QPT has unveiled the qMicroModule IBC, an intermediate bus converter architecture intended to take 800V DC power used in emerging AI data centre racks and convert it directly to the 6V supply feeding processor voltage regulators.
The Cambridge power electronics company is positioning the design as a licensing platform for semiconductor manufacturers and power system suppliers rather than simply as a finished converter. QPT says a two-module stack can deliver roughly 5kW and is designed to provide up to five times the power density of current 800V converter approaches.
The development targets a part of the AI infrastructure stack that is becoming more constrained as rack power rises. NVIDIA is developing an 800V DC architecture for future AI factories, with the higher distribution voltage intended to reduce current, copper requirements, conversion stages, and the volume occupied by power equipment. Its roadmap includes rack and row architectures intended to support increasingly dense compute systems.
The 800V distribution still has to be stepped down to the low voltages used by accelerator and processor power stages, and the conversion hardware competes for the same physical space, cooling capacity, and electrical budget as the compute equipment it serves. QPT’s proposition is that switching frequency can be used to reduce the size of magnetics and associated components, provided the losses created by fast switching can be controlled.
Conventional resonant converters typically operate around the 1MHz level in this part of the power chain. QPT is instead applying a hard switching architecture built around several technologies it has developed for high frequency GaN power electronics. The company says the combination of its ZEST transformer, an energy harvesting network, and fast control allows the converter to keep scaling into the multi-MHz range without the steep loss increase that would otherwise make the approach impractical.
Two smaller converter modules are stacked so that the 800V input is divided between them, leaving each module handling around 400V. That allows the first design to use established 650V gallium nitride devices rather than waiting for higher voltage GaN products, while retaining a path to 1200V devices as those technologies mature. The outputs are combined to provide the low voltage supply required downstream.
QPT is also using cycle-level control to address load transients. AI processors can change power demand rapidly, forcing conventional designs to carry significant bulk capacitance to ride through those changes. QPT says its controller can respond within a single switching cycle, reducing the amount of stored energy required while also allowing faults to be shut down on the same timescale.
James Cannings, chief executive of QPT, said: “The whole industry agrees the AI rack is moving to 800 volts. The real question is how you convert that power without giving up the space the processors need.” He said the company intends to work with partners on designs based on the architecture.
The qMicroModule extends technology QPT is already using elsewhere. Its MicroDyno motor control platform operates at 1MHz using hard switching GaN, while its qAttach die attach technology is intended to improve heat transfer from power semiconductor dies. Earlier company work has positioned the same technology stack for power supplies as well as motor drives, giving the converter a technical lineage beyond a paper architecture.
The 800V transition is moving from concept towards an equipment ecosystem. NVIDIA says more than 80 companies are working around its 800V specifications, while its Kyber rack architecture is intended to use the higher voltage distribution model. At the same time, EMEA data centre capacity has moved beyond 12GW, with a much larger pipeline dependent on access to power and the equipment required to convert and distribute it efficiently.
QPT’s claimed advantage will ultimately depend on efficiency, thermal performance, electromagnetic compatibility, reliability, and manufacturability under sustained data centre loads. Those tests will determine whether multi-MHz switching can deliver the promised density once the architecture is integrated into production power systems rather than evaluated as a development platform.
The company is opening a limited number of lead design partnerships with semiconductor and power system companies, with licensing forming the commercial route into products designed for 2027 rack deployments and beyond.
Rob Gwynne, co-founder and chief technology officer at QPT, said: “Resonant converters run into a wall as you push them to higher frequencies and higher power. Ours keeps scaling: double the switching frequency and you roughly double the power density, with no redesign.” Partner designs will provide the next test of that scaling claim inside the emerging 800V data centre power chain.
