NVIDIA describes computing centres capable of shifting certain workloads when the power grid is constrained. For business users, this approach puts task priority back into the discussion.
Not every AI task needs to finish at the same moment. In a publication dated 15 September 2026, NVIDIA describes the automatic adjustment of certain computing workloads to power grid signals. The Santa Clara example relies on Emerald AI Conductor: lower-priority jobs yield capacity while the services deemed critical continue.
Distinguishing demonstration from generalisation
NVIDIA also reports a 24% gain in token throughput in a validation carried out at Lambda, at a constant power budget. This result corresponds to a specific configuration. It is not a promise of an identical gain for every company.
The publication distinguishes the system already in use in Santa Clara from DSX Flex, which is intended to extend this logic. This nuance avoids presenting a product roadmap as a generalised deployment.
Priority is also a business decision
For a house, the point is not to choose for itself how a data centre operates. It is to specify the deadlines genuinely required by the services it buys. A reply to a customer who is present and the overnight filing of a document archive do not carry the same time requirement.
Defining these differences makes it possible to discuss service quality differently. The contract can distinguish urgent operations, those that tolerate a delay, and the evidence expected in the event of an interruption. Technical optimisation then becomes compatible with an explicit choice about the experience.
Electrical efficiency does not, however, sum up the environmental impact of a use. One must also look at the volume of work requested and its usefulness. Producing more with a given amount of power does not establish a fall in total consumption. The question therefore remains twofold: making better use of the available computing, and deciding which jobs deserve to mobilise it.
Cette publication est également disponible en :
