AI/GPU Supply Chain — Japan's Hidden Infrastructure Layer
8 stocks powering the physical layer of AI
2026-04-15 · 8 stocks · Claude Opus deep DD. J-Quants + StockAnalysis. Nidec removed (accounting investigation).
Global AI GPU demand is reshaping Japan's semiconductor materials and component industry. NVIDIA GB200 Blackwell requires 30,000 MLCCs, Ibiden FC-BGA substrates, Resonac bonding materials, and Fujikura optical cables per server rack. Japan controls 90% of global photoresist (TOK), dominates CMP slurries (Resonac), and supplies critical packaging substrates (Ibiden). Total hyperscaler capex exceeds $600B in 2026. HBM demand growing 5-10x. Nidec (6594.T) removed due to active accounting investigation.
Portfolio Overview
| # | Company | Theme | Conv | Wt | PE | Fwd PE | PB | ROE | OpMar | D/E | DY | FCF |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Fujikura Ltd5803.T | AI/GPU Supply Chain | HIGH | 12% | 11.00 | 10.90 | 3.04 | 27.7% | 16.6% | N/A | 0.5% | N/A |
| 2 | Ibiden4062.T | AI/GPU Supply Chain | MEDIUM | 8% | 67.00 | 74.80 | 5.04 | 7.5% | 14.9% | N/A | N/A | -¥8B |
| 3 | TDK CORP6762.T | AI/GPU Supply Chain | MEDIUM | 8% | 17.90 | 22.70 | 2.05 | 11.5% | 12.4% | N/A | 1.5% | +¥105B |
| 4 | RESONAC HOLDINGS CORPORATION4004.T | AI/GPU Supply Chain | MEDIUM | 7% | 84.10 | N/A | 3.49 | 4.0% | 3.5% | N/A | N/A | +¥43B |
| 5 | Tokyo Ohka Kogyo4186.T | AI/GPU Supply Chain | MEDIUM | 6% | 34.70 | N/A | 5.09 | 13.8% | 20.0% | N/A | N/A | N/A |
| 6 | Nitto Denko6988.T | AI/GPU Supply Chain | MEDIUM | 6% | 15.50 | 16.00 | 1.96 | 12.8% | 18.8% | N/A | 1.9% | +¥49B |
| 7 | JCU Corporation4975.T | AI/GPU Supply Chain | MEDIUM | 5% | 17.70 | 18.20 | 3.01 | 17.2% | 41.3% | N/A | N/A | N/A |
| 8 | Murata6981.T | AI/GPU Supply Chain | LOW | 5% | 37.90 | 35.80 | 3.01 | 8.0% | 14.8% | N/A | 1.4% | +¥165B |
Portfolio Construction
Stock-by-Stock Analysis
Fujikura Ltd
5803.THIGHCoreAI/GPU Supply Chain · Weight: 12%
Why this stock
The hottest AI infrastructure play in Japan — up 1,400% in 2 years on insatiable datacenter optical cable demand. ¥300B investment to triple optical fiber/cable capacity under Japan-US framework agreement. PE 11.0x with 27.7% ROE is extraordinary value for this growth trajectory. FY3/25 revenue +22.5% to ¥979B. Supplies high-count optical cabling for AI datacenter interconnects (GPU-to-GPU communication).
Why 12%
12% as HIGH conviction Core. PE 11.0x with 27.7% ROE is the best quality-to-value ratio in the entire AI supply chain. ¥300B capacity tripling investment shows multi-year growth visibility. Stock near ATH but PE 11x means earnings have grown faster than price. Only risk is that +1,400% move may have front-loaded returns.
What could go wrong
1) Stock up 1,400% in 2 years — momentum reversal risk if AI capex cycle peaks. 2) ¥300B capacity investment is massive — execution risk on tripling production. 3) Competition from Furukawa Electric and Sumitomo Electric in optical cables.
