20% in early 2020s to 30% last year and 50% this year, and will outgrow the market. b. Revenue unfolds in two phases: earlier, flash was commodity-priced with ASP declines offsetting volume growth, yielding ~$60 bn avg. annual industry revenue with ~$20 bn cycle swings. Now flash is a key multi-year DC component, with industry size exceeding $300 bn in CY26 and approaching $500 bn in CY27. This lens includes China; back-outs for other regions will be distorted.

c. Supply: late-2010s targets aimed for >30% bit growth with wafer starts peaking at 1.8 mn wpm in 2022. Post-COVID, inventory digestion cut demand and utilization by 0.5 mn wpm, structurally resetting capacity to <70% of peak, yet mid-to-high teens bit output growth continues via node migration. d. Cloud: US DC CapEx for major hyperscalers and Neocloud has been revised up for 15 consecutive quarters since 2023, with combined 2026–2027 spend now at $1.9 tn. AWS noted capacity is insufficient for near-term demand, with annualized revenue potentially approaching $1 tn.

e. Edge: PC and smartphone shipments are expected to decline mid-teens YoY, entirely from the low end. OEMs are shifting to higher-end models, with ASP and revenue growth continuing in 2026–2027. Recent major OEM YoY revenue growth ranges from 13%–24%.

11. AI data centers and KV cache

a. Flash now permeates multiple AI infrastructure layers, led by large QLC data lakes, near-GPU direct-attached buffers and checkpointing with high-performance TLC, and the fastest-growing use case of KV cache. By 2030, TLC will dominate NAND mix with QLC still sizable. b. Infra metrics are shifting from TCO to TVO, where output is intelligence. KV cache expansion in inference is driven by longer contexts, deeper chains, multimodality and persistent model memory.

c. KV cache tiers: HBM and system DRAM serve short-lived ephemeral cache with capacity limits. The overflow persists on flash as persistent KV cache spanning tiers, from ‘ultra-cold’ in data lakes to ‘hot/warm’ in direct-attached SSDs (G3) and ‘lukewarm’ at rack/network-level persistent layers (G3.5). Without persistent layers, intelligence cannot scale, as models are trained once but serve billions of interactions.

d. Four variables drive capacity: session count, retention time, cache miss rate, and storage per session (a function of context length and per-token size, at tens to hundreds of KB per token). Based on this, persistent KV cache deployments reach 1 ZB by 2030, with KV cache at 35% of the market. The 2030 AI DC TAM equals total industry shipments in 2026.

e. In an in-house AI Lab test (server-class, hundreds of user sessions), SSD-equipped systems consumed 75% less energy and delivered 75% higher tokens/sec vs. HBM/DDR-only. Power to generate 1 mn tokens was one-fifth of volatile-only memory systems, as capacity limits otherwise force recomputation. f. Quantization and inference optimizations reduce per-unit KV storage, but trigger Jevons paradox, adding use cases and lifting total demand.

12. Data center product portfolio

a. Both TLC eSSDs and high-capacity QLC SSDs have been qualified and shipped at major hyperscalers, large accounts and OEMs. b. TLC eSSD is shipping across PCIe Gen5 form factors, with Gen6 shown at FMS, targeting high performance and low power to alleviate inference bottlenecks.

c. High-capacity eSSD showcased a 256TB E3 product at FMS, with 128TB giving way to 256TB as mainstream next year.

