# Demand Response Demand response is the practice of reducing or shifting consumption when the grid is under stress, in exchange for a lower price, a faster connection, or a payment. For most industrial loads it is awkward, because a furnace or a production line resents interruption. For compute it is unusually practical, and that fact is one of the strongest cards the industry holds. > [!note] The one-sentence version > A load that can bend when the system is tight is worth more than a load that cannot, and compute can bend further than almost anything else its size. ## Why compute can bend Three properties make it possible. Training and batch work are not latency-sensitive: a job that finishes an hour later is still finished, and [[wiki/Checkpointing|checkpointing]] means pausing costs minutes of recomputation rather than the whole run. Work is geographically portable: a request or a job can be served in another region where power is plentiful, so the load moves rather than stops. And the facility holds energy storage and often generation of its own, so reducing grid draw does not require reducing computing at all; see [[wiki/Data Center Batteries|Data Center Batteries]] and [[wiki/Behind-the-Meter Generation|Behind-the-Meter Generation]]. Inference serving a live product is a different matter, and honest flexibility commitments separate the two: a facility can offer deep curtailment of training and shallow curtailment of serving, and say so in the contract. ## Prerequisites - **Firm against flexible service:** firm service guarantees delivery and requires the network to be built for the full peak; flexible service trades curtailment rights for a lower price and an earlier connection. - **Coincident peak:** the facility's draw at the moment the system peaks, which is what the grid must actually be built for; shaving it is worth more than reducing average consumption. - **Notification window:** how much warning the operator gets, from a day ahead to seconds; shorter windows pay more. - **Baseline:** the consumption the reduction is measured against, a quietly contentious matter in every program. - **Ancillary services:** fast products such as frequency regulation and reserve, where storage and rapidly curtailable load are valuable. - **Spatial shifting:** moving work between regions rather than reducing it, the mechanism with no cost to total output at all. ## How it works in practice The utility or grid operator signals an event. The facility's control system responds according to a prepared plan: pausing or slowing training jobs, deferring batch work, draining a region's queue to another region, discharging storage to serve the load, or starting on-site generation. Computing capacity is reduced only as a last step, and often not at all. Measurement and settlement follow, and the facility is paid or receives the tariff benefit it contracted for. The scheduler is where this becomes real. A cluster manager that understands power as a schedulable resource alongside accelerators and memory can accept a megawatt limit as an input and plan the work within it; see [[wiki/Job Scheduling|Job Scheduling]] and [[wiki/Slurm|Slurm]]. That is the piece of engineering that converts flexibility from a promise into an operating capability. ## Why it is good practice rather than a concession A utility planning for a large new load must build for its peak. A load that reliably shaves its coincident peak requires less new transmission, fewer transformers, and less generation, which lowers the cost borne by everyone on the system and shortens the queue for everyone behind it. The facility gets a faster connection and a better price; the ratepayers get less infrastructure to pay for; the grid gets a resource. The arrangement is one of the few in infrastructure where every party gains, and it is available now rather than pending. ## Relationships - **traded in:** [[wiki/Interconnection Queue|Interconnection Queue]] · [[wiki/Utility Interconnection|Utility Interconnection]] - **enabled by:** [[wiki/Checkpointing|Checkpointing]] · [[wiki/Job Scheduling|Job Scheduling]] · [[wiki/Data Center Batteries|Data Center Batteries]] · [[wiki/Behind-the-Meter Generation|Behind-the-Meter Generation]] - **supports:** [[wiki/Grid Inertia|Grid Inertia]] · [[wiki/Electrical Supply and Reliability|Electrical Supply and Reliability]] - **priced through:** [[wiki/Power Purchase Agreement|Power Purchase Agreement]] - **community context:** [[wiki/What a Data Center Gives Back|What a Data Center Gives Back]] - **source articles:** [[articles/Vertically Integrating an AI Superpower from AI Factory to Citizen|Vertically Integrating an AI Superpower from AI Factory to Citizen]] · [[articles/The Gigawatt Is the New Measure of War|The Gigawatt Is the New Measure of War]] - **collection route:** [[collections/American Infrastructure|American Infrastructure]] · [[collections/American|American Collection]] ## Evidence posture - **Established:** firm against flexible service structures; coincident peak as the planning quantity; notification windows and their pricing; baseline measurement as the settlement basis; ancillary service products; checkpointing permitting low-cost interruption of training. - **Established (datestamped):** as of October 2026, flexible large-load tariffs trading curtailment rights for earlier connection are in force or in active proceedings in multiple United States jurisdictions. - **Strongly indicated:** power-aware scheduling as an operating capability at scale; operators describe it and sustained performance records remain short. - **Unresolved:** how much flexibility the industry ultimately delivers against what it commits to; the deciding evidence is settlement data across several system peaks. ## Sources - [How the Electrical Grid Is Being Rebuilt for AI](https://www.bloomberg.com/) — Bloomberg Primer, October 2026; demand growth and grid constraint. - Utility flexible-load tariff filings and regional grid operator demand-response program rules, as reviewed October 2026.