The Regenerative Strategist

Welcome and thank you for reading The Regenerative Strategist. A weekly update on harnessing the power of decarbonization, digitization, and enhanced data collection, delivering buildings of the future powered by renewable energy & automation.

Want to stay informed?

Subscribe to this series using the button above and let us know what you want to hear about next week using #UrbanAO in the comments below.

Introduction

AI data centers just learned a cruel lesson: the grid is not a plug; it’s a courtroom. At gigawatt scale, “interconnect” stops being an engineering step and becomes a political event—visible to ratepayers, communities, and regulators alike. If you lean on the public grid, you inherit public backlash and rate fights. If you go off grid, you inherit air permits, emissions scrutiny, and lawsuits.

This is not metaphor. It is an emerging operating reality across multiple regions: queues, tariffs, hearings, and litigation are now embedded in the build path for compute. The bottleneck is no longer only megawatts or transformers; it is permission—often expressed as process: interconnection study timelines, cost allocation rules, permitting windows, and enforceable reliability obligations.

Meanwhile, as the energy system becomes more constrained and contested, AI is showing up on the other side of the table: inside grid operations themselves. Grid operators are adopting machine learning for load and renewable forecasting, anomaly detection, and decision support; vendors are pushing generative AI into control-room tooling and planning workflows; and regulators are beginning to ask a new kind of question: if an AI-influenced decision materially affects reliability, prices, or access, who is accountable—and what evidence is admissible when challenged?

What follows is a rigorous map of how we got here, what AI is doing in grid operations, why disputes are escalating, and how governance can catch up—before the next “interconnection” becomes a referendum on legitimacy.

1. How grid control became a courtroom

The electric grid has always been governed as much as it has been engineered, but the balance has shifted repeatedly over the last 60 years.

Key inflection points in modern reliability governance traces back to cascading outages, including the 1965 Northeast blackout, which helped catalyze industry-wide coordination and the formation of reliability institutions that later became the modern reliability regime. In North America, that evolution culminated in mandatory reliability standards under Section 215 of the Federal Power Act (added by the Energy Policy Act of 2005) and the certification of the North American Electric Reliability Corporation as the Electric Reliability Organization, with enforcement oversight by the Federal Energy Regulatory Commission.

In parallel, competitive market reforms pulled the grid further into legal and administrative adjudication. U.S. open-access orders created the foundation for competitive wholesale markets and changes in “who controls what” across generation, transmission access, and market rules. Once you have market rules, you also have market disputes—about jurisdiction, cost allocation, market design, and the boundaries between wholesale and retail authority. A canonical example is the Supreme Court’s decision upholding demand response compensation rules: a reminder that “dispatch” is not only an algorithmic outcome, but also a justiciable policy choice embedded in tariffs and law.

Europe’s integration followed a different institutional path, but similar dynamics. Regional coordination and harmonized network rules were built through a framework that assigns roles to TSOs, national regulators, and cross-border bodies—helping turn technical coordination into codified governance. The ENTSO-E and the Agency for the Cooperation of Energy Regulators[17] formalize cross-border governance and dispute-adjacent processes around market coupling, balancing, and enforcement against market abuse.

So why does this matter for AI and data centers?

Because the same legal “surface area” that constrained and shaped competitive power markets is now being activated by load growth that is both (a) unusually large and (b) unusually fast. The International Energy Agency has repeatedly highlighted the scale and pace of data center-driven electricity demand growth, stressing both infrastructure implications and the need for smarter integration via operational flexibility. In its analysis, data centers are becoming “major players” in electricity systems and are exploring flexibility measures to shorten lead times for grid access in capacity-constrained regions.

When a new facility is 240 MW, 1 GW, or 2 GW, the grid conversation changes. Recent examples underline the point:

• Meta began construction on a roughly $10B, ~1GW data center campus in Lebanon, Indiana.
• Nebius announced a 240MW data centre near Lille, redeveloping a former industrial site in Béthune.
• In Louisiana, a regulator-facing dispute emerged over a Meta data center project and associated gas generation investments, including concerns raised by Earthjustice about ratepayer risk and cost responsibility.

These are not “just” engineering projects. They are arguments about who gets to use the system, on what terms, and with which obligations.

The grid is becoming a courtroom not because stakeholders became more litigious, but because the grid has become a scarce public resource under stress—where the legitimacy of allocation decisions must be defended.

2.What AI is actually doing in grid operations

Despite the hype, AI in the grid today is mostly not a robot dispatcher pulling levers. It is primarily machine learning that improves (or sometimes destabilizes) the decision inputs that operators and market systems rely on.

