Pricing Power Arriving Faster and Stronger Than Expected

研报英文原文证据摘录

Pricing Power Arriving Faster and Stronger Than Expected

IdeaM

Investment Thesis

AI accelerators are becoming increasingly transient-current

intensive

Peak current densities on modern AI accelerator boards have reached levels that place

significant demands on the power delivery network (PDN). The highly dynamic nature of

AI workloads generates rapid current excursions (high di/dt), resulting in voltage

transients that can persist from sub-microsecond to multi-microsecond timescales.

As accelerator power levels continue to rise while operating voltages remain extremely

low, maintaining voltage stability within increasingly tight tolerance bands has become a

critical design challenge. During these transient events, the power delivery network must

supply large amounts of charge before voltage regulators can fully respond.

Consequently, board designers are relying on greater amounts of local decoupling

capacitance to support transient load requirements and preserve power integrity.

This trend is driving increased adoption of higher-capacitance MLCCs, including 47µF,

100µF, and larger devices, in addition to traditional high-frequency decoupling capacitors.

As a result, total installed capacitance per accelerator and per rack is growing faster than

MLCC unit count, reflecting a shift toward higher-capacitance content rather than simply a

larger number of components.

This aligns with our supply chain checks on MLCC content growth for VR200 (vs GB300),

which we discussed previously.

Total MLCC unit demand is expected to increase approximately 80% in a VR200

NVL72 rack relative to a GB300 NVL72 rack: Based on our estimates, a complete Rubin

NVL72 rack contains approximately 570,000 MLCCs, compared with roughly 320,000

MLCCs in a GB300 NVL72 rack ( Exhibit 5 ).