AI That Saves Power: A 2026 Field Guide

It's a conversation I've had more times than I can count this year. A customer calls, but it’s not about performance or features, it’s about electricity. "We're out of headroom," they say. Not CPU headroom. Power headroom.

Their AI initiatives are working. Customers love the new features, the pipeline is healthy, the board is excited. But the power budget isn't keeping up. Facilities is paging Finance. Finance is paging IT. And everyone is asking the same question: do we slow down the roadmap, or speed up the efficiency?

I find myself giving the same answer: you do both. And here's the twist: AI is the tool that makes it possible.

The paradox we’re living

There's an irony at the heart of the current moment. AI's energy appetite is real and growing. But the same technology that's straining our power infrastructure is also becoming one of the most effective tools we have to cut waste at scale. The question isn't whether AI consumes energy, it does. The question is whether we're smart enough to point that intelligence inward, toward the systems that sustain it.

I've started thinking about this as a walkabout problem. When I work with teams grappling with power constraints, we don't start with a spreadsheet. We start with a tour: data center, network, buildings, and operations. One circuit around the estate to see where the opportunities actually live.

Let me take you on that walk.

Stop 1: The data center, where heat meets math

The first thing you notice in a data center isn't the servers. It's the air. Cooling is the tax every compute cycle pays, and lately that tax feels steep.

For years, we've treated heating, ventilation, and air conditioning (HVAC) like a thermostat: set it and forget it, overcool just in case. But cooling systems don't behave like thermostats. They behave more like living systems. They can learn.

AI-assisted control changes the equation. Instead of fixed rules, the plant reads telemetry and predicts what the hall will need minute by minute. Coils and fans modulate just enough to keep temperatures stable. It's boring in the best way: no drama, fewer spikes, measurable kWh pulled off the bill.

But here's where it gets interesting. The second lever isn't about how you cool—it's about when you compute.

Google has been publicly documenting an approach they call carbon-aware computing. Using forecasts from services like Electricity Maps, they shift flexible workloads toward the hours and regions where the grid is cleanest: training runs, batch analytics, anything that can wait . Solar and wind peak at predictable times. Why not schedule your heavy compute to ride that wave?

The team I was working with recently started classifying their jobs by flexibility: minutes, hours, days. The ones that can wait get quietly pushed toward cleaner windows. They didn't change their models. They changed when and where they run, without missing the SLA. The room literally breathes easier.

What this looks like in practice

• Add AI supervisory control for chillers and computer room air handler (CRAH) / computer room air conditioning (CRAC) equipment where you have rich telemetry.
• Classify workloads by flexibility and begin carbon-aware scheduling for the flexible tiers.
• Surface carbon-intensity data to your SRE dashboards and capacity planning tools—make it visible, not buried.

The network, teaching capacity to nap

If the data center is a climate system, the network is circulatory. And it has the same flaw as any system built for peaks: most of the day, it runs hotter than it needs to.

I was looking at a timeline of radio and switch activity with a network architect recently. "Here's last night," he said. "East wing: five access points broadcasting full-tilt to nobody. South distribution: idle links lit like a runway."

The fix isn't heroic. It's attentive. AI watches occupancy and traffic patterns, then does two humanly impossible things at once: it gets conservative about power while staying fearless about user experience. Idle radios slip into sleep and wake before anyone notices. Flows get steered so cold links can truly go cold. Power over Ethernet budgets obey policy instead of habit. Badge readers and safety gear stay prioritized after hours; wall displays don't.

The mobile operators are proving this works at city scale. In London trials, Vodafone UK and Ericsson used AI/ML orchestration to dynamically manage RAN energy. The result? Up to 33% reduction in daily Radio Unit power at selected 5G sites, without degrading user experience. That's a real field deployment, not a lab result.

The symmetry is striking. Whether it's a 5G sector or a floor of access points, the principle is identical: capacity should nap when demand does.