A useful baseline is the reliability community’s framing: AI/ML can help, but it must be integrated with clear human oversight and responsibility. In a 2024 white paper, the North American reliability community emphasizes that operators should be involved in decision-making when AI/ML is tested and implemented, and that operators should retain final authority over AI/ML-generated actions.

From there, the dominant operational use cases cluster into four practical buckets:

Load and renewable forecasting (the “probability layer” for dispatch and reserves).

Forecasting is the most mature AI/ML deployment area because it is measurable, repeatable, and directly monetizable in reliability outcomes (reserve sizing, commitment decisions, congestion anticipation). For example, the PJM Interconnection has publicly discussed machine-learning approaches in load forecasting (including methods like XGBoost and neural networks) and links forecasts to operational and market timelines (e.g., day-ahead processes and very short-term forecasts used in security-constrained dispatch).

In Great Britain, the system operator (now operating under the National Energy System Operator brand) has collaborated with the Alan Turing Institute to use machine learning to improve renewable power forecasts—explicitly positioning ML as a tool for balancing a renewables-heavy system. Vendor deployments also highlight “ML for operational needs,” such as inertia metering and forecasting solutions used to inform short-term grid operation and longer-term assessment.

Dispatch and operational decision support (from “optimization only” to “optimization + learned recommendation”).

Traditional grid dispatch is already heavily automated through optimization (e.g., security-constrained economic dispatch). The more meaningful AI shift is upstream: better forecasts, better state awareness, and better anomaly triage so the operator’s dispatch decision is based on richer, faster evidence.

This is also where generative AI is being pilot-tested—not as a replacement for EMS/SCADA math, but as a translation layer: summarizing alarms, mapping between models, improving operator decision support, and generating “explainable” narratives for why a recommendation was made. The National Renewable Energy Laboratory has published work exploring generative AI concepts for grid operations (including “eGridGPT”) while explicitly cautioning that deploying generative AI in critical infrastructure requires careful treatment of risks and trust.

Market bidding and automated participation (where AI becomes an economic actor).

AI intersects with markets in two ways: (1) forecasting price and congestion, and (2) automating bids for flexible assets like batteries, demand response portfolios, or hybrid plants.

A concrete example of automated bidding platforms in storage is the “real-time trading and control” concept used to submit market bids and perform dispatch control for value maximization. As these systems scale, regulators increasingly treat automated strategies as part of market conduct, not merely “software.” In Europe, the strengthening of cross-border market abuse investigation procedures under REMIT (and the ongoing focus on algorithmic trading governance) reflects this direction: market monitoring is adapting to faster, more automated behavior.

Fault detection and asset health (where AI changes outage probability, not just response time).

Utilities and grid operators have always used protection systems and deterministic rules. AI/ML adds pattern recognition at scale: detecting anomalies across phasor measurements, equipment sensors, and operational logs—often as part of predictive maintenance and diagnostic tooling. Reliability research communities have specifically surveyed practical AI/ML applications in protection and control, emphasizing the need for validation, integration discipline, and security in deployment.

In China, the “AI as dispatcher assistant” pattern is becoming more explicit at the distribution level. State media and industry accounts describe AI-enabled dispatch assistants that integrate anomaly detection, alerts, and workflow automation for distribution network dispatch.

The connective tissue across all four buckets is the same: AI changes how decisions are prepared, and—if governance is weak—AI can start to change who is blamed when things go wrong.

3.When AI decisions are contested

If the grid is a courtroom, then “contested AI” is less about a single model being sued and more about AI becoming a material contributor to decisions that are already litigable: interconnection access, cost shifts, environmental compliance, market outcomes, and reliability risk.

Below are the recurring dispute patterns, illustrated with real-world examples across jurisdictions.

Dispute pattern: Who pays for power—and who gets priority?

This is the most immediate “courtroom” flashpoint. Large loads can impose network upgrade costs and reliability obligations, and the allocation of those costs becomes a political fight.

In the U.S., the federal regulator has opened a docket on “Interconnection of Large Loads to the Interstate Transmission System,” explicitly positioning large-load interconnections as a topic requiring reforms and stakeholder input. It has also directed the PJM region to implement new rules governing connections of AI-driven data centers and other large consumers co-located near generation—explicitly citing reliability and cost concerns.

At the state level, regulators are also adapting. The Virginia State Corporation Commission approved creation of a new rate class for the biggest electricity users, including data centers—an explicit attempt to draw clearer boundaries around cost responsibility and rate design in a data-center-heavy system.