13. HBF and 3D Matrix Memory (excluded from revenue guide; OpEx/CapEx only)

a. HBF delivers HBM-class read bandwidth with 8–16x capacity, targeting long-context, large KV cache MoE sparse models. Frontier models now reach 1–2 tn parameters and million-token contexts, shifting to memory-centric AI where memory needs rise as compute needs ease. b. Four published architectures: all-HBF replacing HBM, hybrid HBM+HBF, a small HBM as cache in front of HBF, and prefill/decode decoupling (prefill on high-perf GPU + standard DDR, decode on mid-tier GPU + HBF).

c. Simulation vs. a market-leading GPU (192GB HBM per card, replaced by ~4TB HBF): for multi-turn agentic code dev using a 490 bn-parameter Qwen3, the HBM setup needs at least eight GPUs to fit the model, while HBF runs on a single card. Four HBF GPUs match eight HBM GPUs, lifting CapEx efficiency by 8x at minimum config and doubling GPU efficiency at full token throughput. Mechanism: workloads quickly exhaust high-bandwidth capacity and spill to system memory, cutting GPU utilization by nearly 50%.

d. Ecosystem: announced partnership with SK Hynix in Aug last year and formed an OCP alliance in Feb (with Google and Tenstorrent), releasing a first public spec last week for XPU integration. Meta has joined, with specs iterating over the next two years. The tech advisory board includes David Patterson and Raja Koduri, with Jim Keller added.

e. Progress: first HBF storage die has taped out (image redacted; details withheld), with initial inference customer samples in 2027. A second-gen ‘HBF for the Edge’ is co-developed with multiple customers to run 100 bn+ parameter models on edge devices, enabling ‘never-forget AI’. f. 3D Matrix Memory: started in a development fab in 2017 and moved to IMEC’s 300mm line in 2024. Delivered 300mm wafers and packaged samples with multi-Gb functional arrays near product specs, though on a longer timeline.

II. Key Q&A

Q: How do you diversify NBM customers to avoid overexposure to any single end market?

A: Signings follow rigorous screening across data center and edge markets. Even among hyperscalers, biz. models differ widely, and such diversification is itself a selection criterion. We back long-term winners, with integration depth at a historical high across technology and commercial fronts.

Q: Before HBF launches, if GPU customers cut HBM, how much HBF or standard eSSD content is needed to maintain performance?

A: The market is moving fast, making ‘what to build and how much’ hard, especially under the old build-first price-later paradigm. Inference at scale is the sharpest tech shift in our careers, and customers relentlessly drive per-unit needs lower for power, space and efficiency to scale faster and cheaper. Being close to customers is critical, and NBM gives us their quarterly deployment lists for 3–5 years. This offers unique guidance during architectural flux on what to build and where to allocate resources.

In lab runs on multi-user sessions of a 1.2 tn-parameter model, we shmoo-scanned HBM and DRAM capacities from 4GB up to server max. For complex inference, HBM plus SSD sufficed, with minimal DDR needed for batch services. This was a surprise internally. Industry chatter suggests Vera Rubin halves system DRAM (SOCAMM). Our experiments indicate systems with eSSD plus HBM could cut further. eSSD capacity depends on workload and the persistent KV cache pool, so it is use-case specific; however, reducing system DDR for inference is clear, whether by half depends.

Q: HBF is not in the model and bit growth is only mid-to-high teens. If HBF succeeds, how will it be modeled, what gets displaced, and how do you size the upside?

A: We will address that once products are in customers’ hands and needs are clear. We follow customers—build what they ask for rather than anchor on industry supply. Aligning incentives is key; a year ago incentives were misaligned, amplifying volatility to everyone’s detriment. Visibility has moved from 3 months to 4 years in two quarters. We will reassess a year out after customer feedback. With full-stack control from NAND IP to front/back-end, options exist, but it is premature to detail them now.

Q: HBM is getting customized with compute on the base die. Will HBF follow?

A: Compute near or in memory is a natural trend given data-movement costs. Near-memory or in-memory compute will happen, and we can support it. We already have memory-management capabilities and welcome compute adjacent to or within memory—this is our strength. Scope depends on what customers want precomputed or computed in controllers adjacent to memory. Once aligned, we can deliver.

Q: If HBF ramps in 2028/2029, will CapEx plans change? Is advanced packaging for HBF capital-intensive?

A: HBF reuses our established NAND capabilities, which is why we could lead with HBF. Today’s NAND roadmap productivity can raise output easily when chosen to expand, and HBF provides an outlet, while being orthogonal to the storage biz. We have the related capabilities, and will reassess once products hit the market next year. These are the kind of high-class problems we are willing to solve.

