Murata (6981) vs Musashi (7220) — The AI Power-Delivery Stack

One line: These are not competitors. 6981 (MLCC) and 7220 (Hybrid SuperCapacitor) sit at opposite ends of the power-delivery frequency spectrum and are complements enforced by physics — Murata smooths the nanoseconds at the die, Musashi smooths the seconds at the rack.

Built 2026-06-03 from two Substack sources (Made in Japan + Irrational Analysis) plus a claude.ai "Murata + PDN" first-principles thread. Companion to 6981, 7220, and the upstream-materials work in mlcc-sector-2026-05-23.

The frequency-domain split

A Blackwell-class GPU pulls ~1,400A at ~0.7V and swings that current in nanoseconds. V_droop = L·(di/dt), so even picohenries of parasitic inductance crash the rail unless charge is delivered faster than the wires "feel" the inductance. That requirement is layered by frequency, and the two companies own different layers:

• Murata MLCC — the nanosecond layer. ~nF–µF per part, kHz to hundreds of MHz, sitting millimeters from the die. Ceramic (BaTiO₃) dielectric. Low ESR is mandatory — it has to discharge at GHz. A material-science game (see moat below). The chat's load-bearing point: "you cannot just stack more MLCCs to fix it — the dielectric has to hold capacitance under DC bias. That is material science, not count."
• Musashi HSC — the multi-second layer. ~3,800 farads per cell (nine orders of magnitude more energy than an MLCC's microfarads), 3.8V/cell, sitting at the rack/system level. Activated-carbon + lithium-doped electrodes — electrochemical. Its job is peak shaving: buffering the second-to-minute power oscillations synchronized GPU clusters create (the OpenAI/Microsoft/Nvidia "Power Stabilization for AI Training Datacenters" paper is the demand driver). High ESR is its defining drawback — but for this job ESR doesn't matter, volumetric energy density does.

The boundary is physics, not marketing. A high-ESR HSC physically cannot dump charge at GHz (discharge current caps at V/R), so it's locked out of chip-level decoupling. No MLCC stores enough energy to ride through a multi-second grid sag. They are complements; both are levered to the same AI-power wave from opposite ends of "765kV → 0.65V."

Two different kinds of moat

This is the sharper investment contrast than the physics:

• Murata's moat is manufacturing — ~50 years of ceramic process IP, a 4-supplier oligopoly, 12–18mo AEC-Q200 qualification, ¥150–200B/2-yr per miniaturization generation. Durable, replication-proof from below.
• Musashi's moat is a product/datasheet monopoly in a niche so new no competitor has shown up — Irrational Analysis (May 28 2026): "an uber monopoly of a rapidly growing, mission-critical niche... massive technical/engineering moat," Musashi "vaporizes everyone else" on volumetric energy density. That is a more fragile moat — one breakthrough datasheet from a challenger — which is exactly why Made in Japan, despite liking the setup, hasn't bought (see below). Caveat carried from 7220: 3.8V/cell is NOT the differentiator (Vinatech 006440.KS is also 3.8V); the edge, if it holds, is 3.8V + 3,800F + the volumetric-density claim — and that claim is still internal-verification-needed.

Source notes

• Made in Japan — "A Hidden AI Data Center Opportunity (Trading at 6x EBITDA)," Apr 4 2026 (madeinjapan.substack.com). A business/valuation thesis on 7220: HSC peak-shaving as a hidden growth option inside an "unattractive" auto-parts shell; DC customers approached Musashi, not the reverse; capacity 1.5M→6.5M cells/yr; at full 6.5M, a back-of-envelope ~¥85bn run-rate revenue × 20% EBITDA × 15x = ~¥255bn for the HSC segment alone, ~eclipsing today's EV. Author explicitly has not pulled the trigger — too many known-unknowns (pricing not set, ramp delayed a year, depreciation schedule unsettled, possible advance capex). A more cautious read than the price action implies.
• Irrational Analysis — "Musashi Seimitsu ($7220.T) Supercap Note / UNLIMITED POWER," May 28 2026 (the proximate catalyst for 7220's +260%-off-low run; author discloses a personal position +57% after a prior -25% drawdown). The technical-moat thesis. Density > ESR; caps stack in series (TI/ADI app notes) to reach hundreds of volts; "these capacitors can easily kill a human."
• The claude.ai "Murata + PDN" thread (2026-05-31/06-01) is where this first-principles framing originated; an earlier turn mis-pulled Irrational Analysis when "Made in Japan" was meant — both are now captured here.

Murata material moat — and the "rotate down the stack" question

The damnang2 ("Murata: the stock you missed") comment asked: if Murata defends share, does more value sit upstream? Premise is currently inverted — Murata is raising prices (15–35% on AI/auto-grade, April 2026; GS Conviction List May 27, demand cycle extended to ~2030), not cutting. But the deeper logic holds: upstream materials capture the whole industry's volume growth regardless of which capacitor maker wins.

The investable upstream chain (corrected — see below):

Why Murata captures most of this internally: vertical integration — it makes ~50% of its own BaTiO₃ powder (hydrothermal synthesis, sub-100nm grain control no Chinese commodity player matches) plus in-house dielectric film and electrode paste. It is the upstream bottleneck, which is why the rotate-down-the-stack value largely accrues to 6981 itself rather than to a listed supplier.

The one real long-term threat to the material moat: embedded deep-trench capacitors (eDTC) / HD-MiM on-chip integration pulling decoupling onto the package/die — a 24–36mo watch item per Banyan Lane, not a near-term killer. Chinese commodity encroachment is bounded (stuck at the 0402/0603 tier; can't ship 008004). The material moat is Murata's most durable pillar — far more so than its eroding SAW/RF franchise.

Errata (why this note exists in this shape)

Shoei Chemical was initially tabled as a listed upstream play. It is private (Bloomberg 7867113Z:JP). Removed from the investable list; retained only as a structural chokepoint fact. See memory verify-listed-before-tabling. Distinct names: Shoei Chemical (private, Ni paste) ≠ Sakai Chemical 4078.T (BaTiO₃ powder) ≠ Shoei Yakuhin 3537.T (unrelated trading co).