AI for Energy and Utilities That Respects Grid Physics

The April 2025 Iberian blackout, where 2,200 MW of generation tripped in 20 seconds and left 60 million people without power, settled a debate that had been running for years. The ENTSO-E Expert Panel final report (March 2026) confirmed the root cause was cascading overvoltage from generators in fixed-power-factor mode, not renewable intermittency. The lesson for every utility deploying AI into grid operations is direct: an optimization model that ignores voltage stability, reactive power flows, and inverter control dynamics can recommend actions that look economically optimal in software and cause cascading failures in hardware. Grid AI is not a software problem. It is a physics problem with a software interface.

The NERC CIP Problem Nobody Planned For

NERC CIP standards were written for deterministic control systems. An AI inference engine sitting on the ICCP link between your control center and a balancing authority can trigger "medium impact" BES Cyber System classification under CIP-002-5.1, creating compliance obligations that neither the utility nor the AI vendor anticipated. CIP-003-9, effective April 1, 2026, expands governance for vendor electronic remote access. CIP-015-1 mandates internal network security monitoring for high and medium impact systems with external routable connectivity, with full compliance by October 2028. CIP-013-2 requires documented supply chain risk management plans, with penalties up to $1M per day for non-compliance. When we work with utilities integrating AI into operations, the compliance architecture comes first. Not bolted on after the pilot. We map every AI system component against CIP classification criteria, design network segmentation that satisfies CIP-005 while giving the AI the data access it needs, and build the compliance documentation artifacts that auditors actually review.

Load Forecasting in a World That Changed Overnight

U.S. data center grid-power demand will reach 75.8 GW in 2026 (S&P Global) and could hit 106 GW by 2035 (BloombergNEF). PJM attributes 12 GW of additional data center demand in 2026/27 alone, explaining a doubling in capacity costs. Goldman Sachs estimates $720B in grid upgrades needed through 2030. The problem for utility planners is that AI load does not follow historical patterns. Data center project timelines are shorter, load additions are larger, and demand profiles are flatter than traditional commercial loads. Average interconnection queue times have ballooned from under 2 years in 2008 to over 5 years now, with 2,000+ GW of resources waiting. FERC Order 2023 addresses the backlog, but PJM is only beginning to explore AI-enabled tools for streamlining interconnection studies. Most utility load forecasting models were trained on a world where demand grew 1% annually. That world is gone. We build forecasting systems that incorporate the non-traditional load shapes, rapid ramp profiles, and correlated demand patterns that AI infrastructure creates, because getting the forecast wrong by even 5% at this scale means billions in misallocated capital.

DER Integration: Protocol Fragmentation Meets Physics

FERC Order 2222 requires ISOs/RTOs to allow distributed energy resource aggregations into wholesale markets. Implementation timelines vary wildly: CAISO went live November 2024, NYISO in April 2024, but PJM targets February 2028 and SPP not until Q2 2030. On the ground, DERMS platforms must coordinate inverters, batteries, and controllable loads across IEEE 2030.5, OpenADR, SunSpec Modbus, and DNP3, often simultaneously. The protocol itself is still evolving: CSIP versions based on IEEE 2030.5-2018 and 2023, with CSIP 3.0 and CSIP AUS 1.2 introducing further changes. Meanwhile, California's energy storage capacity grew from 500 MW in 2018 to 16,900+ MW by mid-2025, with batteries delivering 12+ GW during the evening ramp on March 29, 2025. CAISO curtailed 738,000+ MWh of solar in Q1 2025 alone. At this scale, AI dispatch quality is grid-critical, not academic. We build the coordination layer that sits between DERMS, ADMS, and the physical DER fleet, handling protocol translation, real-time constraint enforcement, and dispatch optimization that respects both market rules and electrical limits.

