Semiconductors
PrivateThe single best "shovel" in AI advanced packaging — a ~70%+ dicing/grinding monopoly with 42% margins and razor-and-blade annuity — but priced at ~57x for a business whose FY-Mar-2026 guide (+6.5% rev) just decelerated hard and whose China (~30-36%) and hybrid-bonding tails are real. Great business, demanding price: WATCHING for a cyclical entry, not a chase.
Research
The verdict
The single best "shovel" in AI advanced packaging — a ~70%+ dicing/grinding monopoly with 42% margins and razor-and-blade annuity — but priced at ~57x for a business whose FY-Mar-2026 guide (+6.5% rev) just decelerated hard and whose China (~30-36%) and hybrid-bonding tails are real. Great business, demanding price: WATCHING for a cyclical entry, not a chase.
Disco makes the saws, grinders and polishers that turn a finished wafer into individual, packageable die — the "Kiru · Kezuru · Migaku" (cut · grind · polish) back-end of chipmaking. Founded 1937 in Hiroshima as a maker of vitrified industrial grindstones; it pivoted into precision semiconductor tooling and is today the runaway global leader in wafer dicing and grinding.
The product line, in plain terms:
Revenue model = razor-and-blade. Roughly ~60% equipment / ~40% consumables + service, with consumables specifically at ~25–35% of revenue and carrying gross margins above the equipment itself. Every tool Disco installs seeds a multi-year, high-margin stream of blade/wheel replacements that scales with the customer's wafer volume, not their capex budget — so the installed base is an annuity that keeps paying even in equipment down-cycles.
Customers: the leading foundries, IDMs, memory makers and OSATs — TSMC, Intel, Samsung, SK Hynix, plus advanced-packaging houses Amkor, ASE and the power-semiconductor makers. Overseas sales are ~90% of revenue (89.9% in Q2 FYE-Mar-2026) — this is a globally-sold, Japan-manufactured franchise.
Suppliers / inputs: industrial diamond, precision spindles/motion systems, lasers, and the metallurgy for grinding wheels — Disco is heavily vertically integrated on the consumable/blade side (it makes its own blades and wheels, which is the source of both the margin and the switching cost).
Competitors: #2 is Tokyo Seimitsu / Accretech; also ADT (Kulicke & Soffa), ASMPT, and niche laser players Synova (water-guided laser) and Plasma-Therm (plasma dicing). Disco's share dwarfs all of them (Lens 3).
Where Disco sits: it is a back-end (assembly/packaging) capital-equipment + consumables supplier — one tier upstream of the OSAT/packaging step, selling into both the foundry (front-end wafer thinning) and the packaging house (interposer thinning + package singulation).
Full chain, named at every node:
Diamond / abrasives / laser sources / precision spindles (Disco's own inputs; blades & wheels made in-house)
↓
DISCO (dicing saws · grinders · polishers · blades · wheels)
↓
Front-end foundry / IDM: TSMC, Samsung, Intel, SK Hynix, Micron → wafer thinning (backgrinding), stealth/DBG dicing
↓
Advanced packaging / OSAT: TSMC (CoWoS), Amkor, ASE → interposer grind-to-reveal-TSV, package singulation
↓
Fabless AI designers: NVIDIA, AMD, Broadcom, Google, Amazon (pull-through demand)
↓
Hyperscaler / enterprise buyers
Chokepoint role — this is the thesis. Advanced packaging (CoWoS + HBM), not logic dies, was the binding constraint on AI-chip production in 2025 — the four largest AI designers consumed >90% of global CoWoS packaging capacity and HBM by value but only ~12% of advanced logic die production. Disco's grinders and dicers are inside that CoWoS flow: the molded interposer is ground and polished down to ~100µm to reveal the TSVs, then diced into individual packages. TSMC's CoWoS capacity ramp — ~35–40K wafers/mo (2024) → 75K (end-2025) → 130K target late-2026 — is a direct volume driver for Disco tools and consumables.
Single-source / concentration on the chain: Disco is itself the single-source chokepoint for high-end dicing/grinding — the risk runs the other way (customers depend on Disco, Lens 3). Disco's own upstream dependencies (diamond, lasers) are commoditized and multi-sourced. The real geographic concentration is demand-side China (~30–36%) — see Lens 4/13.
