AI-Powered Cooling: How to Cut Peaks and Bills

PJM — the grid operator serving 65 million Americans — hit its highest peak demand since 2011 during the summer of 2025. That’s not a fun trivia fact for utilities; it’s a warning label. When heat and humidity spike together, air conditioning becomes less of a comfort choice and more of a public-health requirement. And the grid feels every thermostat adjustment.

Here’s the thing about peak load: it’s not driven by “average efficiency.” It’s driven by what equipment actually does in real buildings on the hottest, stickiest days of the year. That’s why RMI’s 2025 cooling work matters for anyone in energy and utilities who’s serious about grid optimization, demand forecasting, and reliability. The headline isn’t “we need more efficient AC.” The headline is: we need AC that performs under real-world humidity—and we need controls that can coordinate millions of units without making customers miserable.

This post is part of our AI in Energy & Utilities series, and I’m going to take a clear stance: super-efficient hardware is necessary, but AI-driven operation is what turns that efficiency into grid relief. 2025 gave us a case study showing how standards, field data, and smarter controls can line up.

2025 made the cooling-grid problem impossible to ignore

Extreme heat is now a planning assumption, not an edge case. In 2025, cities from Delhi to Osaka to Washington, D.C. saw searing heat and humidity. The RSS source highlights real consequences: record peaks, prolonged blackouts in places like Kolkata, and grid stress across parts of Southeast Asia.

For utilities, the operational challenge is straightforward:

• Cooling demand is highly coincident (everyone needs it at the same time)
• It’s weather-sensitive (load ramps fast)
• It’s humidity-sensitive (latent loads can drive energy use even when temperature is “handled”)

And there’s a trap that many planners still fall into: treating cooling like a simple temperature-control problem. The reality is that humidity management is often what forces systems to overcool, pushing energy use up precisely when the grid is most constrained.

“Overcooling to manage humidity drives excessive energy use — even in the most efficient units. Managing humidity and temperature together is key.”

— Bill McQuade, President, ASHRAE (as quoted in the source)

If you work in utility programs, grid operations, or DER orchestration, that quote should stick. It’s the heart of why “high-SEER” alone doesn’t guarantee peak reduction.

Real-world AC performance is the metric that matters (and standards are catching up)

If test standards don’t match real operation, markets optimize for the wrong target. One of the most consequential parts of the 2025 progress described in the source is the push to modernize how ACs are tested and rated.

Why load-based testing is a big deal

Traditional test procedures can reward performance at steady-state conditions that look nothing like actual homes and offices. Load-based methods aim to reflect how systems cycle, modulate, and respond to varying conditions.

In 2025, the Global Cooling Efficiency Accelerator (GCEA) contributed to momentum in three ways:

1. A draft proposal to an ISO working group for load-based testing and an updated performance metric (July 2025)
2. Engagement around ISO 21280, including (notably) a requirement related to sensible heat ratio so units are evaluated for both temperature and humidity behavior
3. Consensus-building with three international research groups to validate replicability and practicality, including approaches to factor humidity (latent load) into methods

Utilities should care because standards shape products. And products shape the achievable ceiling for:

• demand response performance
• customer comfort during load management events
• peak shaving reliability (how often a strategy works as designed)

A sentence you can take to your next program design meeting: “We can’t procure peak reduction with equipment that was never tested for humidity control at part load.”

The strongest 2025 lesson: humidity is where energy goes to die

Most cooling waste shows up when the system is fighting moisture. The source points to a nine-month field test in Palava City, India, featured in the report Bringing Super-Efficient ACs to Market. While the scraped RSS text doesn’t include every number in the chart, it does make the central point clear: real-world testing demonstrated meaningful differences between super-efficient approaches and typical/high-efficiency market options.

From an AI in Energy & Utilities perspective, that’s the bridge: field performance is the dataset we need to train better control strategies.

What “better control” actually means

Smarter cooling control isn’t about fancy apps. It’s about a system making good decisions under constraints:

• holding comfort bands (temperature + humidity)
• avoiding overcooling as a dehumidification hack
• shifting compressor/fan behavior to reduce kW at peak without violating comfort
• coordinating response across many devices while preserving feeder and transformer limits

GCEA also noted work on integrating smart control logic into existing products to manage both temperature and humidity. That’s exactly where AI becomes practical instead of theoretical.