Smart Metering Beyond Vendor Dashboards

Itron holds 35% and Landis+Gyr 32% of North American smart meter market share (GlobeNewsWire, July 2025). Between them, they cover the majority of the 100+ million AMI endpoints deployed across U.S. utilities. The analytics platforms they ship are designed for their own hardware ecosystems. Landis+Gyr's Revelo is the first IoT grid-sensing meter with edge computing, but the analytics beyond basic load profiling are limited. Itron's analytics platform handles demand response and billing, not the kind of multi-stream anomaly detection that catches revenue losses without generating false positives that trigger disconnections violating state PUC consumer protection rules. Global electricity theft costs $89.3B annually (IEEE Smart Grid), and the ML models deployed for detection suffer from extreme class imbalance: theft events are rare in training data, so false positive rates are high. We build AMI analytics that fuse meter data with network topology, weather, and consumption pattern analysis to improve detection accuracy while keeping false positive rates below the threshold that creates regulatory and customer-relations exposure.

Wildfire, Methane, Water: AI for Physical Risk

Hawaiian Electric committed $450M over three years (2025-2027) for wildfire safety including AI cameras and grid hardening. SDG&E trains wildfire models on a decade of historical records to predict failure and ignition probability at the pole and span level. Technosylva and AiDASH partnered to integrate wildfire modeling with satellite-based vegetation intelligence. But most utilities still rely on coarse weather-driven PSPS decisions that de-energize entire circuits rather than targeting specific spans, because their risk models lack the sensor fusion to be granular. For gas utilities, EPA OOOOb mandates AI-compatible leak detection monitoring with initial surveys required by 2025 and ongoing LDAR compliance tightening through 2026. Next-generation optical gas imaging systems from Bridger Photonics and Opgal use AI for automated frame-by-frame methane signature detection. For water utilities, AI leak detection has produced documented results: Walla Walla reduced non-revenue water from 40% to under 10%, and single-utility deployments show $213,000/year savings (Oldcastle Infrastructure). SVM-based leak detection achieves 94-100% accuracy across 53 peer-reviewed studies (MDPI). We build the sensor fusion and risk scoring systems that turn coarse safety models into targeted, defensible operational decisions, whether the asset is a transmission span, a gas main, or a water distribution line.

Agentic AI and the Control Room Question

Argonne National Laboratory built GridMind, a multi-agent AI system where specialized agents handle scheduling, monitoring, and dispatch independently but coordinate through a shared reasoning framework. This is the DOE's prototype for the "control room of the future." The agentic AI market in energy and utilities sits at $0.64B in 2025, forecast to reach $3.14B by 2030 (Mordor Intelligence). 42% of utilities plan targeted AI deployments in the next two years, but the overwhelming majority lack internal AI expertise (Utility Dive). Duke Energy's AI chatbot handled 280,000+ interactions in its first three months and reduced manual submissions by 90%, but UK utility customer satisfaction with AI-handled complaints dropped from 47% to 44% between July 2024 and January 2025 (Qualtrics). The gap between what agentic AI can do in a lab and what utility operators trust in a control room at 3am during a polar vortex is real, and closing it requires explainability, guardrails, and human-in-the-loop design that most AI vendors skip. We build agentic systems with reasoning traces that operators can audit, constraint boundaries that prevent autonomous actions from violating operating procedures, and graceful degradation that keeps the lights on when the AI is uncertain.

Why Not the Platform Vendor or the Big Consultancy?

GE Vernova, Schneider Electric, and Siemens each build excellent grid software. They ranked 1-2-3 in ABI Research's 2026 Grid Digitalization ranking. But each platform's AI capabilities are designed to keep you inside that vendor's ecosystem. A GE Vernova GridOS shop gets GE AI. A Schneider ADMS shop gets Schneider AI. If your utility runs mixed infrastructure, which most do, neither vendor has an incentive to build the cross-platform integration layer. Oracle and Hitachi Energy round out the top 5 in ISG's Power and Utilities Grid ranking, each with similar ecosystem lock-in dynamics. The large consultancies (Accenture acquired Orlade Group in September 2025 specifically for utility capital projects advisory; Deloitte, McKinsey, Capgemini all run active energy AI practices) bring program management and strategic framing, but the people writing the SOW rarely debug ICCP protocol issues or design edge deployments that survive substation EMI. We work vendor-neutral across GE, Schneider, Siemens, and legacy SCADA environments. We staff with engineers who understand NERC CIP compliance architecture, grid physics, and the specific OT constraints that make utility AI different from enterprise AI. Smaller team, faster delivery, no platform lock-in.