Disco is one of the cleanest near-monopoly + annuity structures in all of semiconductors. Four reinforcing moats:
Dominant, durable share. Estimates cluster at ~52–70% of global dicing/grinding, with the top-3 (Disco + Tokyo Seimitsu ~15–18% + ADT/K&S ~7–9%) controlling >75% of dicing-equipment revenue. In specific consumables Disco is even more dominant: >70% of precision dicing blades, >50% of wafer-thinning grinding wheels by unit. In power-semi and in the hardest materials (SiC), share approaches "lockdown."
Razor-and-blade switching costs. Because Disco makes both the machine and the exactly-matched consumable, it guarantees compatibility and yield — qualifying a rival's blade on a Disco tool (or vice-versa) is a yield risk no fab wants to take mid-node. The installed base compounds the consumable annuity: ~30–35% of revenue is recurring, high-margin, and grows with the customer's wafer starts.
Process/IP moat, not just brand. Disco's edge is a stack of proprietary processes: stealth dicing (subsurface laser, near-zero kerf — dominant in laser dicing, >60% of that sub-market ); DBG (dicing-before-grinding) and Taiko selective-backgrind for ultra-thin die; plasma dicing leadership for die-to-wafer hybrid-bonding singulation; and KABRA — a patented laser SiC-slicing process ("Key Amorphous-Black Repetitive Absorption," 40 related patents) that quadruples SiC wafer productivity and slashes kerf loss. Each is the enabling technology for a next-gen node or material — sold at a price only Disco can command.
Bargaining power. Mission-critical + irreplaceable-once-qualified + a tiny fraction of a fab's total capex = pricing power. The proof is in the P&L: ~42% operating margin and ROIC consistently >30% even through cyclical downturns — margins that sit at the very top of the semiconductor-equipment industry, above ASML on operating margin. That is a moat expressing itself as pricing.
The one genuine threat to the moat is technological, not competitive: wafer-to-wafer (W2W) hybrid bonding singulates after bonding and could, at the limit, reduce dicing passes vs. die-level approaches — see Lens 13. Near-term this is a tail risk (D2W still needs singulation, and Disco leads plasma dicing), but it is the thing a long-term bull must watch.
Disco reports as effectively a single business (precision processing equipment + tools + service) rather than multiple reportable P&L segments, so a clean segment-EBITDA table is n/a — not disclosed at segment level. What is disclosed / sourceable:
By product/revenue type:
| Type | ~% of revenue | Character |
|---|---|---|
| Equipment (dicers, grinders, polishers) | ~60% | Cyclical, capex-driven, AI-led |
| Consumables (blades, wheels) | ~25–35% | Recurring, wafer-volume-driven, highest margin |
| Service / other | remainder | Recurring |
By geography (Q2 FYE-Mar-2024 disclosure — the cleanest regional split found):
| Region | Share |
|---|---|
| China | 35.9% |
| Americas | 13.1% |
| Japan | 12.2% |
| Taiwan | 11.9% |
| Europe | 9.7% |
| Korea | 8.8% |
| Rest of Asia | 8.4% |
Overseas total ~90% (89.9% in Q2 FYE-Mar-2026). China is the single largest market and the central risk variable (Lens 13).
Trend / cause: the mix is shifting from power-semiconductors → generative-AI (memory + logic). Through 2025 Disco explicitly flagged strong AI/HBM/logic shipments and a decline in power-semiconductor demand — i.e. the EV/industrial power-chip slump is being more than offset by AI advanced-packaging pull. That is the accelerating leg; the decelerating leg is China mature-node and power. The FYE-Mar-2026 guide (Lens 5) shows the AI leg alone is not enough to sustain FY2024's +28% growth rate.
Full-year FYE Mar-2025 (the last completed year) — a record:
Latest reported quarter — Q2 FYE-Mar-2026 (Jul–Sep 2025):
Q1 FYE-Mar-2026 (Apr–Jun 2025) for context: net sales ¥89.9B (+8.6% YoY), OP ¥32.7B, record shipments ¥93.0B (+8.5% YoY) — but parent revenue −26.4% QoQ, and management guided slowing Q2 shipments (the lumpiness is real).
