Semiconductors
PrivateThe purest liquid bet on the NAND super-cycle — no DRAM cushion means it prints the most in the up-leg and bleeds the most in the down-leg; at ~8x trough-of-cycle-inverted earnings the tape has already discounted a peak, and a ~30% ex-PE Bain overhang plus 100% cyclical exposure make it a trade, not a hold.
Research
The verdict
The purest liquid bet on the NAND super-cycle — no DRAM cushion means it prints the most in the up-leg and bleeds the most in the down-leg; at ~8x trough-of-cycle-inverted earnings the tape has already discounted a peak, and a ~30% ex-PE Bain overhang plus 100% cyclical exposure make it a trade, not a hold.
Kioxia is the world's #3 NAND flash memory maker and the only pure-play among the majors — it makes flash, and essentially nothing else. There is no DRAM division, no logic foundry, no HBM, no diversification. When NAND is good, Kioxia is the best-levered name in the industry; when NAND is bad, it has nothing to hide behind. That single fact is the whole thesis.
Lineage. Formerly Toshiba Memory — the flash business Toshiba was forced to sell in 2018 to plug a hole blown open by the Westinghouse nuclear disaster. A Bain Capital-led consortium bought it for ~$18B, renamed it Kioxia in 2019, and after two aborted attempts (killed by the 2023-24 NAND collapse) finally floated it on the Tokyo Stock Exchange on 18 Dec 2024 at ¥1,455/share for a ¥886B ($5.6B) initial market cap.
What it sells (FY2025, ended Mar 2026):
Customers. Named/known: Apple (Bloomberg repeatedly frames Kioxia as "the Japanese supplier to Apple" ); NVIDIA (deep SSD-architecture collaboration — a joint program targeting 100 million IOPS SSDs vs today's ~3M ceiling ); hyperscalers and enterprise-server OEMs buying the LC9-series 122.88 TB NVMe drives. Contract structure is memory-commodity: short-cycle, ASP-driven, no take-or-pay — except that in the current shortage Kioxia confirmed its entire 2026 NAND output is pre-booked/sold out, a rare demand-visibility window for a commodity maker.
Structure. The manufacturing base is a 50/50 joint venture with Sandisk (spun out of Western Digital in Feb 2025) — see Lens 2. This JV is the defining operational feature of the company.
Upstream → Kioxia → end customer, named at each node:
Chokepoints & single-source risks: (1) Total dependence on the Sandisk JV — Kioxia does not own 100% of its own production; capex, output cadence, and technology roadmap are co-decided. (2) Geographic concentration in Japan — Yokkaichi and Kitakami are both in Japan; a single-country earthquake/typhoon exposure (Yokkaichi has taken contamination hits before, historically). (3) Equipment lead-times — adding NAND capacity means 12-18 month Lam/TEL/ASML tool orders, which is why 2026 supply is fixed and cannot flex to demand. This lens is names-complete: ASML, Lam, AMAT, TEL, KLA, Shin-Etsu, SUMCO, Sandisk, Apple, NVIDIA.
Be honest: NAND is a commodity, and commodities have thin moats. What Kioxia has:
Bargaining power: weak-to-neutral vs customers in normal times (buyers multi-source), temporarily strong now (sold out through 2026, ASPs doubling). Over suppliers (ASML/Lam/TEL): weak — those are the true chokepoints. Verdict on moat: shallow-but-real via oligopoly discipline; the only durable edge would be HBF winning a new memory tier.
No segments.csv exists (web-only), so all figures are ``. Kioxia reports two core segments:
| Segment | Q4 FY25 rev (Jan-Mar 2026) | QoQ | Share | Character |
|---|---|---|---|---|
| SSD & Storage | ¥600.3B (~$3.77B) | +99.8% | ~62% of the two | Enterprise/DC/client SSD — the AI engine; higher margin |
| Smart Devices | ¥337.3B (~$2.18B) | +81% | ~35% | Smartphone embedded (UFS/eMMC); seasonal, lower margin |
| Other | ¥65.2B (~$0.41B) | +14.3% | ~6% | — |
The trend that matters: the mix swing is the P&L. In the Sep-2025 quarter, seasonal smartphone demand pushed lower-margin Smart Devices to ~35% of sales, and that mix shift (not weak demand) is what caused the earnings miss and the -23% crash. As the mix rotated back toward server/PC SSD through Q4 FY25 and into FY26, margins exploded — Q1 FY26 guided OP margin ~74%. Geographic breakout is not disclosed granularly; the demand pull is US hyperscaler AI capex + global smartphone replacement. Whoever models Kioxia is really modeling the enterprise-SSD share of the mix.
