Navitas, NVIDIA, and 800 VDC: What the Public Evidence Actually Shows

/ Insights / By

The headline version of the story sounds simple: Navitas, NVIDIA, 800 VDC, GaN, AI servers.

The public evidence is more interesting, and more limited, than that. NVIDIA publicly names Navitas as part of its 800 VDC AI data center power ecosystem. Navitas says its GaN and SiC technologies were selected to support NVIDIA’s next-generation 800 V architecture. Navitas has also demonstrated relevant 800 V-to-low-voltage GaN power hardware at NVIDIA events.

What the public record does not yet prove is the stronger claim that Navitas already supplies a disclosed production “800 V GaN package” inside an NVIDIA AI server. For power semiconductor strategy, that distinction matters.

The Common Misunderstanding

The common misunderstanding is to treat every NVIDIA ecosystem announcement as a confirmed production socket.

That is not how high-power AI infrastructure usually moves. First comes the architecture. Then the ecosystem. Then reference designs and demonstrations. Then qualification, second-source decisions, production system disclosure, and volume procurement. Public materials can confirm the early and middle parts of that path without proving the last one.

In this case, the safe reading is:

  • NVIDIA is publicly promoting 800 VDC as a next-generation AI factory power architecture.
  • NVIDIA publicly lists Navitas among its 800 VDC ecosystem partners.
  • Navitas publicly says its GaN and SiC technologies support NVIDIA’s 800 V architecture.
  • Navitas has shown 800 V-to-6 V GaN-based power delivery hardware at NVIDIA-related events.
  • The exact production socket, timing, volume, and bill-of-material position are not publicly proven from the sources reviewed.

That is still important. It is just not the same as saying “Navitas won the NVIDIA production GaN socket.”

Why NVIDIA Is Talking About 800 VDC

The power problem in AI data centers is no longer just an efficiency problem. It is a current, copper, connector, thermal, and serviceability problem.

Low-voltage rack distribution was reasonable when rack power was measured in tens of kilowatts. It becomes painful when racks move toward hundreds of kilowatts and, eventually, megawatt-class AI factory power levels. At fixed power, higher voltage means lower current. Lower current means less copper, lower distribution loss, smaller bus structures, and easier thermal management.

The simple math explains the architecture pressure:

  • 120 kW at 54 V is about 2,222 A.
  • 120 kW at 400 V is about 300 A.
  • 120 kW at 800 V is about 150 A.

Because resistive loss scales with current squared, the move from 54 V to 800 V is not a small packaging tweak. It changes the entire power distribution problem. NVIDIA’s own 800 VDC materials frame the shift around copper reduction, efficiency, power density, and the ability to scale AI factories beyond what legacy 54 V rack power can comfortably support.

Where GaN Fits

Moving power around the rack at 800 V does not remove the need to feed GPUs and accelerator boards at much lower voltage. It simply moves the hard conversion problem closer to the compute load.

That is where GaN becomes strategically interesting. High-frequency GaN switching can reduce magnetics and passive size, support dense resonant conversion, and help make a late-stage converter small enough to sit near the GPU board or compute tray. In plain power-system terms:

  • 800 VDC helps solve the rack current problem.
  • GaN helps solve the converter density problem that appears after high-voltage distribution reaches the compute node.
  • SiC is more naturally suited to higher-voltage infrastructure stages such as facility, rectifier, SST, or rack-front-end conversion.

This is why the Navitas story is not only about GaN. Navitas positions both GaN and SiC into the AI factory power chain: GaN for dense high-frequency conversion, SiC for higher-voltage and higher-power infrastructure stages.

What Navitas Has Actually Shown

Navitas publicly announced an 800 V-to-6 V power delivery board for AI data center applications, using 650 V GaNFast devices on the primary side. The company has described the board as targeting high switching frequency, high peak efficiency, and high power density, and has shown it in NVIDIA event contexts including GTC and the AI Factory MGX ecosystem showcase.

This is strong evidence that Navitas is technically relevant to the 800 VDC architecture discussion. It is also strong evidence that Navitas wants to be seen as an NVIDIA ecosystem power partner.

But a demonstration board is not automatically a production design win. NVIDIA’s own architecture language describes ecosystem development and future Kyber-era deployment. Navitas’ public demonstrations show credible hardware, not a publicly disclosed server bill of materials.

The Boundary That Matters: Rack, Board, or Package?

One reason wording gets slippery is that “AI server power” covers several layers:

  • Facility and power room: grid AC conversion and centralized high-voltage DC distribution.
  • Row, rack, or sidecar: HVDC distribution, protection, busway, and conversion blocks.
  • Compute tray or node entry: where high-voltage power reaches the compute hardware.
  • Board-level conversion: high-ratio conversion down to 12 V, 6 V, or another intermediate rail.
  • Point of load: multiphase VRM and final rails for GPU, memory, and local logic.
  • Package-level power: power delivery inside or immediately around the accelerator package.

The public sources reviewed support Navitas’ relevance across infrastructure and board-adjacent conversion. They do not show Navitas devices inside an NVIDIA silicon package. Unless new evidence appears, avoid “package-level” language.

Vendor Claim Versus Field Reality

  • “Navitas provides 800 V GaN for NVIDIA AI server packages.” Too strong. Public evidence supports ecosystem participation, collaboration, and demonstrated hardware. It does not prove package-level production insertion.
  • “Navitas is an NVIDIA 800 VDC ecosystem partner.” Safe. NVIDIA publicly names Navitas in its 800 VDC ecosystem materials.
  • “Navitas’ GaN and SiC technologies were selected to support NVIDIA’s 800 V architecture.” Safe if attributed to Navitas. It is Navitas’ public statement, also reflected through investor materials.
  • “Navitas won the Kyber power socket.” Avoid. The production socket and bill of materials are not publicly disclosed.
  • “800 VDC means GaN replaces all AI server power stages.” Avoid. GaN, SiC, silicon multiphase controllers, DrMOS, eFuses, isolation, protection, and telemetry all remain part of the stack.

What ChinaSemiOps Would Verify Before Treating This as a Design Win

For an analog or power IC company looking at this ecosystem, the important question is not whether the story is exciting. It is where the socket actually sits, whether the requirement is accessible, and how long qualification will take.

Before calling this a production design win, ChinaSemiOps would want evidence for:

  • the exact power-conversion boundary: facility, rack, tray, board, or package;
  • whether the demonstrated 800 V-to-6 V board is a production candidate or a technology demonstrator;
  • whether NVIDIA or an ODM has confirmed the design in a shipping system;
  • which voltage rail the final Kyber-era implementation actually uses near the GPU;
  • whether second sources from EPC, Infineon, TI, Renesas, onsemi, MPS, or others are being qualified;
  • what qualification, reliability, safety, and serviceability requirements apply at 800 VDC inside AI racks.

The Practical Takeaway

The clean takeaway is this:

800 VDC matters because low-voltage rack distribution becomes a current and copper problem at AI factory scale. GaN matters because high-voltage distribution still needs dense, efficient conversion near the compute load. Navitas is publicly in NVIDIA’s 800 VDC ecosystem and has shown relevant GaN hardware. Publicly, it is not yet proven that Navitas already supplies a production NVIDIA AI server package.

For suppliers, the opportunity is real but should be mapped by power-chain layer. The facility and rack layers may favor SiC, protection, isolation, and high-voltage conversion. The tray and board-adjacent layers are where GaN becomes especially attractive. The final point-of-load layer remains a brutally competitive multiphase power IC market.

That is the difference between a useful technology roadmap and a headline-driven overread.

Sources