The global surge in AI development is transforming not only how we compute, but how our power systems behave. In their 2025 study, Instability Risks from Programmable AI Load Ramping in Low‑Inertia Grids, Dr Dlzar Al Kez and Professor Aoife Foley reveal a challenge hiding in plain sight: AI data centres are no longer passive electricity consumers. They are fast acting, logic‑driven electrical loads capable of reshaping grid dynamics in real time. 

Drawing on detailed time‑domain simulations of a modified New England test system, the authors show how rapid AI training ramps especially in grids dominated by inverter‑based resources can trigger voltage depressions, RoCoF spikes, and system‑wide instability even without a generator fault. Their work exposes a new class of demand‑driven disturbance and argues that without visibility, coordination, and storage, AI loads could become a critical stressor on tomorrow’s low inertia grids. 

This “Closer Look” unpacks the paper’s insights, the authors’ motivations, and the implications for utilities, policymakers, and the rapidly expanding AI sector. 

Q: What motivated you to study AI data centers as a potential risk to power grids?

A: What motivated this work was a clear gap between how we model demand and how it behaves today. Traditionally, loads have been treated as something passive, slow-moving, and largely predictable.

But what we’re seeing with AI data centers is very different. These are large, power-electronic loads driven by software, not by system conditions. They can ramp rapidly, and at large scale those ramps can reach system-relevant magnitudes, often without sufficient visibility for system operator.

Within the Avantern Group at The University of Manchester, Professor Foley and my broader work on resilient net-zero infrastructure and low-inertia power systems made it increasingly clear that this emerging form of demand could not be treated using legacy assumptions.

At the same time, system inertia is decreasing as solar and wind replace conventional generators. That combination, faster, less predictable demand and lower system resilience, is what really raised the concern. It’s not just a scaling issue; it’s a change in demand itself. This builds on my continuing research on data centers, fast frequency response, and high-IBR power systems.

Q: In simple terms, what is the main takeaway about how AI workloads could affect grid stability?

A: The main takeaway is that AI workloads can behave more like disturbances than traditional demand.

Instead of slowly varying, they can change rapidly and autonomously, which can trigger frequency and voltage deviations even when no conventional fault or generator trip has occurred.

In low-inertia systems, those fast changes become harder to absorb, meaning instability can emerge purely from how demand behaves, with no generator fault required. That’s a new kind of risk, and one that current grid codes and protection systems weren’t designed with in mind.

Q: Why is battery storage important in your scenarios, and what does it mean when you say “coordinated” storage across sites?

A: Battery storage helps by absorbing or smoothing those rapid changes in demand.

In our scenarios, uncoordinated storage, where each data center acts independently based only on its own conditions, only partially mitigates the problem. But when sites share timing signals or respond to a common grid frequency threshold, the combined effect is much more stabilizing. That alignment is what we mean by coordination: not just having storage, but having it act as part of a coherent system response.

So it’s not just about having storage, it’s about how that storage behaves as part of a wider system response rather than acting independently.

Q: How feasible is it today for data centers to share timing or load information with grid operators, and what are the practical challenges?

A:Technically, it’s feasible. The challenge isn’t really the technology; it’s the framework around it.

A lot of these facilities operate behind the meter, and their internal processes are driven by commercial and operational priorities. There’s often limited obligation to share real-time information about how workloads behave.

So the issue is less about capability and more about incentives, standards, and regulatory structures. Until there’s a clear framework requiring or enabling that visibility, sharing will remain voluntary and inconsistent, which is precisely where policy has a role to play.

Q: If a regulator asked for practical steps, what would be the easiest first actions to reduce the risk you identified?

A: The first step is improving visibility. If operators can’t see how demand is behaving, they can’t plan or respond effectively. Even basic transparency requirements around ramp rates and operational patterns, potentially through extensions to existing grid codes, would be a significant step forward.

From there, introducing soft limits on ramp rates, or requiring facilities above a certain size to demonstrate local mitigation capability such as storage or demand controllability, would bring large AI loads closer into alignment with system needs. These don’t require new institutions; they require existing ones to extend their scope.

Q: What are the main limitations or uncertainties in your study that readers should keep in mind?

A: Like any modeling study, this is based on representative scenarios rather than exact real-world behavior.

At the Avantern Group we’ve tried to capture realistic dynamics based on available data and reported events, particularly in how we represent fast-ramping load profiles and system conditions, but actual data center behavior can vary depending on design, workload type, and control strategies, and granular operational data from live facilities remains difficult to access, which itself is a limitation worth noting.

The results should therefore be seen as indicative of risk and system sensitivity, rather than precise predictions of specific events.

Q: How generalizable are your findings to other grids, regions, or configurations beyond the specific case you analyzed?

A: The specific numerical results will vary by system, but the underlying mechanisms are broadly applicable. Any grid with growing inverter-based generation and large, fast-ramping loads will face similar pressures, and that increasingly describes systems across GB, Ireland, parts of Europe, and data center-dense regions globally.