Q: NBM covers two-thirds of FY28 shipments. Is there a longer-term target for NBM mix? Similar in FY29 and beyond?

A: The mix will be optimized over time. We are comfortable with FY27 and FY28 levels based on current data, and will keep engaging customers and reassessing. The ratio may evolve, with the goal of sustainably maximizing shareholder value. FY29 appears similar to FY28 in what we see today, but we did not disclose it because optimization continues.

Q: How do you define ‘excess cash’ under the 100% return policy? Is it equal to 100% of FCF?

A: It is straightforward: cash generated minus cash reinvested in the biz., nothing else, and 100% of that excess goes back to shareholders. Last quarter we generated $5.0 bn and repurchased $4.5 bn, reflecting this policy.

Q: On using NAND for KV cache, concerns include tail latency hurting throughput and heavy KV writes shortening device life. Your response?

A: Latency must be viewed by tier. The top high-bandwidth tier is capacity-limited, which forces context to spill to lower tiers. Overall latency is performing well and is workload-dependent—latency-critical tasks need not descend to lower KV tiers and can stay on higher bandwidth layers. It is not either-or. Endurance continues to improve, benefitting from core tech and HBF learnings. Warm cache requires about 3 DWPD, and we offer KV cache DWPD 10x standard SSD for warm workloads. Read-heavy workloads continue to use standard NAND.

Q: Data centers are roughly one-third of the mix, but demand could double in 12 months. How will the mix look in a year? Will edge be crowded out, and how do you balance when supply is tight?

A: Contract terms vary, so FY29–FY30 will involve expiries and new starts. We will keep flexibility in the system and closely follow customers for higher visibility, which benefits all parties and improves decisions. The model started in data centers due to their needs, but customers now proactively approach us for LTAs. We must keep the portfolio healthy with flow across edge, consumer and data center, keeping engineering fully utilized across all three. Data center mix will rise, and we will continuously rebalance to optimize returns across time horizons. Final mix depends on deals signed. We have a long opportunity list and have been aligning structure with customers; the earliest few are large.

Q: Will HBF tighten conventional NAND like HBM did to DRAM? Is Gen1 HBF based on BiCS8 or BiCS9? Pricing and capital intensity?

A: Pricing is not disclosed. HBF arrays are primarily based on BiCS8, leveraging existing capital and installed base, while CMOS uses the most suitable node, with higher performance than BiCS8’s own CMOS, similar to BiCS9 logic. On impact, today’s tech enables 27% annual bit-per-wafer growth, while we only deploy about two-thirds of that, leaving room to accommodate HBF via innovation. If HBF scale exceeds that window, it is a high-class problem; we can scale manufacturing.

The CEO added: these questions all ask where capacity comes from and whether we can scale. NAND is highly scalable, with at least a decade of runway, and the most scalable tech wins. The real issue was pulling vast wafer volumes during demand gaps due to poor visibility. 2026 supply/demand is set by fab plans, but further out the knobs can be tuned. KV cache size is the product of four variables, and any change can alter outcomes massively. Those investing billions in infra can answer the 5–10 year build question. We must stay close to them to understand their roadmaps and help them grasp what is possible.

HBF’s key is the same—non-experts cannot be expected to solve the memory wall. We have co-developed one generation with hyperscalers. Current contracts are fully customized collaboration, not NCNR. In two quarters, visibility extended from 3 months to 4 years; we will see results next year.

The CTO added: BiCS10 vs. BiCS8 delivers 65% more bits per wafer in one generation. DRAM gains only single-digit percent per gen every ~1.5 years, meaning one wafer gain in NAND equals 10+ years of DRAM scaling. If demand surges, switching to BiCS10 adds 65%—not difficult. The CEO added: DRAM is excellent tech, but scaling it further mainly requires more capital.

Q: In NBM, do you retain 100% of annual cost-downs, or share with customers?

