#AI #DataCenters #OpenCompute #HVDC #PowerElectronics #SiC #GaN #Diablo400 #OCP #AIInfrastructure
AI isn’t just changing compute, it is rewriting power electronics. With GPU-rich AI racks already consuming 800 kW to 1.1 MW today and 1.5 MW+ on the horizon, traditional silicon-based power architectures can no longer keep up.
Enter wide-bandgap (WBG) semiconductors, Silicon Carbide (SiC) and Gallium Nitride (GaN) and the OCP Diablo 400 ±400 V HVDC architecture, co-authored by Microsoft, Meta, and Google. Together, they’re redefining how we deliver megawatts of power per rack safely, efficiently, and sustainably.
Why Silicon Hits Its Limits
Legacy silicon MOSFETs and IGBTs have been the workhorses of data center power electronics for decades. But as AI accelerators push power envelopes higher, their limitations are exposed:
- Switching Speed Bottlenecks → Silicon struggles to achieve the >1 MHz switching speeds needed for high-density converters.
- High Conduction & Switching Losses → At ±400 V and future 800 V HVDC fabrics, silicon wastes too much energy as heat.
- Thermal Constraints → Lower efficiency drives larger heatsinks and forces higher cooling costs.
For racks drawing 1 MW+, every fraction of efficiency matters and silicon just can’t scale.
SiC at the Front End: Efficient AC/DC Conversion
In the Diablo 400 architecture, SiC-based solid-state transformers (SSTs) sit at the front end of the power chain, converting 3ϕ AC to ±400 V HVDC with unprecedented efficiency.
Why SiC Excels
- High Breakdown Voltage → SiC MOSFETs easily handle 1.2 kV+ and prepare power fabrics for future 800 V HVDC standards.
- Lower Switching Losses → Higher efficiency (>98% at 50% load) reduces cooling overhead.
- High Thermal Conductivity → Handles multi-megawatt conversion without massive heatsinks.
These SSTs form the backbone of the ±400 V HVDC fabric defined by OCP Diablo 400.
GaN at the Edge: Sub-µs Point-of-Load Conversion
While SiC dominates high-voltage conversion, GaN takes over at the point of load (POL) delivering ultra-fast, low-voltage power directly to GPUs, TPUs, and ASICs.
Why GaN Wins at POL
- >1 MHz Switching Frequencies → Enables smaller magnetics and higher power density.
- Ultra-Fast Current Tracking → Matches GPU load transients in nanoseconds, essential for AI inference and training.
- Lower VRM Stress → Reduces thermal cycling on voltage regulator modules (VRMs), improving long-term reliability.
In AI data centers, where GPUs can demand 175% TDP spikes at 10 kHz+, GaN’s speed and efficiency are game-changers.
Diablo 400: ±400 V HVDC Fabrics for Today
The Open Compute Project’s Diablo 400 specification defines a disaggregated power rack architecture purpose-built for AI-scale compute:
- ±400 V HVDC Distribution → Reduces I²R losses and supports 1.1 MW racks efficiently.
- Hot-Swappable 30 kW PSUs → Enables modular scaling and rapid servicing without downtime.
- Telemetry Integration → Real-time monitoring via Redfish, PMBus, and CANbus ensures predictive maintenance.
- Arc-Safe Connectors & SSCBs → Guarantees operator safety during maintenance and hot-swaps.
This architecture solves today’s scaling challenge but also lays the foundation for future 800 V fabrics.
Preparing for 800 V AI Factories
Hyperscalers like NVIDIA, Meta, and Google, alongside infrastructure providers like Vertiv, are already prototyping 800 V HVDC fabrics to power 1.5 MW+ AI racks.
The shift to 800 V demands:
- SiC SSTs capable of handling >1.5 kV breakdown voltages.
- Liquid-cooled busbars for 2 kA+ continuous current delivery.
- GaN-based DC/DC converters switching even faster for ultra-dense GPU trays.
- Solid-State Circuit Breakers (SSCBs) for microsecond-level fault isolation.
Diablo 400 establishes the standards, connectors, telemetry, and safety protocols needed to make this transition seamless.
Key Takeaways
- Silicon is no longer enough → it can’t meet the efficiency, speed, and thermal demands of AI-scale compute.
- SiC enables high-voltage AC/DC conversion, forming the foundation of ±400 V and 800 V HVDC fabrics.
- GaN drives sub-microsecond point-of-load conversion, critical for GPU transient response.
- Diablo 400 ±400 V HVDC architecture solves today’s megawatt-rack challenge and future-proofs AI data centers.
- The AI era demands power electronics that think as fast as the workloads they feed.
AI workloads are redefining compute, cooling, and power delivery. With SiC, GaN, and Diablo 400 ±400 V HVDC fabrics, we’re entering an era where power architectures evolve at the speed of AI innovation.