Home/Reports/AI Photonics / Co-Packaged Optics (CPO) — US vs JP Value Chain DD

AI Photonics / Co-Packaged Optics (CPO) — US vs JP Value Chain DD

2026-05-14 03:30 · 14.8 KB

Date: 2026-05-14 | Time: 03:30 JST | Type: Explore — expand Light/Photonics theme with AI-CPO supply chain | Method: T-Glass (trace from global leader backwards, find Japan structural edge)


Trigger

This DD was triggered by a Chinese-language X post arguing that the total market cap of all optical-module companies combined is smaller than Micron alone, despite the optical-transceiver market growing from $16.5B (2025) to $26B (2026) — over 57% YoY, one of the fastest-growing semi sub-segments. The post identifies 7 US-listed picks across the AI photonics supply chain: $SOI, $TSEM, $LITE, $SIVE, $COHR, $GLW, $NOK. This DD maps those US picks to Japanese counterparts and compares valuations, moats, and catalysts.


Glossary

TermMeaning
CPOCo-Packaged Optics — silicon-photonics transceivers integrated directly into AI switch/GPU packages, replacing pluggable optics. Drives the 2027-2028 thesis.
NPONear-Packaged Optics — interim stage between pluggable and CPO; optics on the same PCB but not in the chip package (2026 H2).
EMLElectro-absorption Modulator integrated Laser — InP-based laser used in 100G/200G-per-lane optical transceivers. The bottleneck for 800G/1.6T pluggables.
InPIndium Phosphide — compound semiconductor substrate for laser sources in optical transceivers.
OTNOptical Transport Network — long-haul / metro / data center interconnect optical backbone.
APNAll-Photonic Network — NTT's IOWN initiative: replace electrical signals with optical end-to-end.
PECPhotonic-Electronic Convergence — NTT's chip-level CPO precursor; 2nd-gen prototype 51.2 Tbps board-to-board.
ZR / ZR+Coherent pluggable transceiver standards for 400G/800G long-haul DCI.
PAM44-level Pulse Amplitude Modulation — modulation scheme for 800G/1.6T transceivers.

The 7 US picks (X post) — mapped to JP equivalents

LayerUS Pick (X post)JP EquivalentJP Status
Silicon photonics substrate (SOI wafer)$SOINo JP pure-play (SUMCO 3436 makes silicon wafers but not pure SOI)Gap — no investable JP name
Silicon photonics foundry$TSEM (Tower)No JP foundry pure-play; NTT IOWN ecosystem has internal photonic chip foundry (not standalone)Covered indirectly via NTT
InP EML laser$LITE (Lumentum)6503 Mitsubishi Electric (200Gbps EML, POET 3.2T partnership)NEW — added in this DD
Vertical integrator (laser → module → system)$COHR (Coherent)6965 Hamamatsu Photonics (PMT + SiPM + cameras + NKT lasers, post-2024 vertical)already covered
CPO-era laser supply (InP arrays)$SIVE (Sivers)5802 Sumitomo Electric (60% global InP wafer share + FlexBeamGuide SiPh coupler)already covered
Optical fiber + connectors$GLW (Corning)5803 Fujikura + 5801 Furukawa Electric + 5802 Sumitomo Electricalready covered (3 names)
OTN backbone$NOK (Nokia + Infinera)6702 Fujitsu (1FINITY Ultra 1.2T) + 6701 NEC (submarine + Open APN)NEW — both added in this DD
AI photonics test equipment(no US pick named)6754 Anritsu (MP2110A for 800G/1.6T transceivers)NEW — added in this DD
CPO national strategy(no US pick — NVIDIA / Broadcom-led)9432 NTT (IOWN APN, Broadcom 2nd-gen PEC commercialization 2026)NEW — added in this DD

Complete Light Theme Value Chain (post-expansion, 12 companies)

Layer 1: Detection / Photon Counting

  • 6965 Hamamatsu Photonics — sole Western 20-inch PMT for Hyper-K
  • 6856 HORIBA — Raman + mass flow controllers

Layer 2: Light Sources

  • 6925 USHIO — UV/EUV metrology + cinema lamps
  • 6503 Mitsubishi Electric (NEW) — InP EML lasers for 800G/1.6T (POET 3.2T partnership)

Layer 3: Inspection & Test

  • 6920 Lasertec — EUV mask inspection (~100% global share)
  • 6754 Anritsu (NEW) — optical transceiver test (MP2110A BERTWave)

Layer 4: Optical Fiber & InP Substrates

  • 5803 Fujikura — fusion splicers + hyperscale fiber
  • 5801 Furukawa Electric — Lightera 13,824-count fiber + HTS REBCO
  • 5802 Sumitomo Electric — 60% global InP wafer share

Layer 5: Optical Transport Network (OTN)

  • 6702 Fujitsu (NEW) — 1FINITY Ultra 1.2T (JP $NOK equivalent)
  • 6701 NEC (NEW) — submarine cable (400,000+ km) + Open APN

