An AI campus can ease a short-lived grid constraint by delaying work that does not need an immediate result. If a grid is stressed for twenty minutes, moving a non-urgent training, evaluation or batch-inference task can reduce the immediate load. The opportunity depends on which tasks can actually be deferred and how their recovery is managed.[1][2]
JERA’s investment in Emerald AI describes software that coordinates computing workloads and on-site energy with real-time grid conditions while maintaining priority applications. EPRI’s Flex MOSAIC seeks a common way to classify large-load flexibility. FERC is separately examining how large loads—generally above 20 MW—should connect, pay and potentially receive faster treatment when they can curtail.[1][2][3]
A flexibility offer needs more than a megawatt number
| Measure | Value | Context |
|---|---|---|
| Power reduction | How many MW can move or stop? | The instantaneous change that the grid may see.Source 2 |
| Notice | How quickly can the change begin? | Seconds, minutes or hours can serve very different grid needs.Source 2 |
| Duration | How long can it last? | A ten-minute response is not a four-hour resource.Source 2 |
| Frequency | How often can it be called? | Repeated calls may delay work or exhaust operational headroom.Source 2 |
| Recovery | What happens afterwards? | Deferred tasks can create a rebound peak if they all restart together.Source 2 |
Shifted energy is not eliminated energy
Consider a 100 MW campus with 30 MW of genuinely deferrable work. Moving that 30 MW for two hours shifts 60 megawatt-hours out of the constrained period. It does not necessarily reduce the energy needed to finish the work. If the same tasks run later, the benefit is timing, not disappearance.
That distinction prevents two accounting errors. First, a temporary drop should not be reported as a permanent efficiency improvement. Second, the delayed work should not be restarted without measuring the rebound. A credible controller must schedule the recovery as carefully as the curtailment.
Which AI work is plausibly flexible?
| Workload | Likely flexibility | Main safeguard |
|---|---|---|
| Interactive user requests | Low | Protect latency and availability commitmentsSource 1 |
| Safety, fraud or infrastructure monitoring | Very low | Do not interrupt time-critical detectionSource 1 |
| Batch evaluation and synthetic-data generation | Potentially high | Preserve deadlines, data integrity and restartabilitySource 1 |
| Model training checkpoints | Conditional | Move only at safe checkpoint boundaries and account for recovery costSource 1 |
| Cooling and storage support loads | System-dependent | Respect thermal and equipment limitsSource 2 |
The contract must describe performance on both sides
FERC’s large-load proceeding asks whether customers that agree to flexibility or curtailment should move through studies faster and how costs should be assigned. A useful arrangement needs enforceable performance: the grid must specify the signal and value, while the data centre must specify response, duration, telemetry and consequences for non-delivery.[3][4]
The same measurement should protect the customer. A campus that gives the grid a dependable resource should not be treated as if its full maximum demand were inflexible at every hour. Conversely, a marketing claim about “AI-aware energy management” should not earn capacity credit until the response has been demonstrated under representative conditions.
Workload flexibility makes grid capacity more useful when the computing service can honour a defined commitment. The job must be able to wait, the grid must be able to rely on the response, and the postponed work must return without creating the next demand spike. Those conditions determine the value of the offer.
Sources & further reading
- JERA Invests in U.S.-Based Emerald AI to Advance Energy Solutions for AI Data Centers JERA · 26 Aug 2026 · Accessed 8 Oct 2026
- Flex MOSAIC Electric Power Research Institute · Accessed 8 Oct 2026
- Interconnection of Large Loads to the Interstate Transmission System, Docket No. RM26-4-000 Federal Energy Regulatory Commission · Accessed 8 Oct 2026
- FERC Launches Aggressive Targeted Action to Speed Large Load Integration Federal Energy Regulatory Commission · 18 Jun 2026 · Accessed 8 Oct 2026
- Key Questions on Energy and AI International Energy Agency · 16 Apr 2026 · Accessed 8 Oct 2026