This is the moment I usually suggest adding "energy as an SLO" to change management. Every rollout gets a user-experience gate and a power gate. Neither moves without the other. Pre-change baselines, post-change measurements, just like we do for latency or uptime.

What this looks like in practice

• Turn on energy-aware profiles in your network controllers: schedule radios and ports by occupancy and business hours.
• Monitor wake-latency and coverage KPIs: make sure sleep states don't create user-facing problems.
• For multi-site SD-WAN, explore traffic consolidation windows that allow deeper sleep on underused edges off-peak.

The building, where comfort is the currency

Facilities teams have a different kind of dashboard: lighting, ventilation, heating, and the subtle tides of people. For years, buildings chased a schedule on paper, e.g. 8 to 6, weekdays, assume everyone's there. But our companies stopped living on paper a long time ago. Some days hum along; others drift. Tuesday doesn't look like Thursday anymore.

AI reads the building's pulse—sensor hints, badge swipes, Wi-Fi presence, weather forecasts—and nudges comfort into the Goldilocks zone with less energy than a fixed schedule ever could.

Research from Lawrence Berkeley National Laboratory, published in a Nature-affiliated program, estimates 8-19% long-run reductions for commercial buildings from AI-enabled control alone, with larger impacts when combined with strong efficiency policy and low-carbon power. HVAC is consistently the top opportunity area in meta-analyses across the field.

But here's the thing I always emphasize: there's no magic number to paint on the wall. You're not promised "50% savings by Friday." What you get is steady, defensible reductions that add up over time and survive contact with reality. The building runs like good operations run: continuously commissioned, never set-and-forget, always measured against comfort as well as kWh.

When someone complains a meeting room feels stuffy, the system doesn't get defensive. It adjusts, and it learns.

The goal isn't to win the austerity Olympics. The goal is a building that feels right while spending less to feel that way. Measure outcomes the way operators measure reliability: energy per square meter, yes—but also comfort scores and trouble tickets.

What this looks like in practice

• Start with retrofit AI control on existing BMS/BAS before expensive plant changes.
• Use occupancy-aware setpoints and schedules (from badge data, Wi-Fi presence, or sensors) with clear comfort guardrails.
• Track "energy per m² vs. comfort score" as your north-star metric, not just energy alone.

Stop 4: The model, smaller because it’s smarter

Back upstairs, there's an unglamorous proposal waiting. "We don't need a bazooka for every nail."

A 2025 analysis from UNESCO and UCL shows that practical changes—using compact, task-specific models, applying quantization, and trimming prompt/response length—can materially reduce energy. In scenarios studied, savings approached the high two-digit range, with up to 90% cited as an upper bound. The key levers are straightforward: don't default to a giant LLM for routine tasks; prefer smaller models at the edge or in your data center; reduce unnecessary tokens.

It's the AI equivalent of swapping floodlights for LEDs and turning them off when you leave the room.

I don't sell this as philosophy. I sell it as practice. Teams keep the quality bar; they just measure power alongside latency and accuracy. The win is cumulative and cultural. Over months, the platform sheds watts the way an athlete sheds seconds, one honest improvement at a time.

What this looks like in practice

• Map your AI use cases; replace "one-size-fits-all" models with compact specialists where latency and quality allow.
• Enforce IO budgets (prompt and response length) in product teams and agents.
• Use low-precision inference with accuracy guardrails for production paths where it's proven.

The twin, useful when it closes the loop

The phrase "digital twin" has traveled a long way from hype to help. I've seen plant rooms with models on screens that look like video games: pumps, valves, heat exchangers, flows ghosting through pipes.

The twin earns its keep not by being pretty but by being connected: to sensors, to controls, to change logs. When you test a new setpoint strategy for shoulder seasons, you try it in the twin first, then in a single loop, then everywhere. The learning flows back, and the system responds.

Reviews across 2024-2025 conclude digital twins add value where sensor coverage and models are mature—continuous commissioning, scenario testing, predictive maintenance. Benefits vary by sector and maturity, but the direction is positive and growing.