Full-year FYE-Mar-2026 guidance (issued 2026-01-21 — notably Disco's first-ever forecast beyond one quarter out):
The signal: growth is decelerating sharply — from +27.9% revenue (FYE Mar-2025) to a guided +6.5% (FYE Mar-2026). Margins hold near-record (~41%), the balance sheet is pristine, but the rate of change has rolled over. The stock is a rate-of-change instrument (Lens 8), which is why it de-rated despite record absolute numbers.
Balance-sheet flags (Mar-2025): cash & deposits ¥229.2B, total assets ¥654.1B, total liabilities ¥161.4B (non-current only ~¥1.0B → effectively debt-free), net assets ¥492.7B. Net-cash, fortress balance sheet. No red flags in inventory/receivables surfaced in the summary data (verify against the actual FR PDF before sizing a large position).
No transcripts on the research shelf (transcripts=0); sentiment is inferred from IR materials + coverage. The tonal arc across the last ~4 quarters:
Recurring phrases: "generative AI," "advanced packaging ramp-up," "record shipments." What they started saying: full-year guidance (a governance/confidence shift). What softened: the power-semiconductor / China mature-node narrative.
Peer set = the semiconductor-equipment majors (front-end + back-end).
| Company | Ticker | USD mkt cap | P/E (TTM) | EV/Sales | EV/EBIT | Div yield | 5-yr avg ROE |
|---|---|---|---|---|---|---|---|
| Disco | 6146.T | ~¥8T ≈ $50–55B | ~57.7 trailing / ~43.3 fwd | n/a | n/a | ~0.7–1% | ~25% ROE (TTM 25.1%) |
| ASML | ASML | ~$300B+ | n/a | n/a | n/a | ~1% | n/a |
| Applied Materials | AMAT | n/a | n/a | n/a | n/a | n/a | n/a |
| Tokyo Electron | 8035.T | n/a | n/a | n/a | n/a | n/a | n/a |
| Tokyo Seimitsu (Accretech) | 7729.T | n/a | n/a | n/a | n/a | n/a | n/a |
| Kulicke & Soffa | KLIC | n/a | n/a | n/a | n/a | n/a | n/a |
Read: I could only cleanly source Disco's own multiples this pass, so the peer columns are honestly marked n/a rather than filled with plausible fabrications. What is sourced tells the story: Disco trades at ~57x trailing / ~43x forward earnings — a premium to the machinery/SPE sector, explicitly attributed to its "monopoly-like position and structural growth". Against a guided +6.5% revenue / +3.2% OP for FYE-Mar-2026, ~43x forward is a quality-at-a-demanding-price setup: you are paying a monopoly multiple for a year of single-digit growth, underwritten by the annuity + the AI-packaging option value. ROE ~25% and ROIC >30% justify a premium; they don't obviously justify 43x on a decelerating year. (A future refresh should populate the peer multiples to test relative value — that is the missing analytical piece here.)
Disco (6146.T) is a high-beta AI-sentiment proxy on the TSE. Moves >5% cluster around:
Price action (conflicting data — surfaced per provenance rules): sources disagree on the current level. One reports ATH ¥81,000 on 2026-02-26 and a "current" ¥63,760 (~−21% from ATH); another reports ¥84,520 on 2026-07-01 with a ~57.7 P/E. These conflict — ¥84,520 exceeds the stated ¥81,000 ATH, so at least one snapshot is stale or mis-dated. I do not silently pick one. The coherent read: the stock is up ~88% over the trailing year but has corrected ~13–21% from its early-2026 peak on the deceleration + downgrades. For position-sizing, pull a live 6146.T quote before acting — treat the price here as indicative, ~¥64k–85k range.
What the pattern reveals: the market pays Disco for AI-packaging shipment acceleration. It is not valued as a steady annuity compounder; it is valued as a levered call on the second derivative of AI-back-end demand. That cuts both ways.
Grounding caveat: no filings on the shelf, so this is a web-only forensic read of a Japanese GAAP/IFRS filer. It cannot substitute for reading the actual 有価証券報告書 (yūkashōken hōkokusho / annual securities report) and FR PDFs — flagged as an open item.
n/a — not quantified this pass.n/a — verify in the annual report.Regulatory findings (required sub-section) — read from regulatory/regulatory-findings.md + web:
Building from actuals + Disco's own guidance. All outputs `` with arithmetic; inputs labeled. Shares outstanding ~114–115M (mkt cap ~¥8T / ~¥70–84k price → ~95–114M; using company guidance NI/EPS relationship below) — treat EPS precisely off the guided NI ÷ share count when sizing.