Dispute pattern: “Bring your own power” meets air permits and public law.

When a data center builds or procures dedicated generation, it may reduce exposure to grid queue delays—but it increases exposure to permitting, emissions scrutiny, and local legal action.

In Louisiana, controversy around a Meta data center project included concerns that costs from new gas-fired plants linked to the project could be shifted onto consumers, and that changes to financing terms could alter who bears risk over time—issues raised publicly by Earthjustice after regulators declined to open a probe.

Another vivid example involves xAI. Reporting in early 2026 describes regulator findings and threatened litigation tied to the use of methane gas turbines to power large data center facilities, with allegations that required air permits were not obtained and that the company argued exemptions.

These cases illustrate a structural point: “off-grid” is not “off-law.” It simply moves the locus of contestation from transmission tariffs to environmental and local administrative regimes.

Dispute pattern: Co-location and “behind-the-meter” becomes a jurisdictional puzzle.

Co-locating load next to generation (including nuclear units) is attractive: it can shorten time-to-power, reduce reliance on congested networks, and improve predictability. But it also raises hard questions: Is the load effectively bypassing broader network responsibilities? Does it reduce supply available to other customers? How should transmission service and cost allocation work?

This is not hypothetical. Corporate deals and regulatory scrutiny around DBM/BTM arrangements have already emerged in high-profile contexts, including data center investments adjacent to nuclear plants and arrangements subject to federal attention. In addition, regional operators are publishing stakeholder processes specifically on “co-located load and generation behind the same point of interconnection,” a sign that market rules and planning assumptions are being rewritten in real time.

Dispute pattern: AI as an economic actor triggers market conduct scrutiny.

Where bids are automated, the legal question becomes less “was the model accurate?” and more “was the behavior compliant?” Enforcement regimes for market manipulation were not built with modern ML systems in mind, yet they apply to outcomes regardless of whether a human or software initiated the behavior.

U.S. enforcement reporting continues to emphasize bans, disgorgement, and penalties for market conduct, illustrating that regulated energy markets remain enforcement-forward and punitive when behavior is deemed manipulative. Europe’s evolving REMIT procedures similarly underline that cross-border investigations and monitoring are adapting to complex, fast market behavior—including algorithmic trading concerns.

Dispute pattern: “Clean power” becomes a strategic control lever.

One of the most important shifts is that major AI and cloud actors are contracting for long-duration, firm electricity in ways that look increasingly like industrial-era vertical integration.

Examples include large nuclear PPAs and SMR agreements tied explicitly to data center growth:

• Amazon Web Services and Talen Energy announced a long-term agreement for up to 1,920 MW of carbon-free electricity from the Susquehanna Nuclear Power Plant to support AWS operations, alongside exploration of additional nuclear capacity pathways.
• Google signed an agreement with Kairos Power for advanced nuclear power from multiple SMRs, positioning nuclear as a long-term clean supply option for rising data center demand.
• Microsoft backed a restart pathway for Three Mile Island Nuclear Generating Station via a long-term power purchase agreement (as publicly reported), with subsequent federal lending support reported for the restart.
• Constellation Energy has described data center-driven demand growth and highlighted deals with major tech firms as part of its strategy, reinforcing that “AI load” is reshaping generation investment narratives.

The contested question is no longer “are you allowed to consume the power,” but “are you allowed to reorganize the system’s risk, cost, and carbon profile around your load?”

4. Governance, accountability, and the power shift

If AI is reshaping both grid operations and data center energy strategies, the central governance challenge is deceptively simple:

When an AI-influenced decision changes reliability outcomes, market prices, or access to infrastructure, can the decision be explained, audited, and legitimately defended?

Four governance layers are now converging on the grid:

Layer: AI risk management as an organizational discipline.

The strongest cross-sector baseline is the AI risk governance framing from the National Institute of Standards and Technology, which positions AI risks as socio-technical and calls for lifecycle risk management to promote trustworthy AI. For grid actors, the practical implication is that “model performance” is not enough; governance must cover context, monitoring, human oversight, and incident response.

Layer: “High-risk AI” regulation for critical infrastructure.

The EU AI Act explicitly classifies certain AI systems used as safety components in the management and operation of critical infrastructure—including electricity supply—as “high-risk,” reflecting a regulatory view that failures can have large-scale safety and societal impacts. For operators and vendors serving European energy systems, this implies structured obligations around risk management, documentation, and oversight—not optional “best practices.”

Layer: Reliability and cybersecurity regimes that already exist.

