The Power Chain

Electricity does not reach a processor in the form the grid delivers it. It steps down through a chain of conversions, each built from power semiconductors and each losing a percentage. That chain is being rebuilt, because the voltage it was designed around has run out.

The proposal is to convert medium-voltage alternating current — 13.8 kV, or 13.8 to 35 kV depending on whose account you read — directly to 800 volts of direct current at the data-centre perimeter, removing intermediate stages, and to distribute at that voltage to racks of a megawatt and beyond. Google describes the same shift in its own terms: from the 48 VDC it championed a decade ago to plus or minus 400 VDC supporting up to 1 MW per rack, developed with Meta and Microsoft under a project called Mt Diablo. The detail worth keeping is why that nominal voltage: to draw on the supply chain electric vehicles already established. The power architecture can be rebuilt this fast because somebody else paid to industrialise the parts.

Silicon carbide and gallium nitride are doing the work, and the reason is switching speed rather than any headline efficiency claim. A device that switches faster needs smaller magnetics to move the same power, which is why a solid-state transformer can be a fraction of the size of the iron-and-copper unit it replaces. A silicon-carbide manufacturer puts the gap starkly: SiC devices operate above 10,000 hertz where conventional silicon IGBTs rated at 6500 volts are generally limited to a few hundred.

Every participant publishes an efficiency figure and none can be checked here — 98% per conversion stage, product ratings of 97 and 98%, power-factor correction peaking at 99.3%, an end-to-end projection of up to 5%. All are recorded as claims and asserted nowhere, for a specific reason: per-stage and end-to-end efficiency are different quantities, the number of stages differs between the architectures being compared, and no source read states its baseline. What can be said without one is structural. Removing a conversion stage removes its losses, and at a megawatt a rack, one percentage point is ten kilowatts.

The problem that actually decides whether 800 volts is deployable gets almost no coverage, and it is not efficiency. It is whether a technician can pull a board from a live rack. At 54 volts that is routine; at 800 VDC connecting or disconnecting on a live bus needs the board's capacitance charged and discharged under control, or the event is a fault. Infineon's stated contribution is a hot-swap controller on silicon-carbide JFET technology that lets boards be exchanged while the rest of the rack keeps running — and the economics are stated plainly alongside it: an AI server costs as much as thirty times a traditional server, so a rack powered down for service is a rack that stops earning.

The demand curve the components are being specified to is published by the people building them. Infineon puts AI rack power at around 120 kilowatts now, rising to 500, and to 1 megawatt by the end of the decade. Google projects more than 500 kW per rack before 2030. Navitas describes supporting 1 MW racks and beyond. Three companies with different interests landing on the same order of magnitude is meaningful corroboration — but they remain projections, and this entire architecture is being built in anticipation of a load rather than in response to a measured one.

What is missing is named and it is substantial. The multi-vendor specification behind all of this returned HTTP 403 to both an automated fetcher and a browser and was not circumvented, so an architecture designed by several companies through a standards body is described here through the announcements of individual participants — four of the five sources read being suppliers or a customer describing their own parts.