Anchor (guided) — FYE Mar-2026: Revenue ¥419.0B · OP ¥172.1B · NI ¥126.4B. At ~114M shares → EPS ≈ ¥1,109.
Three-year EPS path (fiscal years ending Mar-2026 → Mar-2028):
| Scenario | FYE Mar-2026 (guided) | FYE Mar-2027 | FYE Mar-2028 | Logic |
|---|---|---|---|---|
| Base | NI ¥126.4B (EPS ~¥1,109) | NI ~¥145B (EPS ~¥1,270) | NI ~¥162B (EPS ~¥1,420) | AI-packaging + HBM4 ramp re-accelerates revenue to ~+12–15%/yr off the FYE-26 base; margins held ~41–42%; ~flat share count |
| Bull | as guided | NI ~¥160B (EPS ~¥1,400) | NI ~¥195B (EPS ~¥1,710) | CoWoS to 130K wafers/mo (late-2026), power-semi/SiC recovers, share gains; revenue ~+20%/yr, mix-driven margin +100–200bps |
| Bear | as guided | NI ~¥118B (EPS ~¥1,035) | NI ~¥112B (EPS ~¥985) | AI-capex digestion + China export-control step-down; revenue −5% then flat, margin −200–300bps on under-absorption |
Base case ≈ ¥1,270 EPS FYE-Mar-2027. Against a ~¥64–85k price that is ~50–67x forward — the multiple is the risk, not the earnings. The business will very likely grow EPS; whether it grows into a 50x multiple fast enough is the entire question.
Forecast tracking: per --watchlist rules, not logging a forecast.ts Brier forecast in this unattended breadth pass (log one only on a genuinely committed base case via /thesis). Suggested forecast to log later: "6146.T FYE-Mar-2027 net income ≥ ¥145B," p≈0.55, resolves 2027-03-31, tags disco,deep-dive.
Bull case. Disco is the irreplaceable toll-booth on AI advanced packaging — the one company whose ~70% dicing/grinding monopoly and razor-and-blade annuity mean it gets paid on every HBM stack and every CoWoS interposer, regardless of which fabless designer or foundry wins. Advanced packaging is the binding constraint on AI compute through 2026, CoWoS capacity is set to nearly 4x from 2024 to late-2026, HBM4 is ramping — and Disco's tools + consumables scale directly with that wafer/package volume. Layer on KABRA's SiC lock for the eventual EV/power recovery, ~42% margins, ROIC>30% through cycles, net-cash, and a founder-family owner with top-of-industry capital discipline. The consumable annuity (~30–35% of revenue, growing with the installed base) means even an equipment air-pocket doesn't break the model. This is a compounder you rarely get to buy cheap — and every AI-capex up-cycle re-rates it.
Bear case (permanent-impairment candidates).
Pre-mortem (18 months out, thesis broke). It's early 2028. AI-capex hit an air-pocket in 2026–27 as hyperscalers digested a build-ahead; CoWoS ran ahead of end-demand and utilization fell. Simultaneously, Japan tightened export controls to cover advanced-packaging dicing/grinding, knocking out a chunk of the ~35% China book. Disco's shipments went sideways for four quarters; the consumable annuity cushioned earnings but not the narrative. The 50x multiple that assumed perpetual acceleration collapsed to 25–30x. The business was fine; the stock halved.
Are multiples too high? For a single-digit-growth guided year, yes on a 12-month view — you're paying a monopoly premium for a cyclical trough-ish year. On a 3-year view, if AI-packaging volume compounds as the KB bottleneck data implies, the growth catches the multiple. It's a timing/entry question, not a quality question.
Contrarian view (what the market is refusing to see): the bulls treat Disco as a pure AI-beta call and trade the shipment second-derivative — so they under-weight the consumable annuity, which is the least cyclical, highest-margin, most-monopolistic part of the business and which compounds mechanically with the entire installed base's wafer volume (AI and everything else). The market is mispricing Disco as "cyclical AI equipment" when the ~35% razor-blade stream is closer to a wafer-volume royalty that deserves a higher, steadier multiple. The bear's mirror-image blind spot: nobody is discounting hybrid bonding because it's a slow tail — but slow tails are exactly what a 50x multiple has no margin of safety against.
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.