Start where the load lives. Central plant first, then the rest. It’s also important to remember that the twin isn't a silver bullet; it's an amplifier for a team that already knows what good looks like.

What this looks like in practice

• Twin your highest-load systems first; link to live KPIs and close the loop to automation.
• Use twins for "what-if" studies before seasonal setpoint changes or major retrofits.
• Publish twin-derived changes with measured deltas (kWh, comfort, maintenance tickets) to build organizational trust.

The two external forces you can’t ignore

While you've been tuning your estate, the policy world has been tuning you.

In Europe, the EU Energy Efficiency Directive now requires annual reporting for data centers ≥500 kW. Regulators are exploring minimum performance standards and labeling schemes. This isn't punitive, rather it's a nudge toward ongoing efficiency improvements and data discipline.

In the United States, a July 2025 Executive Order directs agencies to accelerate permitting for data-center infrastructure, with closer scrutiny of grid impacts and stronger incentives for clean-energy alignment.

Transparency and speed are the new defaults. That's not a burden if you build with them in mind. Publish your carbon-aware scheduling policy internally. Ship energy telemetry with new services. Treat compliance artifacts as product features, not paperwork.

The conversation with finance

By the week's end, you find yourself in the meeting that truly decides whether any of this happens: the one with finance.

I've learned to bring two pictures. The first is a growth curve marching straight into a power wall. The second is the same curve bent gently downward by a dozen small levers: AI-assisted cooling and scheduling; networks that sleep when people do; buildings that serve comfort with less; models that fit the job; twins that prove changes before they cost you.

The investment is modest because most of the wins are operational, not capital. The results are credible because they're measured like SLOs, not slogans. And the story is coherent because it's not a crusade against AI: it's AI cleaning up after AI.

Finance doesn't need poetry. Finance needs risk retired and capacity unlocked. Show them both.

What changes on Monday

You don't need to launch a moonshot. You just start by turning a few dials:

Your schedulers start classifying jobs by flexibility and follow the cleanest hours and regions they can. Your controllers learn to treat energy the way they already treat performance—targets, alerts, post-change reviews. Your networks practice good sleep hygiene and wake without drama. Your buildings stop pretending Tuesdays look like Thursdays. Your models get smaller where they can, and precise where they must. Your twins earn their keep in the one place that matters: closing the loop to automation.

As small wins accumulate, no single headline claims the victory. The system does.

What this means for Extreme Networks customers

Energy efficiency is becoming a first-class objective in network and platform strategy. For enterprise Wi-Fi and switching, policy-driven sleep states, idle-link consolidation, and PoE budgeting can deliver measurable reductions with minimal risk—when paired with occupancy awareness and fast-wake guardrails.

For industrial and private wireless deployments, a 2025 Nokia/GlobalData study finds enterprises adopting private wireless with on-prem edge are reporting quick ROI alongside energy and emission reductions as part of their automation programs. Treat energy and safety outcomes as co-equal targets.

For cloud and application workloads, carbon-aware placement and scheduling are now table stakes. Adopt the hyperscaler pattern internally for batch analytics and model training.

The quiet payoff

Six months from now, the graphs tell a different story. There's still growth—features land, customers scale—but the energy line is flatter than the forecast you showed the board last year. The team spends less time firefighting and more time fine-tuning. The regulatory packets that once felt like a scramble now ship with the product.

And your roadmap? The one that kept stalling on power? It starts to feel roomy again.

The paradox remains: AI can be hungry. But it can also be disciplined. The difference is whether you point its intelligence inward as well as outward—whether you let it optimize the systems that sustain it.

This isn't a call to austerity. It's a call to craft. In 2026, the winners won't be the loudest models or the flashiest builds. They'll be the teams that teach their infrastructure to think—about comfort as well as kilowatts, about timing as well as throughput, about right-sizing as well as right-now.

And they'll look back on that Tuesday afternoon, the one where the power budget blinked red, and remember it as the day the system started to learn.