NVIDIA published three technical explainers on the same day. They sit at different layers - model, execution, interconnect - and none of them is about making the computation faster.
The first is about choosing a model. Nemotron 3.5 Lightning, its example, activates only 3B of its 30B total parameters to process a single token. The piece compares dense models, which run everything every time, against Mixture-of-Experts models, which switch on only the parts they need, framed by active parameter count and throughput. The second covers the deterministic execution of Groq 3 LPX on Vera Rubin, which fixes timing in advance to remove wasted waiting. The third is NVLink 6. In large-scale training every GPU in the cluster moves in lockstep, so one failed node halts the entire run; NVLink 6 stacks redundancy and recovery from the link layer up to the system layer so that training survives a failure.
Three separate technologies that, lined up, point the same way. Not faster arithmetic, but how to spend a fixed budget of resources and how not to stop. The caveat is that only the latter two actually name power as the constraint; the model-choice piece stays in the language of throughput. The measure of a system is shifting from how fast it goes toward whether it can keep going.
What was published is design reasoning, not measurements. The test comes when operators can produce their own numbers - work done per watt, and how far a training run actually got after a failure - from their own environments.