The Power Conversion Layer: How GaN, B-TRAN and High-Density Power Modules Are Reshaping AI Data Center Infrastructure

GridReadiness documents the power infrastructure bottlenecks that block AI data center deployment — transformers, grid connections, GOES supply chains. But inside the data center, a parallel transformation is occurring in the power conversion layer: the systems that take grid electricity and deliver it precisely to GPU racks. Understanding this layer matters for anyone evaluating the complete AI infrastructure stack.

THE POWER PATH FROM GRID TO GPU

Electricity arriving at an AI data center from the grid follows a specific path before it reaches a GPU. Each conversion step introduces losses — heat generated, efficiency reduced. The companies working on this layer are attacking these losses at multiple points:

For a 100 MW data center, a 20% aggregate conversion loss means 20 MW consumed in power conversion infrastructure rather than computation. At €60/MWh, that is over €10 million per year in electricity that generates zero compute revenue. The companies improving this efficiency are attacking a large and real cost.

GAN — THE SILICON REPLACEMENT FOR POWER CONVERSION

Gallium Nitride (GaN) semiconductors are replacing traditional silicon MOSFETs and IGBTs in power conversion applications. The physics of GaN allow faster switching at higher voltages with lower energy losses — translating directly into smaller, more efficient power conversion systems.

Navitas Semiconductor is the leading pure-play GaN company for data center applications. Their GaNFast technology enables power conversion circuits that are smaller, run cooler and waste less energy than silicon-based alternatives. For data centers where cooling is a major operational cost, reducing the heat generated by power conversion has direct economic value.

VICOR — HIGH-DENSITY POWER FOR AI SERVERS

Vicor Corporation operates at a different point in the power conversion chain — between the data center power distribution and the individual server. Their factorised power architecture delivers high-density power modules that convert electrical power as close to the GPU as possible, minimising distribution losses.

The relevance for AI infrastructure is direct: NVIDIA's high-end GPU systems consume power at densities that exceed what legacy server power architectures were designed for. A single H100 GPU system can draw 700W or more. Rack densities in AI data centers are reaching 30–100 kW per rack, compared to 3–10 kW in traditional enterprise data centres. Vicor's technology is designed for exactly this environment.

IDEAL POWER AND B-TRAN — THE BIDIRECTIONAL SWITCH

The most architecturally interesting development in data center power conversion is the bidirectional switch. Traditional power semiconductors (IGBTs, MOSFETs) are fundamentally unidirectional — they allow current to flow in one direction. Creating bidirectional power flow (essential for battery systems, UPS units and energy storage) currently requires multiple components in complex arrangements.

Ideal Power's Bidirectional Bipolar Junction Transistor (B-TRAN) addresses this directly. A single B-TRAN component handles bidirectional current flow natively, with dramatically lower conduction losses than multi-component alternatives.

For AI data centers, the UPS (Uninterruptible Power Supply) application is particularly relevant. A 100 MW data center requires 100+ MW of UPS capacity. Current UPS technology using conventional bidirectional switching loses significant energy in both charge and discharge cycles. B-TRAN in UPS systems could reduce these losses substantially — but the technology is still in the validation phase with industrial customers.

ROGERS CORPORATION — THE THERMAL MANAGEMENT LAYER

Between power conversion and computing, heat management is the constraint that limits how densely components can be packed. Rogers Corporation produces advanced substrate materials — laminates, ceramics and thermal interface materials — that manage heat in high-power applications.

Their relevance to AI infrastructure is indirect but real: as GPU power density increases and power conversion components run at higher frequencies, the thermal requirements at the component level become more demanding. Rogers materials are embedded in the substrates of power semiconductors, the circuit boards of server power supplies, and the thermal management systems of high-performance electronics.

THE GRIDREADINESS CONTEXT

These power conversion developments operate at a different level from the grid infrastructure that GridReadiness primarily covers. But they connect to the same fundamental question: how efficiently can a data center convert grid electricity into useful AI computation?

The six-layer Physical AI infrastructure model that GridReadiness has documented places power infrastructure (Layer 6) as the foundation that all other layers depend on. The power conversion layer sits within Layer 6 — between the transformer (external grid interface) and the GPU rack (compute output). Improvements in GaN, B-TRAN and high-density power modules make the internal half of Layer 6 more efficient, while GridReadiness focuses on the external half: the grid connections, transformers and sites that bring power to the data center in the first place.

For infrastructure investors evaluating the full AI data center power chain, both halves matter. The external bottleneck (transformers, grid connections) is where the deployment delays are measured in years. The internal efficiency gains (GaN, B-TRAN) are where the operational economics are won over the life of the asset.

INVESTMENT FRAMEWORK SUMMARY

The power conversion layer is not the bottleneck that is blocking AI data center deployment in 2026 — that bottleneck remains the external transformer and grid connection supply chain. But it is the layer where the efficiency wars of the next decade will be won, and understanding it completes the picture of how power flows from the European nuclear grid to a GPU running AI inference.