It's a school day!
Nvidia is moving its future AI data centres towards an 800V power architecture, and the reason is pretty simple: the amount of power these systems need is getting ridiculous.
Today's AI racks typically use much lower-voltage power distribution(sub 60V), but as Nvidia moves towards racks needing hundreds of kilowatts, and eventually around 1MW(that's super massive-1 rack using the power of 600 houses!), the amount of current required becomes enormous. At 800V, you can deliver the same amount of power using far less current.
That brings some pretty big benefits. You need less copper, thinner cables and smaller busbars, which saves space, weight and cost. More importantly, lower current means much lower electrical losses, because resistive losses rise with the square of the current. So there is less energy wasted as heat.
The 800V approach should also allow Nvidia to simplify the power chain, with fewer stages of voltage conversion. That improves efficiency and reduces the amount of power infrastructure taking up valuable space in the rack, so they can stuff more GPUs inside
Ultimately, it's about making much more powerful AI systems practical. As racks move towards 500kW and eventually 1MW, the old architecture becomes increasingly difficult to manage. 800V gives Nvidia a much more scalable way of delivering all that power efficiently.
