ABB's Infinitus and the rack-heat problem: same war, different front

ABB just launched Infinitus, a source-to-rack DC portfolio built around solid-state transformers. The thesis is simple: every power conversion step bleeds energy as heat, so cut the conversion steps. Their BCG-cited modeling says a 5% efficiency gain frees 25 MW for compute on a 500 MW campus, with the first full sites expected in two to three years.

Last week we published a lab note, "Four levers on AI rack heat," arguing a ~100 kW rack envelope could in principle come down to ~18.75 kW if four levers all move: 0.8V→0.4V voltage scaling, co-packaged optics, 48V distribution, and adiabatic CMOS on the control plane. It's a planning envelope, not a measured rack, and it says so on the page.

The overlap is real: our 48V distribution lever is the same thesis as Infinitus — cut conversion stages, cut heat — applied one level down, inside the rack. ABB productizing this layer means the industry now agrees on it.

The difference is where it stops. Infinitus ends at the rack door. The chip-level levers — voltage scaling, co-packaged optics, adiabatic logic — are where the rest of the heat lives, and that's the part nobody is productizing yet.

Stack them honestly: ABB fixes the facility side, we described the rack side. Theirs is vendor modeling with a product roadmap; ours is a conditional envelope. Both point the same direction.