Where AI fits: turning efficient AC into a grid asset

AI makes cooling flexible, not just efficient. Hardware improvements lower the baseline. AI and analytics create operational value on top of that baseline — especially during peaks.

1) Demand forecasting that treats cooling as a humidity problem

Most utility load forecasting models already ingest temperature. Fewer treat humidity (or dew point) as a first-class driver of demand. That’s a mistake in hot-humid regions and increasingly a mistake everywhere.

Practical approach I’ve found works:

• Forecast cooling demand using temperature + dew point (or wet-bulb), not temperature alone
• Segment customers by building type and HVAC type (single-speed vs inverter-driven)
• Use field telemetry (where available) to learn response curves by feeder

The payoff is tangible: better day-ahead peak prediction, fewer surprise constraint violations, and more confidence in dispatching demand response.

2) AI-driven load management (without customer backlash)

The grid needs kW reduction. Customers need comfort. AI is the negotiator.

Instead of crude cycling (which can cause humidity rebound and discomfort), AI-based strategies can:

• pre-condition intelligently when renewables are abundant
• limit compressor power while maintaining humidity targets
• stagger response to prevent “rebound peaks” when events end

The key is using comfort as a constraint, not an afterthought:

• temperature band (e.g., 23–26°C)
• humidity band (e.g., 45–60% RH)
• maximum ramp rate (avoid sudden changes that people feel)

When humidity is explicitly constrained, you reduce the “overcool now, pay later” pattern that drives both bills and complaints.

3) Predictive maintenance for cooling fleets

A poorly performing AC is a hidden peak-load multiplier. Dirty coils, refrigerant issues, sensor drift, and failing fans can all increase kW draw while reducing delivered comfort.

Utilities and ESCOs don’t need to become HVAC contractors, but they can support ecosystems that make predictive maintenance normal:

• anomaly detection on runtime vs weather
• power signature changes indicating mechanical problems
• humidity control failure detection (a big one)

For large commercial portfolios, maintenance analytics can be the cheapest “new capacity” you’ll ever buy.

4) Making renewables easier to integrate

Variable renewables are a scheduling problem. Cooling is a controllable load. Put them together and you get a cleaner grid.

AI can align cooling demand with renewable output by:

• shifting some cooling earlier (pre-cooling) when solar is high
• using thermal mass in buildings to ride through evening ramps
• targeting specific feeders where PV backfeed or evening constraints are worst

The grid benefit is not abstract: lower net peak, fewer curtailment events, and reduced reliance on peakers.

A practical playbook for utilities and large building operators

If you want cooling to become a reliability tool in 2026, start with these five moves. They’re designed to work even if you don’t control the hardware roadmap.

1. Update your cooling load models to incorporate humidity (dew point/wet-bulb) and building segmentation.
2. Prioritize real-world performance data when evaluating program savings; lab ratings are not enough for peak planning.
3. Design DR programs around comfort constraints, including humidity. Avoid strategies that create rebound peaks.
4. Pilot “smart humidity control” in a few representative climates and building types. Measure kW, kWh, and comfort outcomes.
5. Invest in ecosystem enablers the source calls out: testing infrastructure, digital modeling tools, and controls expertise. If manufacturers and integrators can’t test advanced logic reliably, adoption stalls.

A blunt observation: the utility that treats cooling as an orchestrated resource will out-plan the utility that treats it as an uncontrollable nuisance.

What to watch in 2026: standards, pilots, and market activation

The source closes with a 2030 vision: 5% of the global room AC market using super-efficient, climate-friendly technologies. The 2026 focus areas listed — manufacturing capability, scaled pilots, and market activation — are exactly the levers that determine whether super-efficient cooling stays niche or becomes normal.

From the AI in Energy & Utilities angle, 2026 is also where the winners separate themselves:

• Standards that reflect humidity performance will push better equipment into the market.
• Scaled pilots will generate the operational datasets needed for better forecasting and control.
• Market activation (finance, aggregation, awareness) will decide whether utilities can count on these resources in IRPs and reliability planning.

If you’re building an AI roadmap for grid optimization, don’t treat cooling as “just another load.” Treat it as one of the few loads that can deliver fast, customer-aligned flexibility at scale.

The forward-looking question I keep coming back to: When the next record-breaking heat wave hits, will your territory be protected by more generation—or by smarter buildings that simply don’t ask the grid for as much at the worst possible moment?