Semiconductors
PrivateA twice-cyclical Japanese power-semi laggard that just wrote off its SiC bet (¥163B impairment) and simultaneously round-tripped to an all-time high on an AI-server SiC re-rating — the story got re-priced two quarters before the P&L can confirm it, and at ¥5,950 the tape now demands the AI-power pivot actually shows up in the numbers. WATCHING, tilting BEARISH into the print.
Research
The verdict
A twice-cyclical Japanese power-semi laggard that just wrote off its SiC bet (¥163B impairment) and simultaneously round-tripped to an all-time high on an AI-server SiC re-rating — the story got re-priced two quarters before the P&L can confirm it, and at ¥5,950 the tape now demands the AI-power pivot actually shows up in the numbers. WATCHING, tilting BEARISH into the print.
ROHM is a ~¥480B-revenue ($3.2B) integrated device manufacturer (IDM) founded in Kyoto in 1958 by Kenichiro Sato — who started by inventing a small precision resistor in 1954 and built it into a global semiconductor house ``. It designs, fabricates, and sells its own chips (IDM, not fabless) across four reported segments:
The business model in one line: ROHM sells the "energy-saving + downsizing" analog and power layer that sits between a battery/grid and a motor/processor — the boring-but-critical silicon that switches and regulates power. Its self-declared vision is "power and analog solutions" ``.
Contract structure: classic broad-line semi — no take-or-pay, no single mega-contract; revenue is a long tail of automotive OEM/Tier-1 design-wins (multi-year, sticky once designed-in) plus catalog/distribution volume. Payment terms are ordinary trade, not recurring/subscription. Concentration is by end-market, not by single customer (see Lens 4).
ROHM is unusually vertically integrated for its size — a deliberate strategic choice that is the whole SiC story. Named map, upstream → downstream:
. For silicon wafers, standard Japanese/global suppliers (Shin-Etsu, SUMCO). Gold/copper for packaging — a real cost line (material costs +¥12.8B YoY partly on **gold prices**) .Chokepoints / single-source: ROHM's own SiCrystal substrate line is both a moat (self-supply) and a risk (it built 6-inch capacity into a market that moved to 8-inch and then oversupplied — the direct cause of the impairment). Distribution runs through the big global cats (Digi-Key, Mouser, Arrow) plus direct.
Where the moat is real:
. Owning SiCrystal (substrate) + wafer + device + the analog control IC around it is a genuinely rare full-stack position — only Infineon, STMicro, Wolfspeed and onsemi sit in the same tier; the top-5 control **>90% of SiC revenue** .Where the moat is weaker than bulls think:
By business segment (Q3 of the year ending March 2026, i.e. Oct–Dec 2025 quarter — the cleanest recent split) ``:
| Segment | Quarterly revenue | YoY | Read |
|---|---|---|---|
| Semiconductor Devices (Discrete/Power incl. SiC) | ¥55.8B | +19.2% | The growth engine — SiC-for-inverters + AI-server pickup |
| LSI (analog/mixed-signal ICs) | ¥55.1B | +6.7% | Steady recovery, the ballast |
| Modules | ¥7.7B | −1.9% | Structural laggard |
| Others | (residual) | — | Passives/foundry/printheads |
Full-year (ended March 2025) segment growth ``: LSI +7.1%, Semiconductor Devices +9.7%, Modules +3.0%, Others +3.6% — a broad, unspectacular recovery masking the SiC write-down beneath it.
By end-market (year ended March 2025) ``: Automotive 49.2%, Consumer 21.9%, C&S/Components-&-Systems 12.3%, Industrial 12.7%, Communication 3.8%. The trajectory that matters: automotive has climbed from ~45% (FY-ended-Mar-2023) to ~49% — ROHM is levering itself into EV/auto power right as the EV growth curve flattened. That is the crux of the whole thesis.
No segments.csv on the shelf — every number here is , not .
