Silicon Valley Power treats an AI factory as a dispatchable grid asset — 200+ automated load cuts

Silicon Valley Power has repeatedly signaled an AI factory to cut power on demand — more than 200 successful demand-response events so far — with Emerald AI’s Conductor platform throttling flexible workloads while priority jobs keep running. On the first live cut, consumption dropped from ~4 MW to ~3 MW in under a minute with no operator intervention. Nvidia says the run is on its Eos AI factory in Santa Clara and frames it as commercial proof that AI factories can behave as dispatchable grid resources, not just fixed loads. Why it matters: This is the opposite playbook to permitting freezes: instead of pausing new megawatts, utilities recruit existing AI load to flex when the grid is stressed. Nvidia is generalising the pattern as DSX Flex (Emerald Conductor as an early implementation), with a dedicated ~96 MW Vera Rubin deployment planned at its Manassas, Virginia AI Factory Research Center. Verified facts (Nvidia / Emerald / SVP): SVP’s Flexible Load Interconnect Program is described as the first commercial utility programme designed to treat AI factories as dispatchable resources.
Conductor responds to grid signals in under a minute against a predefined workload hierarchy (batch/low-priority yields; high-priority inference continues).
Separately at the AI Infra Summit (15 Sep 2026), GPU cloud Lambda reported a DSX MaxLPS validation: 19 nodes at ~85% power inside the same facility budget as 16 full-power nodes to ~24% more cluster token throughput and ~23% better performance per watt on HGX B200. Caveats / inference: The Santa Clara run is Emerald Conductor on Eos — Nvidia presents it as a precursor to DSX Flex, not a finished DSX Flex install. Open Nvidia materials do not disclose which tenant workloads were cut, nor how often priority inference would have blocked deeper reductions. Lambda’s +24% figure is a controlled cluster validation, not a production hyperscale guarantee; Nvidia’s “up to 40% more GPU capacity” claim for Vera Rubin NVL72 is a projection for suitable environments. The signal: Power is becoming a software scheduling problem. The companies that can shed load on cue — and squeeze more tokens from a fixed megawatt budget — may clear interconnection queues faster than those that only ask for more capacity. Complementary to Texas’s pause on new data-centre approvals amid grid crunch. Sources: Nvidia — From Megawatts to Tokens (15 Sep 2026) · Nvidia — AI Infra Summit / DSX · Nvidia case study — Lambda / DSX MaxLPS