Grid AI governance does not start from zero; it must interlock with existing reliability and cyber obligations. In North America, mandatory reliability standards and enforcement mechanisms exist under federal law and the reliability regime, while cybersecurity requirements for the bulk power system are operationalized through Critical Infrastructure Protection standards. This matters because new AI tooling expands the cyber-attack surface—even when it improves operational decision support.

Layer: Adversarial risk and attack taxonomies.

Machine learning in critical infrastructure is not only vulnerable to ordinary software failures; it is vulnerable to targeted manipulation: data poisoning, evasion attacks, and integrity attacks on sensors or training pipelines. NIST’s work on adversarial ML taxonomy provides a shared language for these threats and mitigations, enabling more disciplined risk controls.

From these layers, accountability models are emerging that can be made concrete. The most defensible approach for grid operations is documented human authority with auditable AI support, implemented through:

1. Decision rights mapping: explicitly define which actions AI can recommend, which actions it can automate, and what requires human confirmation. The reliability community has stressed that operators should retain final input on AI/ML-generated actions.
2. Contestability and evidence: every material recommendation should be reproducible (or at least explainable) enough to stand up to regulator questions after the fact.
3. Audit logs and “model provenance”: maintain records of data sources, model versions, feature sets, and override decisions to support forensic analysis and compliance review.
4. Independent validation and red-teaming: incorporate adversarial testing and scenario-based stress tests aligned with adversarial ML risk models.
5. Operational monitoring and drift management: treat model drift as a reliability issue, not a data science inconvenience—especially under rapidly changing load shapes driven by AI compute.

The power shift is already visible. Consider what is happening simultaneously:

• Grid operators are being asked to revise tariffs and rules to accommodate large AI-driven loads and co-location structures.
• Regulators are changing rate design so that large users do not automatically socialize their costs to everyone else.
• National regulators are conditioning grid access on “bringing new supply,” effectively forcing data centers into quasi-utility roles. The Irish regulator’s policy requires new data centres connecting under the policy to provide new renewable and dispatchable electricity generation, directly or via contracted development.
• Tech firms are locking in long-duration, firm power via nuclear PPAs and SMR agreements, increasing their leverage over generation investment trajectories.

In plain terms: AI is not only consuming power; it is reorganizing power’s governance and capital structure.

Ethically, this raises questions that grid stakeholders cannot afford to treat as secondary:

• Distributional fairness: Are ratepayers subsidizing private compute buildouts through socialized transmission and generation costs?
• Environmental justice: Do “behind-the-meter” gas solutions move emissions burdens to specific communities, and do permitting regimes adequately protect those communities?
• Transparency and democratic legitimacy: If decisions are increasingly algorithm-influenced, can affected communities and regulators understand and contest them on equal footing?

Policy recommendations and best practices (operators, regulators, vendors)

For grid operators and utilities:
Adopt AI where it strengthens situational awareness and forecasting but implement it with reliability-grade discipline: operator override authority, training, audit trails, cyber segmentation, and formal incident processes. Align AI deployment with bulk-system cyber requirements and adversarial risk testing.

For regulators:
Treat large-load interconnection and co-location as system governance issues, not case-by-case exceptions. Standardize “who pays” principles; require transparent queue and upgrade cost rules; and demand AI accountability artifacts (documentation, monitoring, and post-incident disclosure) for AI used in critical infrastructure decision support—as the EU AI Act’s “high-risk” framing suggests.

For vendors and integrators:
Design for auditability and operator trust: explainability tooling, versioned model governance, secure-by-design data pipelines, and robust failure modes (graceful degradation to deterministic baselines). Use recognized AI risk frameworks and adversarial taxonomies to structure testing and threat modeling.

Conclusion

The AI boom is teaching a counterintuitive truth: electricity is not only a commodity—it is a socially governed system. At gigawatt scale, grid access is permissioned via process, and process is increasingly adversarial. That is why hyperscalers are moving toward “powered land,” co-location, behind-the-meter supply, and long-duration procurement: not because they love building power plants, but because the market window for AI is faster than the grid’s legal and institutional timelines.

At the same time, AI is entering the grid itself—quietly but materially—changing forecasts, alerts, and recommendations that shape operational outcomes. The reliability community’s warning is the right one: the grid is the planet’s most complex sociotechnical system, and AI must be deployed as accountable decision support with explicit human authority, not as a black box that dilutes responsibility.

The next decade will be defined less by whether we can build enough megawatts—and more by whether we can build enough legitimacy: fair cost allocation, transparent rules, and AI governance that can withstand courtroom-grade scrutiny.