Subscribe to Our Digest
Artificial intelligence is moving from chatbot conversations to autonomous AI agents working together on huge data sets – and the data center is becoming an AI factory. These next generation facilities require unprecedented computer power which in turn brings new challenges of power delivery, thermal management and energy efficiency.
To help address these challenges, NVIDIA has created the NVIDIA MGX™ architecture, a modular platform that accelerates the deployment of AI infrastructure. But the real potential of MGX can only be realized with equally advanced power conversion technology.
NVIDIA MGX is a modular server architecture that allows OEMs and cloud providers to build a wide range of AI systems using standardized components. Instead of redesigning hardware for every new accelerator generation, manufacturers can reuse common building blocks across servers, racks, and large-scale AI deployments.
The architecture supports over 100 system configurations, reducing engineering costs and accelerating time-to-market. MGX enables seamless integration of CPUs, GPUs, DPUs, networking, and storage technologies while providing the flexibility required for rapidly evolving AI workloads.
As AI models continue to scale, however, the primary challenge is no longer compute density alone – it is delivering power efficiently to increasingly power-hungry accelerators.
The new MGX architecture has been designed to meet the needs of emerging agentic AI workloads. These systems will require high-throughput inference, large memory capacity, low-latency interconnects, and continual reasoning.
To address these needs, NVIDIA has developed a number of rack level innovations including PCB based modular rack design, dynamic power steering, liquid cooling and advanced power management technologies.
One of the most important trends is the shift to higher voltage power distribution architectures that are designed to accommodate future growth in rack power density.
Most traditional AI servers use 48 VDC distribution within the rack. Future AI accelerators are expected to require a few kilowatts per device, though suitable for current systems.
Power increases, so low voltage distribution needs big currents. More copper loses , bigger conductors , connector limitations and more thermal challenges. Higher current means more copper losses.
For example, delivering 12 kW on a 48 V distribution network would require ~250 A of current, which poses considerable efficiency and implementation challenges.
These limitations are spurring interest in higher voltage DC distribution architectures.
Traditional data center power architectures are reaching their limits. Future AI processors are expected to consume several kilowatts each, making conventional 48 VDC distribution increasingly inefficient due to excessive current and copper losses.
NVIDIA’s answer is an 800 VDC power architecture that distributes high-voltage DC power directly to AI racks. This approach offers several advantages:
The architecture is also designed to support both new AI factories and existing data centers through hybrid deployment models.
Converting 800 VDC into the low voltages required by GPUs requires extremely efficient power electronics. Silicon-based solutions struggle to meet the efficiency, thermal, and power-density requirements of modern AI infrastructure.
Gallium nitride (GaN) transistors provide a superior alternative through:
These characteristics allow designers to build compact, high-efficiency power converters capable of supporting megawatt-class AI racks.
This approach preserves compatibility with existing server ecosystems while benefiting from higher-voltage distribution throughout the facility.
It provides a practical migration path for operators seeking to increase rack power without redesigning every subsystem.
EPC’s 150 V GaN devices support compact isolated converters that achieve higher efficiency and power density than traditional silicon MOSFET solutions. This makes the architecture attractive for organizations seeking a gradual transition to next-generation AI infrastructure.
Direct conversion from 800 VDC to 12 VDC reduces distribution losses and eliminates intermediate power stages.
This architecture is particularly valuable within NVIDIA’s high-density GPU environments, where minimizing electrical losses and maximizing thermal performance are essential.
EPC recommends an eight-stage Input-Series Output-Parallel (ISOP) converter architecture using both 150 V and 40 V GaN devices to optimize efficiency, thermal distribution, and system reliability.
For maximum performance, some AI systems may convert directly from 800 VDC to 6 VDC before supplying sub-1 V power rails to GPUs.
This approach minimizes conversion complexity while leveraging EPC’s latest low-voltage GaN transistors. Devices such as the EPC2378 and EPC2379 deliver ultra-low resistance and high-current capability, enabling efficient point-of-load (POL) power delivery for AI accelerators.
One of the leading approaches for 800 VDC conversion is the Input-Series Output-Parallel (ISOP) architecture.
In an ISOP design, multiple converter stages share the high input voltage while delivering power in parallel at the output. This approach distributes thermal stress, reduces component voltage requirements, and enables the use of highly optimized lower-voltage switching devices.
Additional advantages include:
These characteristics align well with the space and cooling constraints of modern liquid-cooled AI servers.
The EPC91123 reference design demonstrates the potential of this approach, delivering 6 kW power conversion from 800 VDC to 12.5 VDC while achieving peak efficiency above 98%.
The most promising application of GaN is thought to be in the last power conversion stage feeding GPUs.
EPC’s newest 25 V and 18 V GaN transistors deliver very high current densities while retaining efficiency at switching frequencies over 2 MHz. This allows for smaller voltage regulators, faster transient response, smaller PCB area and less cooling.
These benefits directly translate to lower operating costs and improved compute density for hyperscale AI deployments.
As the scale of AI infrastructure increases and approaches more power dense architectures, power delivery is becoming a primary system level design consideration. The move from conventional 48 V distribution to 800 VDC architectures is driven by the need of the industry for higher efficiency, lower losses and to support future generations of accelerators.
In this context, advanced power conversion technologies such as GaN are emerging as key enabling technologies. Together with modular platforms such as the NVIDIA MGX, they offer a pathway to more efficient and scalable AI data centers that can support the next generation of training and inference workloads.
More information about using GaN-on-Silicon Power FETs