Overview

The escalating power demands of AI-driven data centers are necessitating a fundamental rethinking of power delivery architectures. A recent technical analysis highlights the integration of high-current Silicon Carbide (SiC) power modules into multi-megawatt Solid-State Transformer (SST) architectures. This advanced design enables direct conversion from 34.5 kV medium-voltage grids to 800 V and future 1500 V DC infrastructures, addressing the critical need for high efficiency and power density in modern hyperscaler facilities.

Market Context

Electrical power and grid connectivity have emerged as the ultimate constraints for AI infrastructure. The electricity demand for data centers is projected to increase at least twofold by 2030, with large-scale hyperscalers acting as the primary driver. Currently, 72 percent of data centers operate with a power demand exceeding 50 MW, and blueprints for 100 MW installations are already available. Looking ahead, trends indicate that future hyperscalers will require power in the gigawatt range.

Because AI infrastructure is immensely power-hungry, any improvement in efficiency directly translates into computational throughput. Consequently, performance per Watt, specifically measured as Tokens per Second per Megawatt, has become a key index for evaluating AI throughput and generating revenue.

Technical Details: The Shift to SST

The power architecture supplying server racks is transitioning from a conventional 48 V DC-bus to 800 V. Traditional systems rely on Low-Frequency Transformers (LFT) and multi-stage AC-to-DC conversion at low voltage. In contrast, an SST-based approach employs an AC-to-DC conversion stage with galvanic isolation to energize the 800 V or 1500 V DC-bus directly from the 34.5 kV distribution network.

Replacing LFTs with SSTs offers superior efficiency, increased power density, and enhanced modularity. Modular SST architectures are highly attractive for hyperscalers due to their scalability. The power supply architecture utilizes multiple AC/DC and isolated DC/DC conversion cells connected in series at their input. The evaluation focuses on the isolated DC-to-DC conversion stage using a dual-active bridge topology. With an 1800 V DC-link voltage on the input side, grid voltages like 34.5 kV can be achieved with 16 cells per phase, limiting overall system complexity.

800 V DC-Voltage Output System

For the isolated DC/DC converter stepping down from 1800 V to 800 V, the architecture utilizes specific Mitsubishi Electric SiC MOSFET power modules. The primary side employs the FMF800DC-66BEW (3.3kV/800A) in an LV100 housing, while the secondary side uses the FMF600DXE-24BN (1.2kV/600A) in an NX housing with two devices in parallel.

Operating conditions for this 800 V system include a switching frequency of 10 kHz, an output power of 550 kW, a turns ratio of 2.25, and a stray inductance of 25 µH. Based on liquid cooling with thermal resistances from the heatsink to water at approximately 16K/kW and 30 K/kW for the LV100 and NX packages respectively, simulation results demonstrate that the maximum virtual junction temperatures reach 127 °C for the primary side and 150 °C for the secondary side.

1500 V DC-Voltage Output System

Driven by the need for higher power densities, 1500 V DC-bus systems are being evaluated. This voltage level aligns with the upper limit of the European Low Voltage Directive and is widely accepted in PV inverters and battery-energy-storage systems.

For the 1800 V to 1500 V conversion, the primary side utilizes two FMF800DC-66BEW modules in parallel. The secondary side employs two FMF1600DC-50CW (2.5 kV/1600 A) modules in parallel. Both devices utilize the LV100 housing. The virtual junction temperature is estimated using liquid cooling, maintaining a thermal resistance from the heatsink to the water of around 16 K/kW for the LV100 devices.

Implications and Future Outlook

The integration of high-power SSTs and high-current, highly reliable SiC modules provides a robust solution for multi-megawatt hyperscaler applications. By innovatively combining high power densities with high operational efficiency, these architectures maintain the modularity required to scale output power seamlessly.

  • Direct medium-voltage to DC conversion reduces conversion stages and associated losses.
  • Modular cell-based designs allow for gigawatt-scale data center power delivery.
  • Advanced SiC modules ensure reliable operation under extreme thermal and electrical stresses.

Ultimately, optimizing the power supply architecture is no longer just a hardware challenge; it is a direct enabler of AI performance. By maximizing the Tokens per Second per Megawatt metric, these next-generation SST architectures will define the economic and operational viability of future AI infrastructure.