Navitas debuts 800v GaN board for AI racks

What made October 2026 matter was that the industry stopped talking about higher-voltage AI power in the abstract and started assembling an interoperable stack around it. BigGo Finance captured that shift in the headline, “Navitas Unveils 24 kW 800V AI Power Board as HARTING Joins Nvidia’s MGX Ecosystem,” linking a specific 800V DC product to Nvidia’s modular platform and adding ecosystem-level connector support, while Navitas’s role as a wide-bandgap semiconductor company made clear that this was not just a new board, but a semiconductor-enabled move toward an 800V DC rack power-delivery standard.

The mechanism works because the new architecture and the new semiconductors reinforce each other: MarketBeat reported that Navitas’s MGX-aligned board “executes a single-stage conversion from 800 volts DC directly to 12 volts DC,” whereas “Traditional data center architectures rely on an intermediate 50-volt conversion step,” and “Removing that intermediate step delivers 12 volts directly to the voltage regulator module.” That simplification is only practical because wide-bandgap devices are now ready for the job; as MarketBeat framed it, gallium nitride technology and the broader shift to GaN and SiC are what make streamlined, lower-loss, high-density 800V AI rack power delivery feasible at deployment scale.

The architectural change is straightforward: instead of stepping power through several intermediate conversions before it reaches compute, the new design pushes 800V DC deeper into the rack and collapses stages along the way. As the Profile put it, power traditionally goes “Grid to die in four conversions, 13.8kV down to under 1V, with 800VDC designs merging the UPS and PSU stages into one,” which is why the emerging approach is better understood as direct 800V DC-to-12V DC delivery followed by tightly managed final point-of-load conversion.

What makes that simplification practical is the switching behavior of wide-bandgap devices, especially GaN and SiC, which cut the heat penalty that accumulates every time electricity is converted. The Explainer framed the stakes starkly: “You take a gigawatt of power and you convert it into something like 700 megawatts of heat… There’s maybe 8% lost in power conversion on the way to the chip, maybe 10% depending on the architecture,” so faster, lower-loss switching matters; that is why “Navitas Semiconductor designs GaN and SiC chips… offer a lot of exposure to the 800VDC buildout,” while “TXN makes the vital electrical plumbing that regulates, converts, monitors, and distributes electricity across hardware architecture like 800VDC.”

The scale change is no longer theoretical: BigGo Finance reports that “a single top-spec AI rack now draws roughly 120 kilowatts,” far beyond the legacy 5-to-20-kilowatt range where conventional rack design was comfortable. Using Nvidia’s GB200 NVL72 as a concrete marker, BigGo Finance describes a system where “one rack crams in 18 servers” and weighs roughly 1.3 metric tons, underscoring that AI infrastructure has already crossed into a different class of power density, thermal load, and physical integration than traditional enterprise racks were built around.

Vendors are responding as if this is an immediate market transition, not a distant roadmap: BigGo Finance says Lenovo’s next-generation liquid cooling can move 98% of heat into water lines, enabling a “100-kilowatt-plus rack” to run without computer-room air conditioning. On the power side, Quiver Quantitative News captured the strategic response in Navitas’s move to buy deeper into the stack, with the headline “Navitas Acquires Claros for $232.8M to Smash AI Data Center ‘Power Wall,’ Eyes $8B Market by 2030,” a clear sign that suppliers see higher-density rack power as large enough to justify vertical acquisitions.

The biggest brake on adoption is that Navitas’s 800V push is tied to NVIDIA’s schedule, so any slippage upstream can delay the whole ecosystem’s commercial payoff. SixSigmaCapital noted that a 6 July 2026 SemiAnalysis report said NVIDIA’s Kyber rack architecture, critical for the 800VDC content ramp tied to Rubin Ultra, was “delayed more than a year, to 2028,” even as NVIDIA publicly disputed the report and said its roadmap remained intact, leaving uncertainty around when Navitas’s design wins can convert into volume deployments.

Even if the roadmap holds, Navitas still has company-specific execution risks that could slow customer adoption and scaling. SixSigmaCapital said Wolfspeed sued in July 2026 alleging that products including “GaNFast, GaNSlim and GaNSafe” and “GeneSiC MOSFET and SiCPAK models” infringe five patents, while Navitas also showed “reliance on a single distributor peaked at 71% of net revenue in Q2 2026,” and reported Q2 2026 revenue of $10.53 million, down 27.3% year over year from $14.49 million, underscoring commercialization fragility.