The severity depends on system strength, inertia levels, and whether these loads are geographically concentrated, as clustering amplifies their local impact on the grid.

So while our case study is specific, the dynamics it captures are not.

Q: What roles should different stakeholders (data centers, utilities, policymakers) play in addressing these risks?

A: Data centers need to recognize that at scale, they are no longer just consumers, they are active participants in system behavior. That comes with a responsibility for transparency and some level of coordination with system operators.

Utilities and system operators need updated tools to model and monitor these new demand profiles in real time; current frameworks weren’t built with this kind of load in mind.

And policymakers need to update connection standards and market structures, so these risks are properly accounted for, rather than sitting outside the regulatory perimeter. The message is not that AI data centers cannot be connected, but that they need to be treated as active infrastructure, not ordinary passive demand.

Q: What future research directions do you see as most valuable to pursue next?

A: The most immediate gap is real-world load data. High-resolution behavioral data from operating facilities would significantly improve how we model and anticipate these risks, and right now that data is largely inaccessible. This is a core research and innovation direction for the AVANTERN Group: modeling emerging infrastructure risks before they become operational failures.

Beyond that, a key direction is developing coordination mechanisms that are practical enough to implement, not just theoretically optimal. How do you design something that works within commercial constraints, varying ownership structures, and existing market rules?

More broadly, both questions point to the same underlying need: rethinking how demand is represented in planning and operational models. That’s a research agenda that extends well beyond AI data centers.

Q: If you had to summarize the study for a broad audience in one or two sentences, what would you say?

A: Large AI data centers are changing how electricity demand behaves, making it faster, less predictable, and more dynamic and system active. If we continue to treat that demand as passive, we risk missing a growing source of instability in modern power systems. Our study shows that, unless these fast-ramping loads are made visible, coordinated, and supported by storage or grid-forming capability, they could become a new source of instability in low-carbon power systems.

Q: What do you think SSRN contributes to the world of modern research and scholarship?

A: For research like ours, at the Avantern Group, where the policy implications are immediate and the field is moving quickly, SSRN’s value is in closing the gap between when work is done and when it reaches the people who need it. Peer review is important, but it takes time, and decisions about grid codes, data center connections, and energy policy aren’t waiting.

More broadly, SSRN encourages the kind of cross-disciplinary conversation this problem genuinely requires. It sits at the intersection of power systems, computer science, and policy, and having a platform where researchers across those fields can engage with the work early makes a real difference.

ABOUT THE AUTHORS

Professor Aoife Foley

Professor Aoife Foley is Chair in Net Zero Infrastructure at The University of Manchester, with a joint appointment to the Departments of Electrical and Electronic Engineering and Civil and Engineering Management, and is Chief Executive Officer (CEO) of the AVANTERN Group. Her work focuses on the resilience, operation, financing, and future planning of net-zero infrastructure systems across the energy, transport, telecommunications, and digital sectors, examining how emerging technologies, AI-driven demand, renewable integration, and increasingly complex infrastructure networks interact within rapidly decarbonising societies.

With more than 30 years of experience spanning engineering, infrastructure delivery, policy, and academia, Professor Foley’s work consistently bridges engineering, policy, and real-world system operation. Before moving into academia full-time in 2011, she worked across major infrastructure, telecommunications, and energy projects in both public and private sectors. She later became Professor in Energy Systems Engineering at Queen’s University Belfast and previously served as Editor-in-Chief of Renewable and Sustainable Energy Reviews (Elsevier). Ranked among the Stanford/Elsevier Top 2% of Scientists globally, her work continues to influence international discussions on resilient infrastructure, future electricity systems, and the wider challenges associated with net-zero energy transitions.

You can see more work by Aoife Foley on her SSRN Author page here.

Dr Dlzar Al Kez 

Dr Dlzar Al Kez, CEng, MIET, FHEA, is a Research Associate in Net-Zero Infrastructure at The University of Manchester and Chief Technology Officer (CTO) of the AVANTERN Group. His research focuses on the stability and resilience of low-carbon electricity systems with high penetrations of inverter-based resources, battery energy storage, and rapidly changing electrical demand. He works across power system modelling, IBR integration, low-inertia grid behaviour, frequency stability, grid-code compliance, and the operational impacts of emerging digital infrastructure such as AI data centres. Using DIgSILENT PowerFactory, Python, MATLAB, and system-level modelling approaches, he supports technical assessment of grid connection risks, dynamic performance, system strength, and stability challenges in renewable-dominated networks. Dr Al Kez has authored more than 50 peer-reviewed publications and serves as Associate Editor for IET Smart Grid and Smart Grids and Sustainable Energy, alongside roles as Publishing Ethics Advisor and Subject Matter Expert for Elsevier. His work bridges academic research, industry-facing power system studies, and advisory support for utilities, infrastructure developers, regulators, and organisations managing complex grid integration challenges.

You can see more work by Dr Dlzar Al Kez his SSRN Author page here.