As artificial intelligence workloads scale to rack-level systems and full data center dimensions, power delivery capability has emerged as a core bottleneck constraining data center system performance, density, and total cost of ownership. Within the ecosystem of NVIDIA MGX—an open, modular reference architecture—an efficiency revolution underpinned by All-Gallium Nitride (All-GaN) technology is quietly reshaping the power delivery path, extending all the way from high-voltage distribution down to the GPU core.

The latest development in this technological evolution comes from NVIDIA MGX ecosystem member Innoscience. The company is advancing comprehensive All-GaN power conversion technology to support next-generation high-density AI systems. For investors and data center operators, this upgrade in underlying power semiconductor technology is pivotal for breaking through the upper limits of rack power density and achieving a substantial reduction in the operational costs of high-compute facilities.

Traditional power delivery models are struggling to cope with continuously rising rack power demands. The challenge is no longer merely bringing electrical power into the rack, but efficiently and compactly converting high-voltage electricity to the operating voltage required by GPUs. Leveraging characteristics such as low on-resistance, low gate charge, and zero reverse recovery, GaN technology is becoming a key enabling technology to address this challenge, directly enabling smaller magnetic components, superior thermal performance, and lower total cost of ownership (TCO).

As AI systems advance toward higher-density power architectures, the market is closely watching this power delivery solution that breaks through physical space and thermodynamic constraints. This will not only shorten the engineering development cycles for accelerated computing systems but also significantly accelerate the large-scale commercial deployment of next-generation AI factories.

**Front-End Conversion Breakthrough: 12kW Solution Approaches 99% Peak Efficiency**

With the continuous rise in AI rack power, the front-end conversion stage has become one of the most demanding segments in the power architecture.

Within NVIDIA's 800 VDC power architecture, delivering DC power directly to locations closer to the rack reduces the number of conversion stages. However, this requires the front end to simultaneously handle high input voltage, high conversion ratios, and constrained thermal and board space budgets.

Innoscience's latest data demonstrates the direct benefits of GaN in this stage. In its 12 kW, 800V-to-48V stage design, the primary side utilizes 650 V GaN double-sided cooling (DSC) devices, and the secondary side employs 100 V GaN devices, achieving approximately 99% peak efficiency and 98.2% full-load efficiency at an operating frequency of 1 MHz. Furthermore, newly released 150 V GaN devices further simplify secondary-side design, reducing the required number of synchronous rectifier devices by 50%. The reduction in footprint enabled by this high-frequency operation delivers direct commercial value for AI systems pursuing higher rack density.

Beyond 48V front-end conversion, power architecture choices require high flexibility to meet diverse system design needs regarding board space and thermal budgets. Innoscience has extended its All-GaN solution portfolio to cover a full range of intermediate bus voltage options from 800V down to 48V, 12V, and 6V.

For 800V-to-12V conversion, the market can now leverage 40 V GaN devices to achieve efficient synchronous rectification and improved thermal performance. For 800V-to-6V conversion, 15 V GaN devices serve as a synchronous rectification solution, supporting lower intermediate bus architectures and thereby simplifying the final conversion to GPU core voltage. At the critical 48V-to-12V intermediate bus stage, Innoscience's 100 V GaN solution optimizes multi-phase buck conversion. Under the scale effects of AI factories, even marginal efficiency gains translate to significant reductions in cooling requirements and operational costs.

**Vertical Power Delivery Reshapes Core Response**

At the final conversion stage closest to the compute core, where current demand is extremely high and transient response is critical, traditional lateral power delivery faces severe challenges due to distribution losses and motherboard trace complexity. Vertical Power Delivery (VPD) is emerging as a viable architecture to provide shorter current paths, lower parasitic losses, and higher current density.

To meet the requirements of GPU fast dynamic transients, Innoscience has validated the feasibility of operating 15 V GaN HEMTs at frequencies between 3 MHz and 5 MHz. This capability can dramatically shrink the required size of magnetic components and capacitors. The company is currently developing DrGaN solutions that, by supporting high switching frequencies, significantly increase bandwidth, thereby reducing reliance on traditional large-output capacitors. As future MGX AI systems continue to increase accelerator current density, power stages supporting VPD will become essential foundational blocks for near-core GPU power delivery.

To accelerate customer adoption cycles, Innoscience offers a series of evaluation boards and reference designs to help system designers validate GaN performance across the entire AI power delivery tree. These platforms include a 12 kW 800V-to-48V demonstration board, a 48V-to-12V 4-phase GaN evaluation board, and a 6V DrGaN evaluation board targeting future vertical power delivery architectures.

The NVIDIA MGX ecosystem is driving the deployment of modular and scalable AI infrastructure. In an era where AI infrastructure is increasingly constrained by power, the evolution of power semiconductors must keep pace with rising compute density. Through comprehensive coverage from 800 VDC all the way down to GPU core voltage, higher-efficiency, higher-density AI power delivery infrastructure is accelerating from concept to reality.