Monitoring trigger
If AI datacenter capex guidance from hyperscalers (AWS/MSFT/GOOG) slows >20%, TRIM to 8%. If optical cable order backlog extends beyond 2 years, ADD to 15%. Watch quarterly revenue growth — if <15% YoY, reassess.
What the market misses
PE 11.0x for a 27.7% ROE company growing revenue 22.5% with a ¥300B capacity expansion plan. The market may be pricing in cyclicality that does not apply — AI datacenter interconnect demand is structural, not cyclical. Japan-US framework agreement provides government backing.
Ibiden
4062.TMEDIUMSatelliteAI/GPU Supply Chain · Weight: 8%
Why this stock
Primary FC-BGA substrate supplier for NVIDIA data center GPUs via CoWoS packaging. TSMC has taken a stake in Ibiden. ¥500B three-year capex plan starting FY2026 to expand IC substrate capacity. AI server substrate revenue ~3x YoY in FY2024. Every advanced AI GPU needs an Ibiden substrate — this is a physical bottleneck in the NVIDIA supply chain.
Why 8%
8% as MEDIUM conviction Satellite. PE 67.0x is expensive — the AI substrate thesis is priced in. But ¥500B capex plan and TSMC stake validate the structural demand. ROE 7.5% is below average. Held at MEDIUM because valuation limits upside — need to see earnings catch up to justify PE. Forward PE 74.8x even higher suggests earnings dilution from capex.
What could go wrong
1) PE 67x is extremely expensive — vulnerable to any demand slowdown. 2) FCF -¥8B (negative) from massive capex. 3) Forward PE 74.8x > trailing — market expects earnings to dip. 4) If TSMC diversifies CoWoS substrate sources, Ibiden loses pricing power.
Monitoring trigger
If IC substrate revenue growth decelerates below 20% YoY, TRIM to 5%. If FCF turns positive despite capex, ADD to 10%. If TSMC increases stake or signs exclusivity deal, ADD to 10%.
What the market misses
The ¥500B capex plan is the largest substrate expansion in history — Ibiden is betting the company on AI GPU demand being structural. TSMC taking a stake is a validation that no other substrate maker received.
TDK CORP
6762.TMEDIUMCoreAI/GPU Supply Chain · Weight: 8%
Why this stock
World #1 MLCC maker. AI servers consume 30,000 MLCCs each vs 1,000 per smartphone — 30x volume multiplier. Murata raised AI server MLCC CAGR forecast to 30%. Exploring price hikes on AI server MLCCs due to tight supply. New ¥47B MLCC factory completed April 2026. Price hikes confirmed effective April 1. At 52-week high — demand is pulling the stock up.
Why 8%
8% as MEDIUM conviction Core. PE 17.9x is fair for a quality compounder (ROE 11.5%, FCF ¥105B). The 30x AI server MLCC volume multiplier is a structural growth driver. Forward PE 22.7x suggests near-term earnings pressure from tariffs (-24% NI impact from 15% US tariff). Position for long-term AI server ramp, tolerate near-term volatility.
What could go wrong
1) 15% US tariff impact = -24% net income hit. 2) Forward PE 22.7x > trailing 17.9x — earnings expected to dip. 3) Front-loaded H1 demand could create H2 cliff. 4) Smartphone MLCC demand weak offsets AI strength.
Monitoring trigger
If AI server MLCC revenue exceeds 15% of total at earnings, ADD to 10%. If US tariffs escalate beyond 15%, TRIM to 5%. If price hikes stick and margins expand, HOLD.
What the market misses
30,000 MLCCs per AI server vs 1,000 per smartphone means AI datacenter buildout drives 30x the volume intensity. Even modest AI server penetration creates massive incremental demand that smartphones cannot match. Price hikes are being accepted — pricing power is real.
RESONAC HOLDINGS CORPORATION
4004.TMEDIUMSatelliteAI/GPU Supply Chain · Weight: 7%
Why this stock
Global #1 in ceria CMP slurries and etching gases — core materials for HBM hybrid bonding and CoWoS processes. Won TSMC 2025 Excellent Performance Award for advanced packaging materials. Established local production for TSMC Kumamoto. Every HBM chip stack uses Resonac bonding materials — the picks-and-shovels play for the memory AI boom.