Layer 6: National Strategy / All-Photonic Network

  • 9432 NTT (NEW) — IOWN APN + Broadcom 2nd-gen PEC commercialization 2026

Valuation comparison — US vs JP, mapped by layer

LayerUS PickUS Stock (recent)JP PickJP Stock (2026-05-14)JP Discount
InP EML$LITE (Lumentum)PE ~60x, Nvidia $2B equity stake6503 Mitsubishi ElectricPE 32.5x / Fwd 27.5x~50% cheaper
Vertical integration$COHR (Coherent)PE ~35x, Nvidia $2B equity stake6965 HamamatsuPE 50.4x (TTM trough)not cheaper — trough earnings
InP wafers(via $SIVE small cap)n/a meaningful5802 Sumitomo ElectricPE 34.0x / Fwd 23.9xdirect exposure at scale
Optical fiber$GLW (Corning)PE ~25-30x5803 FujikuraPE 87.7x (post-split, post-rally)priced fully — see errata
Optical fiber #2$GLW (Corning)PE ~25-30x5801 Furukawa ElectricPE 65.7x / Fwd 65.8xpriced fully
OTN$NOK (Nokia)Fwd PE 30.8x6702 FujitsuPE 18.9x / Fwd 19.0x~40% cheaper
Submarine + OTN$NOKn/a6701 NECPE 19.95x / Fwd 18.2x~40% cheaper
Test equipment(not US-listed)6754 AnritsuPE 45.4x / Fwd 37.5xpriced after +168% 52w
National CPO strategyn/a9432 NTTPE 11.9x / Fwd 11.6x / 3.6% yieldMASSIVELY underowned

Key inefficiencies (T-Glass observations)

1. NTT (9432) is the cleanest under-the-radar play. PE 11.9x with 3.6% dividend yield and exclusive IOWN/Broadcom 2nd-gen PEC commercialization slated for 2026 H2. Stock -2% 52w = zero photonics-rally premium priced in. If IOWN/Innovative Devices is ever carved out for IPO, the photonic slice gets re-rated to LITE/COHR multiples (5-10x current).

2. Fujitsu (6702) is the JP $NOK at 60% the multiple. 1FINITY Ultra Optical does 1.2T per wavelength (Nokia's PSE-6s does 1.2T) — comparable tech, half the valuation, no broker coverage on the photonics angle.

3. Mitsubishi Electric (6503) is the cheapest InP EML exposure globally. PE 32.5x vs Lumentum PE ~60x. POET partnership for 3.2T optical engines = real CPO supply-chain integration, not just narrative. Trade-off: conglomerate discount (defense, FA, HVAC dilute the photonics pure-play angle).

4. Anritsu (6754) is the "test-shovels" play already priced. +168% 52w means most of the easy money is gone. But every transceiver shipped tests on this — secular demand, no design risk. Hold not buy.

5. The optical fiber names (5801/5803) are PEAK-EARNINGS RISK after their +713%/+949% 52w rallies. See errata in light_companies.json — both had stock splits + are trading at PE 65-88x on company guidance. Trim, not add.


US picks NOT covered by JP equivalents

US PickWhy no JP equivalent
$SOI (Soitec — silicon-on-insulator substrate)No JP pure-play. SUMCO (3436.T) makes silicon wafers but not SOI; Shin-Etsu (4063.T) does some SOI as a sub-segment but is a much broader semiconductor materials play.
$TSEM (Tower Semiconductor — silicon photonics foundry)No JP standalone foundry. NTT (9432) has internal IOWN photonic chip foundry capacity but it's strategic R&D, not a commercial foundry.
$SIVE (Sivers — small-cap CPO laser arrays)No JP small-cap photonics pure-play. The CPO laser angle goes through Sumitomo Electric (InP wafers — 5802) or Mitsubishi Electric (EMLs — 6503).

These are GAPS in JP coverage — investors wanting the silicon-photonics-substrate or pure-foundry exposure must use the US names directly.


T-Glass framework — applied to AI photonics

Step 1: Demand signal

  • 800G/1.6T optical transceiver market: $16.5B (2025) → $26B (2026), 57% YoY (per X post)
  • 800G/1.6T market: $2.5B (2024) → $10.5B (2029), 33% CAGR (per Mitsubishi Electric IR)
  • NVIDIA NVL72: optical modules = 20% of total rack value
  • Meta OFC 2026: CPO proven more reliable + cheaper + lower-power than pluggables
  • NVIDIA GTC 2026: CPO planned for Scale-Up interconnect 2027-2028

Step 2: Supply chain trace (3 phases)

Phase 1 (now): Pluggable 800G → 1.6T → 3.2T transceivers

→ Mitsubishi Electric 6503 (EMLs) + Anritsu 6754 (test)

Phase 2 (2026 H2): Near-Packaged Optics (NPO)

→ Fujitsu 6702 + NEC 6701 (OTN integration)

Phase 3 (2027-2028): Co-Packaged Optics (CPO)

→ NTT 9432 (IOWN PEC) + Sumitomo Electric 5802 (InP wafers)

