The Role of Capacitors in Powering AI at Scale

In 2024, we wrote about the growing power demands of artificial intelligence. At the time, 60 to 80 kW server racks were already stretching datacenter design assumptions. Now, AI server racks consumer 60 to 140 kW each, with NVIDIA’s GB200 NVL72 operating at roughly 120 kW per rack. Consumption could reach 600 kW by late 2027 with the Rubin Ultra NVL576 system. Beyond 100 kW, traditional server power assumptions begin to break down.

Power Conversion Chain

The transition from 54 VDC in-rack distribution to 800 VDC architectures permits dramatically higher energy density. NVIDIA’s published 800 VDC reference design shows 85 percent more power transmission through the same conductor size, with 45 percent less copper versus traditional 415 VAC distribution.  

Power supply units (PSUs) within AI servers convert facility AC or 800 VDC into low-voltage DC for graphics processing units (GPUs). Inside the power conversion chain, higher voltages reshape what capacitors are required to do. 

DC Link

DC link capacitors are traditionally responsible for storing bulk energy, smoothing rectified voltage, and providing hold-up time during power interruptions. That role hasn’t changed in modern rack servers, but the operating environment has. Film capacitors, like the CDE 944L, are now the dominant technology because they combine high ripple capability with low ESL and long life. In high-density PSUs, the DC link plays a more critical role in maintaining stability as GaN and SiC semiconductors introduce faster switching speeds and higher stress.

LLC Resonant Conversion

In LLC conversion, resonant capacitors are responsible for defining and stabilizing the frequency of the LC tank, enabling the soft-switching behavior that drives efficiency. While the job of resonant capacitors is the same in modern AI PSUs, but the performance window has narrowed. At 98 percent efficiency, even small shifts in capacitance can impact resonance and degrade performance. C0G high-voltage MLCCs are increasingly preferred to X7R dielectrics in modern designs because of their negligible capacitance drift, minimal temperature coefficient, lower loss at high frequencies, and smaller footprint.

EMI Filtering

EMI filter capacitors suppress noise, ensuring regulatory compliance at the AC input. In modern PSUs, faster switching speeds and higher system voltages increase noise and shrink compliance margins. In a more intense electrical environment, these capacitors are being asked to operate closer to their limit. X1/Y2 safety-certified MLCCs rated at 500 Vrms, among the highest available in ceramic technology, can suppress noise while limiting leakage current and maintaining fail-open behavior for safety certification.

Managing GPU Load Transients

Additional pressure points have emerged beyond the power conversion chain. AI workloads introduce synchronized demand that traditional severs haven’t seen. When thousands of GPUs initialize training batches simultaneously, rack-level current demand rises in milliseconds. Utility infrastructure could take 10 to 20 seconds to adapt. Supercapacitors are deployed at the rack level to buffer that mismatch by delivering energy faster and reducing peak demand upstream. At the output stage, aluminum electrolytic and aluminum polymer capacitors provide localized energy storage close to the load, smoothing bidirectional transients so GPUs receive a stable supply while upstream power infrastructure is shielded from abrupt load spikes.

For decades, Knowles has supported demanding power systems operating deep underground, inside the human body, and up in space. AI datacenters represent another extreme domain, defined by higher voltages, greater power density, and synchronized load behavior. The role of capacitors in power conversion is recognizable, but the scale, speed, and architectural level at which their jobs are performed has changed, and Knowles is helping engineers meet these new demands.