AI data center power at the rack — socket economics and component moats
Mandatory in-rack energy storage substitutes for years-long grid interconnection capacity
GPU clusters swing load on millisecond scales, forcing interconnection and switchgear to be sized to peak while average load sits near half. Rack-level storage charges on dips and discharges on spikes so the grid sees average demand; GB300-class shelves cut peak grid demand up to roughly 30% and Rubin-class up to roughly 25% on peak current. Storage per GPU rises from near-zero two generations ago to hundreds of joules, occupies a large share of the power shelf, and is required to operate without reserving full peak capacity. That turns a new bill-of-materials layer into a partial substitute for a five-year interconnection queue, with early supply signals such as broad capacitor price increases when orders exceed capacity.
Rack component revenue compounds faster than hyperscaler capex via rising parts per GPU generation
Supplier revenue is the product of how many racks customers deploy and how many of a vendor’s parts are designed onto each rack. Hyperscaler capex has grown on the order of 50–80% year over year while leading power and analog suppliers have reported datacenter-related revenue growth roughly twice to three times that pace. Each GPU generation draws more current, adds power stages, and increases capacitor and power-semiconductor counts; technical difficulty rises in parallel, so the market sells harder parts in larger quantities rather than merely scaling commodity units.
MLCC demand in AI racks is volume-led with a tight supplier oligopoly but limited spot price leverage
High-grade multilayer ceramic capacitors beside the chip buffer instantaneous current; industry estimates put hundreds of thousands of MLCCs per NVL72-class rack, with further increases expected into the next GPU generation. Only a handful of firms manufacture these grades, with two suppliers estimated to hold the majority of AI-server share. Reported revenue growth for server capacitors can exceed 80% annually while average selling prices move only single digits, implying growth driven mainly by units shipped rather than scarcity pricing unlike contract-locked, non-fungible HBM. Volume-led upside depends on fab expansion timelines; without material, socket, or supply-structure protection, margins risk commoditization like DRAM.
| Instrument | Side | Target | Reason |
|---|---|---|---|
| 6981.T | Long | We believe the leading MLCC vendor combines proprietary dielectric materials, recurring design-win sockets on each GPU board generation, and a supplier set too small for quick entrant displacement—so rising parts-per-rack AI demand should flow disproportionately to volume share before pricing commoditizes. | |
| 009150.KS | Long | We believe the second-largest qualified MLCC supplier in AI server grades sits in the same oligopoly and parts-per-rack expansion dynamic, offering correlated exposure where socket wins and fab capacity—not spot price spikes—drive the economic upside. |
Durable rack profits require non-swappable parts — material, socket, or qualified-supplier scarcity
Design-in at board draw fixes suppliers for an entire GPU generation; growing BOM only enlarges the pie unless a part resists substitution. DRAM showed decades of rising density with cyclical price collapse because any matching spec is interchangeable. HBM diverged because custom logic and multi-year contracts removed fungibility. The same filter applies inside the rack—proprietary materials, repeated socket wins, or a tiny qualified vendor list are necessary for revenue growth to convert to defended margins; rotating sockets generation to generation or hype among named partners without scarcity alignment can leave valuations disconnected from fundamentals.
Themes
The content on this page is for informational purposes only and does not constitute financial advice. Stoquate is not a licensed financial advisor. Always conduct your own research and consult a qualified professional before making any investment decisions.