The defining print — year ended 31 March 2025 ``:
| Metric | Year ended Mar-2025 | YoY |
|---|---|---|
| Net sales | ¥481.1B | +7.3% |
| Operating profit | ¥10.8B | +2.3% |
| Operating margin | ~2.2% `` | — |
| Net income (loss) attrib. | −¥158.4B | (from prior-yr profit) |
| SiC-related impairment (extraordinary) | ¥163.2B | — |
| Dividend | ¥50/share | maintained |
The story is unmistakable: operationally roughly break-even, but a ¥163.2B SiC impairment turned it into a ¥158B net loss — one of the larger single write-downs by a Japanese semi in the cycle. ROHM's own stated cause: "slowdown in BEV market growth and the rise of Chinese manufacturers," plus "excess investment" and declining demand for 6-inch substrates ``. Translation: ROHM built SiC capacity (6-inch) for an EV ramp that decelerated, into a market that then oversupplied and moved to 8-inch — and wrote off the mismatch.
Current year (ending 31 March 2026) — mid-recovery, but guidance walked down twice:
Balance-sheet flags (as of 31 Mar 2025) ``: Total assets ~¥1.12T, total equity ~¥915B, total debt only ~¥40B → net cash position, equity ratio very high (~80%). This is the single most important defensive fact: ROHM took a ¥163B hit and its balance sheet barely flinched. A fortress balance sheet is why the equity survived the write-down.
Market reaction: not a crash — the opposite. The stock is at an all-time high (see Lens 8). The impairment was treated as a kitchen-sink event that clears the SiC overhang; the market is now paying for the recovery + AI-server optionality.
No transcripts on the shelf; sentiment reconstructed from earnings briefs + IR materials ``. The tonal arc across the last ~4 quarters:
Phrases that appeared: "800VDC / HVDC," "AI server," "battery backup unit (BBU)," "power + analog." Phrases that receded: the aggressive "6.5×/35× SiC capacity" ramp targets from the 2023 Toshiba announcement — quietly de-emphasised. The stopped-saying is as informative as the started-saying: ROHM is re-badging the same SiC capacity from an EV story to an AI story.
Global power/analog peers. Multiples are `` with source/date; where I could not source a figure it is n/a (never fabricated).
| Company | Ticker | Mkt cap (USD) | Fwd P/E | EV/EBITDA | ROE | Div yield |
|---|---|---|---|---|---|---|
| ROHM | 6963.T | ~$15.5B (¥2.30T, 2026-07-03) | at-loss TTM; ~45–140x normalized `` | 23.5x `` | −4.5% (TTM, loss year) `` | ~0.8–1.4% `` |
| Infineon | IFX.DE | ~$119.7B `` | 37.0x `` | 27.4x `` | n/a | n/a |
| STMicro | STM | ~$69.7B `` | 39.8x `` | 21.2x `` | 0.9% `` | 0.5% `` |
| onsemi | ON | ~$35.5B `` | 35.1x `` | 17.8x `` | 7.5% `` | n/a |
| Renesas | 6723.T | ~$52.6B `` | 23.9x `` | 23.8x `` | n/a | n/a |
Read: ROHM is the smallest of the credible power-semi set — ~1/8th of Infineon, ~1/2 of Renesas. On EV/EBITDA (23.5x) it sits mid-pack, richer than onsemi/STM, near Renesas, cheaper than Infineon — but that is on depressed EBITDA (a near-break-even operating year). Its market cap doubled in a year (stock +241%), so the "cheap Japanese small-cap" framing is gone; ROHM is now priced as an AI-power growth name, not a cyclical laggard. ROE −4.5% is the ugly number — even in a loss year, it flags that ROHM's capital efficiency has lagged peers for years, not just this cycle.