The FY2025 arc (fiscal year ended 31 Mar 2026) is a textbook memory V-shape — one of the sharpest on record:
| Quarter | Period | Revenue | Net income | Note |
|---|---|---|---|---|
| Q1 FY25 | Apr-Jun 2025 | ¥342.7B | ¥18.2B | OP ¥44.8B (-64% YoY); trough; net profit -74% YoY |
| Q2 FY25 | Jul-Sep 2025 | ¥448.3B (non-GAAP) | ~¥61-89B | OP +11% QoQ (first rise in 3 qtrs); profit -62% YoY; -23% stock crash on soft guide |
| Q3 FY25 | Oct-Dec 2025 | ¥543.6B (record) | ¥89.5B (+114.9% QoQ) | SSD & Storage ¥300.4B; 2026 output "sold out" |
| Q4 FY25 | Jan-Mar 2026 | ¥1,002.9B (~$6.29B) | ¥407.7B (~$2.56B, +84.5% QoQ) | SSD & Storage ¥600.3B (+99.8%) |
Full-year FY2025: revenue ¥2.3376T (~$14.7B), +37% YoY; net profit ¥554.4-554.5B (~$3.48B), roughly doubled YoY from ¥272.3B.
The forward guide is the blow-off top. For Q1 FY26 (Apr-Jun 2026) Kioxia guided net profit ¥869B (a 48x YoY jump), revenue ¥1.75T (~5.1x YoY), operating profit ¥1.298T (~29x YoY, OP margin ~74%) — blowing past the ¥405.6B QUICK consensus by >¥300B. The stated driver: US-dollar NAND ASPs doubled within a single quarter. ROIC guided >60%.
Balance-sheet flags: the up-cycle is repairing the balance sheet fast — Kioxia targets a net-cash position by end of Q1 FY26 and leverage <1x, equity ratio rising toward ~60%. The classic memory-cycle flag — revenue driven by ASP not volume ("significant ASP increase, partially offset by reduced bit shipment") — is explicitly the case here, which is the tell that this is a price cycle, and price cycles reverse. Market reaction to the arc: shares are up ~54x from the ¥1,455 IPO and >660% YTD 2026; the market has fully embraced the up-cycle.
No transcripts on the shelf; sentiment reconstructed from `` coverage of the last ~4 prints:
The shift over time: from "we are bottoming, please be patient" (Aug) → "tight and sold out, but we won't over-guide" (Nov, which the market hated) → "records as far as we can see" (Feb-May). The thing management stopped saying is anything about production cuts (a 2024 refrain); the thing they started saying is "sold out through 2026/2027." Tone is now maximally optimistic — which, for a commodity name, is itself a late-cycle sentiment signal.
Kioxia's cleanest peer is Micron (MU) — the only other name with a real, freshly-sourced multiple. Samsung and SK Hynix are conglomerates where NAND is a slice; Sandisk is the JV partner. Multiples are `` or n/a.
| Company | Ticker | Mkt cap | P/E (fwd) | EV/EBITDA | Notes |
|---|---|---|---|---|---|
| Kioxia | 285A.T | ¥45.59T (~$287B) | ~7.9x (2027E) | n/a | TTM P/E 8.12; pure NAND |
| Micron | MU | >$1T | ~6.8x | ~15.8x | DRAM+NAND+HBM; TTM rev $90.3B, NI $50.5B |
| SK Hynix | 000660.KS | n/a | ~10x (implied — Kioxia at ~20% discount) | n/a | DRAM+HBM leader; best NVIDIA HBM allocation |
| Samsung Elec | 005930.KS | n/a | n/a | n/a | #1 NAND ~32%; conglomerate |
| Sandisk | SNDK | n/a | n/a | n/a | JV partner; pure-play flash comp |
Reading it: both pure-cyclicals (Kioxia ~7.9x fwd, Micron ~6.8x fwd) trade at low headline multiples — the classic "cheap at peak earnings" trap. A single-digit P/E on a cyclical near a cycle top is a warning, not a bargain: the E is inflated. Kioxia sits at a ~20% discount to SK Hynix on 2027E, which is defensible given SK Hynix's DRAM/HBM diversification and superior NVIDIA HBM position. Dividend yield and 5-yr avg ROE: n/a (Kioxia has <2 years of public history). The comp table's honest message: Kioxia is not mispriced relative to peers; it is priced as a peak-cycle pure-play, and the debate is entirely about where we are in the cycle.
With <2 years of listed history the "5-year" window compresses, but the moves are violent and revealing ``:
What the tape reacts to, ranked: (1) NAND ASP direction / supply-demand guidance — the dominant driver; (2) the Bain overhang — mechanical selling pressure independent of fundamentals; (3) AI-demand headlines (NVIDIA collab, hyperscaler SSD). It does not trade on smartphone units or company-specific execution — it trades as a high-beta option on the NAND cycle plus a PE-unwind clock. That is the single most important behavioral fact for sizing a position.
Web-only; every figure ``. No filings to audit line-by-line, so this is a risk map, not a reconciliation.