Why 7%
7% as MEDIUM conviction Satellite. PE 84.1x is very expensive — reflects early-cycle HBM demand ramp where revenue is just starting to flow. Quality is low (ROE 4.0%, OpMar 3.5%) but this is a turnaround story from former Showa Denko. The TSMC award and Kumamoto localization are concrete catalysts. Position for HBM volume ramp 2026-2028.
What could go wrong
1) PE 84.1x — one of the most expensive stocks in our universe. 2) ROE 4.0% and OpMar 3.5% are weak — turnaround not yet proven. 3) HBM demand could be front-loaded if AI capex cycle peaks. 4) Samsung and SK Hynix could develop alternative bonding materials.
Monitoring trigger
If operating margin exceeds 8% at next earnings (turnaround accelerating), ADD to 9%. If PE stays above 80x for 2 more quarters without margin improvement, TRIM to 5%.
What the market misses
Resonac's CMP slurries and etching gases are consumed in every HBM production cycle — as HBM volume grows 5-10x, Resonac's materials revenue scales proportionally. The TSMC Kumamoto localization locks in a geographic moat for Japan-based chip production.
Tokyo Ohka Kogyo
4186.TMEDIUMSatelliteAI/GPU Supply Chain · Weight: 6%
Why this stock
Japan controls ~90% of global photoresist supply. TOK is a top player — every AI chip at 5nm and below requires EUV photoresist. Record FY2025: revenue +17.9% to ¥237B, operating income +43.2%. ¥20B Korea plant expansion (3-4x capacity). Strategic partnership with Irresistible Materials for next-gen EUV. Operating margin 20.0% — high-quality materials monopoly.
Why 6%
6% as MEDIUM conviction Satellite. PE 34.7x is expensive but justified by 20% OpMar and Japan's 90% photoresist monopoly. ROE 13.8% is solid. The moat is nearly impenetrable — photoresist requires decades of know-how and customer qualification. Growth is tied to advanced node adoption which is accelerating for AI chips.
What could go wrong
1) PE 34.7x is not cheap. 2) Korea plant expansion (¥20B) has execution risk. 3) Export controls on photoresist to China could limit TAM. 4) If EUV adoption slows (unlikely), demand plateaus.
Monitoring trigger
If Korea plant comes online and revenue accelerates >25% YoY, ADD to 8%. If export controls tighten and revenue guidance cut, TRIM to 4%.
What the market misses
Japan's 90% photoresist monopoly is the most durable supply chain moat in semiconductors — even TSMC cannot substitute it. Every new AI chip fab increases photoresist consumption. TOK's next-gen EUV partnership positions it for 2nm and beyond.
Nitto Denko
6988.TMEDIUMCoreAI/GPU Supply Chain · Weight: 6%
Why this stock
PE 15.5x with 12.8% ROE and 18.8% OpMar — best value play in AI materials. Supplies optical films for displays AND flexible circuits/materials for HBM packaging. FCF ¥49B provides self-funding. Dividend yield 1.9%. Less hyped than Fujikura/Ibiden but similarly positioned in the AI material supply chain. Dual exposure: consumer electronics (stable) + AI/HBM (growth).
Why 6%
6% as MEDIUM conviction Core. PE 15.5x with 18.8% OpMar and FCF ¥49B is genuinely cheap for this quality. Less AI hype than Fujikura/Ibiden means less momentum risk. Diversified revenue base provides downside cushion. The HBM materials angle is growing but not yet dominant — this is a quality compounder with AI optionality.
What could go wrong
1) HBM/AI materials revenue not yet dominant — still a diversified materials company. 2) Display business faces cyclical pressure. 3) Less direct AI exposure than Fujikura or Ibiden. 4) Korean competition in optical films.
