Energy & Sustainability in the Age of Agency

Executive summary

• Demand and control are now linked. The IEA estimates that data centre electricity use rose 17% in 2025. Use at AI-focused facilities rose about 50%. Total data centre consumption is projected to roughly double to 950 TWh by 2030 ^[2][3]. Meanwhile grid operators are already running AI agents in live control-room work ^[4].
• Value is shifting to whoever controls flexible assets. Home batteries, virtual power plants (VPPs) and flexible data centres are becoming grid resources. More and more, software agents will decide when to charge, export or curtail. That changes who earns the margin ^[3][6].
• Regulators are catching up quickly, and their regimes differ. The EU AI Act treats some grid AI as high-risk ^[16]. US regulators are writing mandatory standards for data centre loads ^[14]. Five Eyes cyber agencies have issued joint guidance on AI in operational technology (OT) ^[15]. Boards need governance that works in every market they operate in.

What's changing

1. AI agents are working in grid control rooms. CAISO acts as Reliability Coordinator for much of the western US. That region handles more than 500 new generation outage requests a day. CAISO and its technology partner OATI have put a multi-agent system into production that reviews next-day outage requests in minutes. Separate agents handle reporting, cited procedural advice and monitoring of critical outages ^[4]. EPRI's control-room research takes a similar approach: AI supports sensing, deciding and acting, with a human kept in the loop ^[4]. In the US's largest market, PJM is working with Google and the Alphabet-backed Tapestry to automate parts of its connection queue, where projects wait about five years on average ^[5].

2. Distributed energy is becoming a fleet that software runs. Australia's 2026 Integrated System Plan (ISP) found that rooftop solar supplied 16% of National Electricity Market (NEM) supply in Q1 2026. That was more than grid-scale solar, wind or hydro. The plan expects about two-thirds of solar homes to have batteries by 2050, and more than half of those to join a VPP ^[6]. Australia's market rule-maker, the AEMC, sees AI coordinating household batteries, EV charging and solar exports through distribution system operators. It has also flagged data security and algorithmic collusion as risks ^[7]. In buildings, Trane launched AI controls in 2025 that adjust HVAC automatically using occupancy, weather and usage data. They come with a conversational agent for facility teams ^[8].

3. AI's own load is reshaping capital plans in Asia-Pacific. Malaysia's national utility, TNB, has supply agreements with 59 data centre projects totalling 8.3 GW ^[10]. Moody's notes that the government expects data centres to reach 31% of Peninsular Malaysia's electricity consumption by 2035, up from 4% in 2025. It estimates MYR 80–95 billion of power investment will be needed over the next decade ^[9]. Singapore takes a different line. It releases new capacity only to operators that prioritise sustainability, and it targets a power usage effectiveness (PUE) of 1.3 or better across all data centres within ten years ^[11].

4. Data centres are being asked to become grid assets. The IEA expects 20–25 GW of battery storage at data centres by 2030. It suggests faster connections in exchange for non-firm access and demand response ^[3]. Ireland now requires new large data centres to bring dispatchable on-site or nearby generation or storage. They must also match 80% of annual demand with additional renewables ^[12]. The EU's June 2026 roadmap adds a data centre energy rating, with first labels expected in 2027, and a "use-it-or-lose-it" principle for grid connections ^[13].

Through the Age of Agency lens

Age of Agency: delegation becomes an operating decision. The CAISO case shows the shift from AI that answers questions to AI that triages, monitors and recommends inside a regulated workflow ^[4]. The board question changes from whether to use AI to which decisions are delegated, and within what limits. For utilities, this means writing down a "delegation map":

• tasks agents may carry out alone, such as reports and screening
• tasks they may recommend, such as outage approvals
• tasks that need a human, such as switching or load shedding.

Under EU rules, AI that supports decisions on load distribution or shutdowns may be classed as high-risk ^[16]. That legal line should shape the map.