Step 3: Bottleneck hunt

  • InP EML lasers — Lumentum's monopoly; Mitsubishi Electric is the only credible non-US-aligned alternative
  • Test equipment — Anritsu + Keysight duopoly for 800G/1.6T
  • InP wafers — Sumitomo Electric 60% share = the upstream choke
  • Photonic-electronic convergence chips — NTT + Broadcom partnership (2026 H2 commercialization)

Step 4: Monopoly check

TickerLayerMoat rating
6920 LasertecEUV mask inspection★★★★★
6965 Hamamatsu20-inch PMT★★★★★
5802 Sumitomo ElectricInP wafers (60%)★★★★
9432 NTTIOWN + Broadcom CPO★★★★
6503 Mitsubishi ElectricInP EML + POET 3.2T★★★★
6754 AnritsuOptical test★★★★
6856 HORIBARaman + mass flow★★★★
5801 / 5803 / 6925 / 6702 / 6701various★★★

Step 5: Capacity timeline

  • Pluggable 1.6T: design-in 2026, ramp 2027 — Mitsubishi EML, Anritsu test
  • NTT 2nd-gen PEC commercialization: 2026 H2 (Broadcom partnership announced)
  • Hyper-Kamiokande PMT delivery (running): 20,000 units, data start 2028
  • CPO mass deployment: 2027-2028 per Nvidia GTC roadmap

Step 6: Catalyst mapping

  • 2026 H2: NTT-Broadcom 2nd-gen PEC commercialization — single biggest near-term catalyst
  • OFC 2027 (March): Next-gen 3.2T standards announcements
  • NVIDIA quarterly capex: 800G/1.6T transceiver shipment guidance
  • Q2/Q3 2026 JP earnings: Anritsu, Mitsubishi Electric, NTT, Fujitsu prints

Step 7: Cycle position

  • Anritsu: peak-rally (+168% 52w), PE 45x — late cycle, hold-not-buy
  • Mitsubishi Electric: mid-cycle, +126% 52w, PE 32x — reasonable
  • Fujikura / Furukawa Electric: peak-rally (+713% / +949% 52w) — TRIM
  • Fujitsu / NEC: early cycle, PE 19x, modest 52w returns — best risk/reward
  • NTT: pre-cycle, -2% 52w, PE 12x — most asymmetric upside if IOWN gets recognized

Trade construction (T-Glass weighting)

Position typeAllocation per namePicks
HIGH conviction12-18%NTT 9432, Mitsubishi Electric 6503, Lasertec 6920
Picks-and-shovels7-10%Anritsu 6754
Cheap OTN exposure7-10% eachFujitsu 6702, NEC 6701
Pre-cycle InP wafers12%Sumitomo Electric 5802
TRIM (over-extended)0-5% eachFujikura 5803, Furukawa Electric 5801 (peak-earnings flag)
Hold (legacy thesis)5% eachHamamatsu 6965, Ushio 6925, HORIBA 6856

Risks

  • Hyperscaler capex pullback — if NVIDIA / Meta / Microsoft / Google cut 2026 capex, 800G/1.6T transceiver demand softens immediately
  • CPO delays — if NTT-Broadcom 2nd-gen PEC slips beyond 2026 H2, the NTT thesis de-risks substantially
  • Currency — JP photonics names earn meaningful USD revenue; yen strengthening compresses OPM
  • Stock-split timing data quality — see prior errata; Fujikura/Furukawa had splits that StockAnalysis lagged on
  • US-JP tech decoupling — if US export controls extend to JP photonics, Sumitomo Electric InP / Mitsubishi EML could face customer-side restrictions

Data Provenance

All on-chain stock data sourced 2026-05-14:

  • Live OHLCV: J-Quants API v2 /equities/bars/daily — Tier 1
  • Listed-info: J-Quants API v2 /equities/master — Tier 1
  • Valuation metrics (PE, fwd PE, market cap): StockAnalysis.com per ticker — Tier 2
  • Cross-validation note: Yahoo Finance JP used for company-guidance Fwd PE on names with recent splits (Fujikura, Furukawa Electric) per prior errata
  • Supply chain evidence: every entry in light_companies.json has source_url field linking to the actual press release, IR page, or trade-press article

Specific URLs cited inline in light_companies.json:


Outstanding follow-ups

1. Run the actual research pipeline scriptsscripts/research_explore.sh would spawn a Claude subprocess; since I ran the workflow inline in this session, no subprocess invocation needed. If user wants cron-driven future updates, fix the script's hardcoded /Users/aqua-bancho/ path (typo or different user) and add it to launchd / GitHub Actions schedule.

2. JSON validation guardrail added in separate file scripts/validate_data_json.py + GitHub Actions step — prevents another silent ev_companies.json-style base64 breakage.

3. Per-stock deep DDs for new names (Mitsubishi Electric, Anritsu, NTT, Fujitsu, NEC) — currently each has thesis/risk/trigger/supply-chain entries in light_companies.json but no standalone DD report. Recommend deep DD on NTT 9432 + Mitsubishi Electric 6503 first (highest-conviction picks).

4. Stock-split watch list — Fujikura (6:1, Apr 2026), Furukawa (10:1, pending). Add automated split-detection cross-check in research pipeline to avoid stale-PE issues flagged in prior errata.