This research paper presents a method for quantifying the electricity-demand flexibility that AI data centers can provide to the power system. The approach reschedules batch computing workloads subject to constraints on delay, uninterrupted execution, and available CPU, GPU, and memory resources. Task-scheduling decisions are translated into a server power profile, which is then used to assess peak-shaving potential and the duration for which a load reduction can be sustained, using real GPU-cluster traces.
Key Takeaways
Shifting batch workloads can position an AI data center as a controllable demand-flexibility resource for the power system without violating workload execution constraints.
Assessing flexibility requires joint consideration of CPU, GPU, and memory utilization: in the scenario examined, considering GPUs alone materially understates the available power-reduction potential.
Permissible delay for batch jobs is a key parameter: increasing it raises the achievable load reduction, but requires shifting more jobs and extending the duration of those shifts.
Finer time granularity preserves more workload-scheduling options, while aggregating intervals reduces the maximum achievable flexibility.
There is a direct trade-off between the required magnitude of power reduction and the length of time for which the data center can sustain that reduction continuously.
Key Figures
Power demand of modern hyperscale AI data centers
свыше 100 MW
Cited in the introduction as a potential power requirement for contemporary facilities.
Scale of prospective AI data centers
гигаваттный
According to the paper, some future facilities may reach this scale.
Power reduction demonstrated in cited experimental studies
25 %
Example cited by the authors of software orchestration of GPU workloads while maintaining quality of service.
Duration of the power reduction
3 часа
Applies to the power-reduction example cited in the introduction.
Number of jobs in the daily Alibaba Cluster Trace GPU v2020 dataset used
7 828 задач
One day of data used in the simulations.
Original time resolution of the load profile
1 с
Second-level resolution of the source computing traces.
Typical power-system scheduling intervals
5, 15 или 60 мин
Intervals against which the processed data center load is compared.
Peak-shaving assessment window
15:00–15:25
Flexibility-provision window specified in the experiment.
Case 1 flexibility at γ=1.0
21,71 kW
Only GPU power is considered.
Case 2 flexibility at γ=1.0
32,47 kW
CPU, GPU, and memory are considered.
Increase in Case 2 flexibility relative to Case 1 at γ=1.0
49,6 %
Increase resulting from accounting for CPU, GPU, and memory.
Case 2 flexibility at γ=0
10,76 kW
No additional delay is permitted for batch jobs.
Case 1 flexibility at γ=0
8,690 kW
GPU-only case.
Number of shifted jobs in Case 1 at γ=0
33 задачи
Result of the short-term peak-shaving analysis.
Average shift duration in Case 1 at γ=0
97,42 мин
For shifted jobs reported in Table I.
Number of shifted jobs in Case 2 at γ=0
37 задач
Result for the case accounting for CPU, GPU, and memory.
Average shift duration in Case 2 at γ=0
90,27 мин
For shifted jobs reported in Table I.
Sustained-reduction duration at γ=0.4 for a required reduction of 10 kW
340 мин
Maximum continuous interval over which flexibility can be provided.
Sustained-reduction duration at γ=0.4 for a required reduction of 15 kW
195 мин
Maximum continuous interval over which flexibility can be provided.
Sustained-reduction duration at γ=0.4 for a required reduction of 20 kW
120 мин
Maximum continuous interval over which flexibility can be provided.
Practical Value for Data Center Owners
For AI data center owners, the work provides a methodology for testing what level of contracted demand flexibility can realistically be offered to a grid operator or used to limit peaks—both in terms of the magnitude of the power reduction and the time for which it can be sustained. Implementation requires CPU, GPU, and memory resource telemetry at sufficiently fine granularity, as well as formally defined permissible delay windows for batch workloads; a GPU-only assessment or coarse scheduling intervals can distort the available potential.
The paper introduces FissionReady, a scheduler for data centers supplied by a multi-module small modular reactor (SMR) plant. It jointly manages the output of each reactor module, deferrable batch workloads, and grid purchases while accounting for fuel age, iodine-xenon dynamics, and reactivity margin. In simulations, the approach eliminates forced outages and missed batch-job deadlines while reducing water use and grid-electricity costs relative to a comparable baseline configuration.
Small Modular Reactors (SMRs)Data Center Power SupplyGeneration Load FollowingData Center Microgrids
Wärtsilä’s presentation examines the use of modular gas and dual-fuel power plants alongside energy storage to supply data centers, including AI workloads. It links constraints on grid connections to growing demand for primary onsite and off-grid generation, and cites several projects in Ireland and the United States. The material emphasizes capacity availability, modularity, part-load efficiency, lower water consumption, and the potential to transition to sustainable fuels over time.
Data center power supplyAI workloadsOnsite generationOff-grid data centers
This AVK and Wärtsilä Energy white paper examines constrained grid capacity as a major barrier to the expansion of European data centres, particularly facilities serving AI workloads. It proposes on-site microgrids combining dispatchable generation, BESS and renewables to accelerate site energisation before—or independently of—a grid connection. The paper compares power-system scenarios and four supply configurations for a representative AI data centre in terms of cost, reliability, land use and carbon footprint.
Data Centre Power SupplyAI Data CentresMicrogridsDispatchable Generation