Agentrification: margin moves to whoever orchestrates. As agents take over coordination work, value moves away from owning assets and towards controlling them:

• Gainers are aggregators and VPP operators that manage household fleets ^[6]. So are building owners whose controls can trade flexibility. So are data centre operators that can show they are flexible and connect sooner as a result ^[3][13].
• Losers are likely to be retailers whose only edge is billing and customer contact. Another group is planners whose work is mainly manual queue review, which PJM is automating ^[5].

The workforce question is redeployment, not cutting headcount. Engineers who used to process requests now need to supervise agents, check their outputs and handle exceptions.

UBInt: compute runs on energy. If capable AI becomes essential infrastructure, then grid capacity is the binding constraint on access to it. The IEA warns that equipment and grid-connection bottlenecks are tightening ^[2]. Malaysia's build-out shows how fast one country's power system can be reshaped by demand for AI capacity ^[9]. Asia-Pacific governments face a choice that Singapore and Ireland have already made explicit. Do they allocate scarce power to AI, and on what conditions? The likely conditions are efficiency, flexibility and additional clean supply ^[11][12]. Energy policy is turning into AI policy, and boards in both sectors should plan accordingly.

Risks and governance

• Autonomous control of critical systems. Cyber agencies from the US, Australia, New Zealand, the UK, Canada, Germany and the Netherlands have issued joint principles for AI in OT ^[15]. The four principles are: understand the AI; consider carefully whether to use AI in the OT domain at all; establish governance and assurance; and embed safety and security in AI-enabled OT systems. Utilities and building operators should treat these as their baseline.
• Cyber exposure. The IEA reports that cyberattacks on energy utilities tripled over four years, and that AI is making attacks more sophisticated ^[1]. The EU will assess systemic risk in solar and wind equipment, restrict EU funding for inverters from high-risk suppliers, and may ban such inverters altogether ^[13].
• Large-load instability. In July 2024 a transmission fault in Virginia caused about 1,500 MW of data centre load to drop off the grid. ERCOT in Texas has recorded more than 25 such events. In July 2026, the US regulator FERC directed NERC to file mandatory reliability standards for computational loads by 31 December 2026 ^[14].
• Regulatory classification. Under the EU AI Act, AI used as a safety component in electricity infrastructure is high-risk. That brings duties on risk management, documentation and conformity assessment. Final Commission guidelines are expected by the end of 2026 ^[16]. The Digital Omnibus on AI entered into force on 27 July 2026, so firms should confirm how it changes their obligations ^[16].
• Market conduct. The AEMC has flagged data security and algorithmic collusion as risks, which is directly relevant when pricing agents act on behalf of many customers ^[7].

Questions for boards and leadership

1. Which operational decisions have we delegated to AI agents, or plan to? Who signs off the limits on each?
2. Could any of our AI systems count as a high-risk "safety component" under the EU AI Act? Does our assurance match the strictest regime we operate under?
3. Do we know how our data centre loads, or our customers' loads, behave during a grid disturbance? Are we ready for NERC's coming standards?
4. In the fight to control flexible assets, are we the orchestrator or the asset being orchestrated? What is our plan to earn that margin?
5. Have we tested our AI-enabled OT against the joint international principles, including the case where the agent itself is compromised?
6. Where we site or buy compute in Asia-Pacific, how exposed are we to grid limits, conditional power allocation and additional-renewables rules?
7. What skills will our operations staff need to supervise agents rather than do the tasks themselves? Are we building them now?

How KIG helps

• KIG Advisory. We help boards and executives build delegation maps for agentic AI in grid, building and portfolio operations. We align these with regulation across Asia-Pacific, the US and Europe, and set governance that meets the strictest market's standards.
• KIG Research. We apply the Age of Agency, Agentrification and UBInt frameworks to energy and sustainability. The aim is to help leaders see where value and risk are moving as agents take on coordination work.
• KIG Ventures. We back AI companies in smart building automation and sustainable technology. That gives us a working view of how autonomous controls are built, deployed and governed in practice, and we bring those lessons into our advisory work.