One August evening, Silicon Valley Power sent a signal to reduce electricity consumption to an AI factory in Santa Clara. Emerald AI’s Conductor platform slowed down and rescheduled lower-priority jobs while keeping critical AI services running. Consumption automatically dropped from four megawatts to three megawatts; no operator intervention was required. Silicon Valley Power has sent the factory more than 200 demand signals since then, achieving results every time.
The aim of the approach introduced by NVIDIA at the AI Infra Summit is to meet the electricity needs of AI factories without waiting for new transmission lines. According to Lambda’s first validation, announced the same day, token generation can increase by 24% when a fixed power budget is managed intelligently.
- NVIDIA DSX MaxLPS: Increases token generation within a fixed power budget by monitoring GPU and rack consumption in real time.
- NVIDIA DSX Flex: Changes job priorities according to grid signals and protects high-priority tasks.
- NVIDIA DSX OS: Provides open-source software for lifecycle management, operational consistency, health automation, and resilience in AI factories.
- NVIDIA DSX Sim: Enables designs to be modeled and bottlenecks to be identified before physical installation.
- NVIDIA DSX Reference Designs: Provides validated architectures covering compute, networking, storage, and facilities.
Why it matters
This approach aims to reduce the need for AI data centers to be nothing more than a constant and inflexible load for the grid. For grid operators, its value lies in being able to reprioritize workloads rather than cutting service off completely when consumption demand changes. Data center operators, meanwhile, gain the option to manage computing capacity and operational continuity together within the same power limit. However, the reported results are based on demonstration and validation examples from NVIDIA Blog and its partners; therefore, it remains unclear how the method will perform across different facilities, different workloads, and at a broader scale. The software, simulation, and reference design components of the DSX family also show that the solution is not limited to real-time consumption adjustment.
Background
NVIDIA is not a new name in the FikirPilot archive: we have published 13 news stories mentioning it in the last 90 days; the most recent one is dated 27 September 2026.