ROHM has round-tripped violently. Key moves ``:
What the pattern reveals: ROHM's stock is now a pure-play sentiment vehicle on the SiC narrative. It sold off through the EV-SiC disillusionment (2024–25), bottomed on the impairment, and has exploded on the AI-server-SiC re-rating (2026). The market reacts to (1) the SiC demand narrative (which end-market is buying), (2) sell-side price-target resets (the JPM double), and (3) forex (¥/$). It does NOT trade on the current P&L — which is still barely profitable. This is a story-ahead-of-numbers stock.
): the founding family's direct stake is now low-single-digit/below-threshold, but influence persists through the **Rohm Music Foundation (10.76% holder)** and the conservative Kyoto-IDM culture . Top holders: Master Trust Bank of Japan 14.84%, Rohm Music Foundation 10.76%, foreign ~36.4%, treasury ~6.3% ``. No single controlling shareholder.Web-only; no filings on shelf to line-check. Accounting-risk read is directional ``.
Regulatory findings (required sub-section). Per regulatory/regulatory-findings.md: ROHM has no SEC CIK — not an SEC filer, so no EDGAR LR/AAER search is possible (total_sec_findings: 0, which reflects non-applicability, not a clean SEC audit). Non-SEC web search surfaced one material historical finding:
Verdict: No going-concern or fraud signal. The books tell a clean story of a bad capital decision — the write-down is honest. One resolved historical antitrust settlement (capacitors). Verified via web search + the regulatory-findings file (SEC non-applicable) as of 2026-07-07.
Building bottom-up from the current run-rate. **All outputs , inputs labeled. No `forecast.ts` logged (watchlist/unattended run — per skill rules, only log a committed base case).** Fiscal years end 31 March; shares outstanding ~386M (ex-treasury) .
Anchor: year ending Mar-2026 tracking to ~¥460B revenue, ¥9B net income (¥23 EPS) on guidance ``; but 9M already did ¥14.8B net / ¥38.4 EPS, implying a soft Q4 or conservative guide.
; op margin recovers to ~6% as SiC quality-assurance costs normalize and Kunitomi (April-2026 start) absorbs → net income ~¥28B → **EPS ~¥72** . FY-end-Mar-2028 revenue ¥525B, net ~¥40B → EPS ~¥104 ``.The number that matters: at ¥5,950, on base-case ~¥72 FY-Mar-2027 EPS, ROHM trades ~83× forward earnings ``. Even on a generous bull ¥145 EPS two years out, that's ~41×. The stock is priced for the bull case to be largely correct. There is little valuation cushion if the AI-server SiC ramp slips or a second SiC impairment lands.
Bull case. ROHM is a rare full-stack SiC IDM (owns the substrate via SiCrystal) with a fortress net-cash balance sheet, that has already taken its pain (¥163B kitchen-sink impairment clears the EV-SiC overhang) and pivoted its SiC capacity toward the fastest-growing power market on earth — AI-server / data-center HVDC, where its 750V SiC MOSFET (SCT4013DLL) is already designed into battery-backup units for ±400V/800VDC architectures, with mass production ramping Q2–Q3 2026 and volume peaking 2027 ``. Automotive is recovering (+13.9% YoY), the new CEO has a clean slate, and 92GW of incremental AI-data-center power demand by 2027 is a secular tailwind that needs exactly ROHM's product. If SiC re-rates as an AI-power story rather than an EV story, the earnings base can triple by 2028 — which is what JPM's doubled ¥8,000 target and Simply Wall St's ¥543B-revenue/¥54B-earnings 2028 path encode.
Bear case (2–3 permanent-impairment risks). (1) SiC is commoditizing, and ROHM is #5. Substrate prices collapsed >20% below $500/wafer and now sit near $400 or cost, driven by Chinese overcapacity backed by $15B of state subsidies and BYD-style localization ``. Being the smallest of the credible players in a price war is a structural disadvantage — the ¥163B impairment may be the first, not the last. (2) The AI-server SiC TAM is real but early and contested — mass production is Q2–Q3 2026 with volume "after 2027," and Wolfspeed, STMicro and onsemi are chasing the same BBU/PSU sockets; ROHM has no disclosed anchor customer yet. (3) Expectations are stretched: the stock is at an all-time high, +241% in a year, ~83× base-case forward EPS, on ~2% operating margins — a story fully priced before the P&L confirms it.