A: Pricing details are sensitive. The key is: even at floor pricing, NBM GPM is not expected to fall below 80%, and should be around 80% each year of the contract, which can be used for modeling.

Q: Among the four HBF architectures, where is customer interest concentrated? Is decode decoupling a focus?

A: The first two (all-HBF and HBM+HBF) are nearly identical, chosen based on needs. From HBF’s perspective, decoupling resembles all-HBF with deployment split at the system level, so one architecture can serve all three. Mobile is different and handled separately. We have not yet seen customers adopt the third (small HBM as cache in a layered design), though academia has published solid data. There are also proprietary architectures we cannot discuss.

Q: Your roadmap supports 27% bit CAGR, well above mid-teens demand guidance. One way to close the gap is lengthening node cycles (historically 18 months). If others do not slow, how do you balance share?

A: These are R&D productivity numbers. Converting to fab plans takes years and cannot be dialed up or down quickly. Everyone runs multi-year, per-node, per-fab, per-yield plans that sum to bit output. Change is possible but slow. We emphasize highly capital-efficient scaling. When WD split, HDD at 4–6% CapEx intensity was a very good biz., while here we commit to low single-digit.

The inverse also holds: if you mis-execute, overproduction happens fast. Pre-2023, plans assumed 30% bit growth and 15% cost-down, which proved unrealistic and would not justify investment. We must throttle this productivity via portfolio planning, while recognizing new businesses and innovations will bring demand. Another lens: scalability attracts TAM—the more scalable, the more adoption—so keep horsepower in the engine without crushing the market.

The fundamental fix is aligning investment cycles. Quarterly auctions do not fit 10-year investments. The industry is shifting rapidly to a new model. Waiting for the old cycle is a waste; methodical change can compound into different outcomes.

Q: With BiCS9 reusing NAND cell arrays and longer node cycles, will future NAND cost declines slow to ~10% or less?

A: We said last time we will not discuss cost-downs and we have stuck to that. Cost-downs are our productivity, not the customer’s. Debating it repeatedly makes everyone assume it belongs to customers, which it does not. Product prices should be market-driven, not cost-driven. Will future costs be lower? Yes, but specifics will not be disclosed. That is the value of 19 generations of NAND accumulation.

Q: Veteran investors doubt NBMs: LTAs hold in upcycles but often break in downcycles. Even hyperscalers over-forecasted during COVID. Why is it different now? Why should 2030 forecasts be stronger?

A: Skepticism is easy, but we must believe change can drive better outcomes. We hire people willing to change; many fear breaking things, but the industry was already broken. Our CFO brought a fresh perspective and asked ‘why do it this way’. Other industries do it differently and get different results, so doing differently can compound benefits when done with confidence and method. Will it be different? It may not be, and we acknowledge the risk. But uncertainty and skepticism also harbor the greatest returns. The past 18 months’ returns came from extreme skepticism off a very low base, below replacement cost for a fraction of the biz., making CAGR highly sensitive to endpoints. We must let go of the past and assess what is most likely, rather than assume a reversion.

The change is in relationship altitude. In 2000, meeting a CEO’s lieutenants got us a next-day scolding from supply-chain teams. Now we have strategic dialogs directly with CEOs and CFOs. They could renege, but that is not the base case. They are public-company executives with boards and fiduciary duties; no one signs contracts just to exit. No guarantees, but this is the most likely path, and our capital return policy signals our conviction.

Q: On physical AI and non-DC demand: assume 20%–30% of bits are still consumed there post-FY28, yet the market hears of client-edge destocking and pricing pain. For customers (some in NBM) how do they see demand beyond CY26?

A: We cannot discuss individual customers. We have deep engagement with physical AI customers, understand their needs and collaborate closely. Relationships are in good shape, but details are confidential. It is a market we are committed to, seeing growth and significant potential.

<End of text>

Risk disclosure and statements:Dolphin Research Disclaimer and General Disclosure