Regulatory findings (required sub-section). Read regulatory/regulatory-findings.md (Step 0):
"Kioxia Holdings" (FTC OR DOJ OR FDA OR CFPB OR consent decree OR settlement OR fine OR penalty) enforcement): no material enforcement action, consent decree, or antitrust penalty against Kioxia surfaced in coverage. (Historical industry note: DRAM/NAND price-fixing probes have periodically touched the memory sector broadly; nothing specific and material to Kioxia found.)No financials.csv, no share count on the shelf, and Kioxia reports in yen with a fiscal year ending March — so EPS-per-share modeling from primaries is not possible web-only. I therefore project operating profit / net profit trajectory (the metrics the company and Goldman actually guide) rather than fabricate a per-share EPS. Every input labeled; output ``.
Anchors (all ``): FY2025 (ended Mar-2026) net profit ¥554.4B, revenue ¥2.3376T. Q1 FY26 guide: revenue ¥1.75T, OP ¥1.298T, net profit ¥869B — i.e. a single quarter guided above the entire prior fiscal year's profit. Goldman models FY2028 consolidated OP toward ¥10T with ~80% gross margin.
FY2026 (ending Mar-2027) — profit path:
The honest projection: FY26 is nearly locked to be a record (2026 output sold out). FY27-28 is a coin-flip on the cycle — the same +66% capex that funds the bull case builds the bear case's oversupply. Kioxia has no DRAM/HBM cushion, so its through-cycle earnings power is structurally more volatile than Micron's or SK Hynix's — it will out-earn them at the top and under-earn (or lose) at the bottom.
Forecast log: skipped per --watchlist rules (no forecast.ts create in the unattended loop; this is not a committed base case worth Brier-scoring). If promoted to a thesis, the scoreable binary would be "Kioxia FY2027 (ending Mar-2027) reports a YoY operating-profit decline" — the cycle-turn bet.
Bull case. The AI data-center build-out has structurally re-rated NAND from a PC/phone commodity to an AI-infrastructure input. Enterprise SSD is exploding (Q4 FY25 SSD segment +99.8% QoQ), 2026 output is sold out, ASPs doubled in a quarter, and the industry is capacity-constrained because tool lead-times mean supply can't flex for 12-18 months. Kioxia is the highest-beta pure-play on exactly this — no diversification means maximum torque. Add the HBF option (a flash tier between HBM and SSD for AI inference), the NVIDIA 100M-IOPS collaboration, a fast-repairing balance sheet (net-cash target, leverage <1x), Japanese sovereign backing (¥150B subsidy), and a ~7.9x forward multiple at a ~20% discount to SK Hynix. If "the era of cheap SSDs is over" is a regime not a spike, the stock is early.
Bear case (2-3 permanent-impairment risks). (1) It's a commodity cycle, and cycles turn. Revenue is ASP-driven on falling bits — the definition of a price spike. Kioxia's own +66% capex + a possible 3rd Kitakami fab + industry-wide expansion is the oversupply seed for 2027-28; when it hits, a pure-play with no DRAM has nothing to cushion the collapse, exactly as in 2023 (losses, -70% contract prices). (2) The Bain overhang is a mechanical share-price weight — Bain has cut from ~51% to <30% and is still selling into strength; every block sale caps the stock regardless of fundamentals. (3) Single-digit P/E on peak earnings is a trap, not a discount — the E is inflated; on normalized through-cycle earnings the multiple is far higher.
Pre-mortem (18 months out, thesis broke). It's early 2028. NAND ASPs peaked in mid-2026 and rolled over as Kioxia's, Micron's, and Samsung's new capacity (all greenlit at the top) came online into a decelerating AI-capex digestion phase. Kioxia's mix swung smartphone-heavy in a soft consumer year; the pure-play that had the most torque up now has the most torque down and prints a loss quarter. Bain finishes exiting into the weakness. The stock, which round-tripped a 54x move, gives most of it back. Everyone knew it was a cycle; the only question was the month it turned, and the +66% capex told you they'd overbuild.
Are multiples too high? Headline no (~8x); cycle-adjusted yes — this is peak-earnings pricing. Contrarian view of what the market is refusing to see: the bulls are treating a price cycle (ASPs doubling on flat/down volume) as a secular demand regime, and are ignoring that the same capex funding today's bull thesis is manufacturing 2027's oversupply. The market is also under-weighting that Kioxia, uniquely among the majors, has zero diversification to soften the turn.
Dismantling the bull case:
Best analog franchise on Earth, mid-cycle, fully priced — the FCF-inflection thesis is now consensus at ~40x forward and above Street targets; you're buying quality at a cyclical-optimism peak, with China share-loss the under-priced tail. WATCHING, not chasing.
The pure-play picks-and-shovels winner of AI-chip test, printing a vertical Q1'26 (+87%, $2.53 EPS) — but the stock fell ~14% on it because Q2 guidance steps DOWN sequentially and a ~54x P/E prices permanent acceleration; great business, demanding price, cyclical tape. NEUTRAL/WATCHING into the next print.
Best-in-class EDA franchise temporarily wearing an Ansys-debt-and-amortization disguise — the GAAP "collapse" is accounting, not the business; the real risk is paying ~35x forward for a name whose Design-IP leg is structurally cracked and whose synergy math doesn't pay until FY2028.