Pre-mortem (18 months out, thesis broke): It's early 2028. AI-server SiC design-wins went to Infineon/Wolfspeed at 8-inch scale; ROHM's 6-inch legacy and sub-scale cost position lost the price war; Chinese SiC took auto share; a second SiC impairment landed; EV stayed soft; the stock round-tripped from ¥5,950 back toward ¥3,000 as the AI-power narrative failed to show up in margins. The tell in hindsight: the market re-rated on a narrative pivot (EV→AI) using the same capacity that had just been written off, and paid an all-time-high multiple for optionality that a bigger, cheaper rival captured instead.
Are multiples too high? On current earnings, unambiguously yes — ~83× base-case forward EPS with 2% operating margins is a growth multiple on a cyclical laggard. Justified only if the AI-power SiC ramp is both large and ROHM-captured.
Contrarian view (what the market refuses to see): The bulls are treating ROHM's SiC capacity as an asset re-pointed at AI. But that capacity was impaired for a reason — it was the wrong node (6-inch) at the wrong scale (#5) at the wrong cost. Re-labeling it "AI-server SiC" doesn't fix the structural cost/scale gap versus Infineon. The market has priced the pivot's upside while ignoring that the same weakness that caused the write-down travels with the capacity into the new market.
Dismantling the bull case. Revenue concentration: ~49% automotive into a decelerated EV cycle — the exact exposure that just caused a ¥163B write-down; if EV SiC stays soft and AI-server volume is "after 2027," there's an air-pocket. Moat weaker than bulls think: the "world-first SiC MOSFET" pedigree is 15 years old; the current fight is about 8-inch scale and cost, where ROHM is sub-scale vs Infineon/STMicro and undercut by subsidized Chinese substrate. Owning SiCrystal is a double-edged sword — it's why ROHM couldn't just walk away from the capacity it had to impair. Most dangerous competitor bulls underestimate: not Infineon (obvious) but Chinese SiC substrate/device makers — they turned SiC into a commodity in 18 months and are localizing the auto supply chain ROHM depends on. Worst capital-allocation move: the ¥289B SiC build-out that produced a ¥163B impairment before a yen of profit — a textbook capacity-ahead-of-demand error, now owned. Assumptions that must hold for ¥5,950: (a) AI-server SiC is large and ROHM wins meaningful share, (b) no second SiC impairment, (c) auto SiC re-accelerates, (d) margins climb from ~2% to low-double-digits. If growth disappoints 20–30%: on ~83× base EPS, a miss re-rates the multiple and the earnings — a plausible path back toward ¥3,000–3,500 (near JPM's old ¥3,900 target). Single scenario that permanently impairs: SiC becomes a fully commoditized, Chinese-dominated component (like solar cells did) and ROHM is stuck as a sub-scale #5 with an impaired asset base and structurally low ROE — a value trap wearing an AI-growth costume.
A fortress-balance-sheet specialty-analog foundry that has been repriced overnight from a $45 cyclical into a $250 AI-silicon-photonics growth story — the operational inflection is real (Q1'26 op-profit +96%, $1.3B of 2027 SiPho contracts, $290M prepaid), but at ~66x forward EPS the stock already discounts most of the 2028 model, and a fresh GlobalFoundries patent-injunction suit at the ITC is an asymmetric, under-priced tail risk. Long the business, cautious on the multiple.
A real GaN/InnoSwitch franchise on a depressed-earnings base that the market has already re-rated +116% YTD onto a 2027-2028 AI-datacenter dream — at ~9x EV/sales and ~54x forward P/E with insiders selling and zero datacenter revenue today, the AI optionality is fully priced; WATCHING for the cyclical-recovery print, not the narrative.