Monitoring trigger
If AI/semiconductor materials segment exceeds 25% of revenue at earnings, ADD to 8%. If display segment drags total growth below 5%, HOLD at 6%.
What the market misses
PE 15.5x with 18.8% operating margin and ¥49B FCF — the market hasn't re-rated Nitto Denko for AI like it has Fujikura (+1,400%) or Ibiden (PE 67x). Similar quality, much cheaper, with downside protection from the diversified base.
JCU Corporation
4975.TMEDIUMSatelliteAI/GPU Supply Chain · Weight: 5%
Why this stock
Near-monopoly in via-filling plating chemistry essential for high-density interconnects in advanced packaging. Operating margin 41.3% — highest in our entire universe. Benefits from a complexity multiplier: as AI chips get more complex (more layers, finer pitch), JCU sells more product per chip. Directly levered to Ibiden/Shinko capex cycles. Small cap (¥154B) with massive operating leverage to AI substrate buildout.
Why 5%
5% as MEDIUM conviction Satellite. PE 17.7x is reasonable for 41.3% OpMar and 17.2% ROE — exceptional quality metrics. Small cap (¥154B) limits position size. The complexity multiplier thesis is unique — more advanced chips = more JCU revenue per unit. Risk is concentration in a niche chemistry that could be disrupted.
What could go wrong
1) Small cap ¥154B — low liquidity, volatile. 2) Niche product concentration — if alternative via-filling chemistry emerges, moat evaporates. 3) Customer concentration — heavily dependent on Ibiden and former Shinko Electric. 4) Limited diversification beyond plating chemicals.
Monitoring trigger
If Ibiden announces further capex expansion, ADD to 7% (JCU directly benefits). If operating margin drops below 35%, reassess moat. If a competitor enters via-filling market, EXIT.
What the market misses
41.3% operating margin in a ¥154B company that is a hidden monopoly supplier to every AI substrate maker. The complexity multiplier means revenue per chip grows as AI chips advance — a built-in growth accelerator that doesn't depend on unit volume alone.
Murata
6981.TLOWSatelliteAI/GPU Supply Chain · Weight: 5%
Why this stock
AI revenue already 10% of total and growing 25-30% CAGR. New uPOL power modules specifically for AI GPU power delivery. MLCCs and passive components for datacenter power systems. PE 37.9x is expensive but AI revenue acceleration could justify re-rating. Diversified base (batteries, sensors, storage) provides stability while AI segment scales.
Why 5%
5% as LOW conviction Satellite. PE 37.9x is expensive for 8.0% ROE. AI revenue at 10% is still small. But 25-30% CAGR in AI segment and new uPOL power modules show directional momentum. FCF ¥165B is strong. Position for AI power component demand growth, not for current value.
What could go wrong
1) PE 37.9x with 8.0% ROE — overvalued on current fundamentals. 2) AI is only 10% of revenue — takes years to become dominant. 3) Battery segment facing EV headwinds. 4) Competition from Murata in AI server passives.
Monitoring trigger
If AI segment exceeds 15% of revenue at next earnings, ADD to 7%. If uPOL wins design slots at NVIDIA/AMD, ADD to 7%. If PE expands beyond 45x without AI acceleration, TRIM to 3%.
What the market misses
The uPOL power module is specifically designed for AI GPU power delivery — each NVIDIA Blackwell rack needs precision power management that TDK is uniquely positioned to supply. The 25-30% CAGR in AI segment compounds on an already large base.
Data Corrections (Errata)
| Ticker | Metric | Original | Actual | Source | Impact |
|---|---|---|---|---|---|
| 6594.T | governance | LOW conviction | AVOID | HBM research agent | Removed from portfolio — active accounting investigation + shareholder lawsuit |
Methodology
Claude Opus deep DD. J-Quants + StockAnalysis. Nidec removed (accounting investigation).
AI-generated for research purposes only. NOT investment advice. Generated 2026-04-15.