Overview

Infineon Technologies is expanding its collaboration with NVIDIA to attack one of the most demanding engineering problems in modern AI infrastructure: power delivery. Timed with NVIDIA's GTC Berlin, the two companies announced a joint focus on the high-voltage intermediate bus converter (IBC) stage of 800-volt direct current (VDC) architectures, with Infineon's gallium nitride (GaN) technology positioned at the center of the design. The announcement builds directly on a partnership first outlined in May 2025 and expanded last October, which laid the groundwork for serviceable, hot-swappable 800 VDC server racks.

The framing matters. This is not a conventional component supply agreement. It is a co-engineering effort aimed at rewriting the power conversion chain inside AI racks at precisely the moment that rack-level power consumption is climbing toward the megawatt range.

The 800 VDC Architectural Shift

Traditional data center power infrastructure was never designed for the load profiles that dense AI training and inference clusters now impose. As rack power draw escalates, conventional distribution architectures run into hard physical limits on current, copper mass, and conversion losses. The industry response has been a migration toward 800 VDC distribution, deployed alongside existing setups during the transition.

Raising the bus voltage reduces current for a given power level, which in turn reduces resistive distribution losses and the conductor volume required to carry them. But higher bus voltage creates a new problem: the silicon-level devices that AI processors depend on cannot accept 800 V directly. Something must step that voltage down efficiently, densely, and reliably. That task falls to the intermediate bus converter.

Technical Deep Dive: Why GaN Wins at the IBC Stage

The IBC sits between the 800 VDC main bus and the lower-voltage rails that downstream stages use to generate precise point-of-load voltages for AI processors. It is a multi-stage approach that slashes distribution losses while operating under strict constraints on board space, thermal headroom, and system uptime.

GaN is proving to be the optimal semiconductor technology for this stage. Compared with traditional silicon, GaN switches significantly faster, reaching frequencies up to 1 MHz at the IBC level while keeping losses exceptionally low. That switching frequency is the lever that changes board design economics.

  • Smaller magnetics and passives: Higher switching frequency shrinks the required size of magnetic components and passive elements, which dominate converter footprint at lower frequencies.
  • Higher power density: The direct result is a much lower board profile, freeing rack space that is exceptionally scarce in AI deployments.
  • Thermal relief: GaN's ultra-low conduction and switching losses deliver best-in-class conversion efficiency, keeping heat generation in check even as rack power densities soar.

Those three effects compound. Less board area consumed by power conversion means more area for compute. Less heat generated in conversion means less cooling capacity diverted from processors. In a megawatt-class rack, those margins are architectural, not incremental.

Portfolio Breadth and Grid-to-Core Integration

Infineon's GaN portfolio supports the full spectrum of high-voltage IBC conversion ratios, stepping 800 V down to 50 V, 12 V, or 6 V. That flexibility lets data center operators tailor rack architectures without sacrificing efficiency or density.

Critically, Infineon pairs its GaN power switches with its PSoC control C3 and EiceDRIVER families, backed by in-house manufacturing and system expertise. The broader portfolio also spans silicon and silicon carbide (SiC) options to cover a wide range of design requirements. The company is therefore not selling a single device but a matched control-and-driver stack around it.

Beyond the IBC, the collaboration covers the entire data center power chain, including advanced power path protection and hot-swap controller functionality. These are the safety and serviceability features that keep 800 VDC AI server boards running seamlessly during maintenance and fault events. The combination positions Infineon as a grid-to-core partner, delivering optimized power solutions from the facility's main supply straight to the AI processor.

Market Context

The timing reflects a broader squeeze in AI infrastructure economics. Power delivery has moved from a background concern to a first-order constraint on how densely operators can pack accelerators into a rack. Every percentage point of conversion efficiency recovered, and every square millimeter of board reclaimed, translates into deployable compute. Vendors that can supply the full chain, from facility-level distribution down to the processor rail, gain disproportionate influence over rack reference designs.

Industry Impact

By anchoring GaN at the IBC stage of an 800 VDC architecture co-developed with NVIDIA, Infineon is effectively arguing that wide-bandgap semiconductors belong at the core of AI power delivery rather than at its periphery. That framing pressures competing power semiconductor suppliers to demonstrate comparable high-voltage GaN capability, matched control silicon, and system-level integration. It also raises the technical bar for rack integrators, who must now treat power architecture as a design variable co-optimized with compute silicon rather than a downstream procurement decision.

Implications for AI Data Centers

For operators, the practical implications are concentrated in three areas: density, serviceability, and thermal budget. A simpler, higher-density 800 VDC design replaces today's multi-stage, inefficient power conversion patchwork. Hot-swappable 800 VDC racks, the foundation laid by the earlier phases of the partnership, preserve uptime during component replacement. And lower conversion losses reduce the cooling burden in facilities already constrained by thermal headroom.

Future Outlook

Infineon's CEO Jochen Hanebeck framed the effort in structural terms: "AI is fundamentally reshaping the demands we place on power infrastructure, and semiconductors are at the heart of solving that challenge. Our collaboration with NVIDIA on 800 Volt power architectures shows what is possible when deep semiconductor expertise meets the pace of innovation required by AI. Gallium nitride-based system solutions are a key enabler of this next generation of data centers, and Infineon is proud to help power the AI era from grid to core."

The open question is how quickly 800 VDC rack designs move from co-development into volume deployment across NVIDIA's upcoming AI platforms. If they do, GaN's position at the IBC stage will shift from promising to default, and the grid-to-core integration model will become the reference architecture the rest of the industry is measured against.