I spend a lot of time talking about what AI can do. How it can automate your invoicing, write your content, analyse your customer data, predict demand. All brilliant stuff. But there’s a conversation happening behind all of that magic, and it’s one that doesn’t get nearly enough attention.
Every time you ask ChatGPT a question, an AI model running on thousands of processors in a data centre somewhere draws power. A lot of power. And the way that data centre manages its energy — how it cools those processors, distributes electricity, and handles waste heat — is rapidly becoming one of the most important infrastructure challenges of the decade.
If you’re running a business that relies on cloud services, AI tools, or any kind of digital infrastructure, this matters to you. Even if you never see the inside of a server room.
How Much Energy Does AI Actually Use?
Let’s start with numbers, because the scale is genuinely staggering.
The International Energy Agency’s Electricity 2024 report projects that global data centre electricity consumption could reach over 1,000 TWh by 2026 — roughly the total electricity consumption of Japan. AI workloads are the primary driver of that growth.
A single query to a large language model like GPT-4 uses roughly 10 times more energy than a standard Google search, according to estimates from Goldman Sachs research. That might not sound like much for one query. But multiply it by hundreds of millions of queries per day, add in model training — which is even more energy-intensive — and you’re looking at electricity demand that’s growing faster than almost any other sector.
In Ireland specifically, the numbers are even more striking. Data centres consumed approximately 21% of all metered electricity in 2023, according to the CSO. That’s more than all rural households in the country combined. And the AI boom is only accelerating demand.
Why Energy Management Matters More for AI Data Centres
Here’s the thing that makes AI data centres different from your typical cloud hosting facility. The power density is dramatically higher.
A traditional data centre rack might draw 5-10 kW. An AI training rack with NVIDIA H100 or B200 GPUs can draw 40-70 kW. Some next-generation configurations are pushing past 100 kW per rack. That’s not an incremental increase. It’s a fundamentally different engineering problem.
| Workload Type | Typical Power Per Rack | Cooling Challenge |
|---|---|---|
| Standard web hosting | 3-5 kW | Manageable with air cooling |
| Cloud computing | 8-15 kW | Hot/cold aisle containment usually sufficient |
| AI inference | 20-40 kW | Requires advanced cooling, often liquid |
| AI training clusters | 40-100+ kW | Demands liquid or immersion cooling; air cooling is physically inadequate |
When your racks are pulling that kind of power, every inefficiency in your energy management system gets amplified. A 10% loss in power distribution that barely mattered at 5 kW per rack becomes a massive waste at 60 kW. Bad energy management doesn’t just cost money at this scale. It can literally prevent the facility from operating.
What Does Energy Management Actually Mean in a Data Centre?
It’s not just about the electricity bill, though that’s certainly part of it. Energy management in a data centre covers the entire chain from grid connection to server chip — and everything that happens to the heat on the way out.
- Power distribution — getting electricity from the grid to each rack efficiently, with minimal conversion losses. Every transformer, every UPS, every power distribution unit introduces some waste.
- Cooling systems — removing heat from the IT equipment. This is typically the single largest energy cost after the servers themselves. In a poorly managed facility, cooling can consume 40-50% of total power.
- Monitoring and optimisation — real-time tracking of power usage, temperature, airflow, and efficiency metrics like PUE (Power Usage Effectiveness). You can’t improve what you can’t see.
- Waste heat recovery — capturing the heat generated by servers and putting it to productive use, like heating nearby buildings. Most facilities still dump this into the atmosphere. That’s starting to change.
- Renewable integration — sourcing energy from wind, solar, or other renewables, either onsite or through power purchase agreements (PPAs).
A sophisticated BEMS for data centres — that’s a Building Energy Management System — ties all of these elements together into a single intelligent control layer. It monitors conditions in real time, adjusts cooling output based on actual load rather than worst-case assumptions, and identifies inefficiencies before they become expensive problems.
The Metric That Matters: Power Usage Effectiveness
PUE is the number everyone uses to measure data centre efficiency. It’s simple: total facility energy divided by IT equipment energy. A PUE of 1.0 would mean every watt goes to computing. A PUE of 2.0 means you’re using as much energy on overhead as you are on actual computing.
- Legacy facilities: PUE of 1.8-2.5. Common in older buildings with outdated cooling and power distribution.
- Average modern facility: PUE of 1.3-1.5. Decent, but there’s significant room for improvement.
- Best-in-class: PUE of 1.05-1.2. Google reports a fleet-wide average of 1.10. Meta’s facility in Clonee, Ireland achieves similar numbers.
The difference is enormous in real terms. On a 50MW AI data centre, reducing PUE from 1.5 to 1.2 saves roughly 131,000 MWh per year. At Irish commercial electricity rates, that’s millions of euro annually. And that’s before you count the carbon reduction.
AI Managing AI: How Machine Learning Optimises Energy Use
Here’s where it gets genuinely exciting — and yes, I appreciate the irony. AI is now being used to reduce the energy consumption of the data centres that run AI.
Google’s DeepMind team demonstrated this back in 2016, when they used machine learning to reduce cooling energy in Google’s data centres by 40%. The AI system learned the complex relationships between temperature, power, pump speeds, and outside weather conditions — relationships too complex for human operators to optimise manually.
Modern data centre energy management systems use similar approaches:
- Predictive cooling — AI anticipates heat load changes based on workload patterns and adjusts cooling proactively rather than reactively
- Airflow optimisation — machine learning models identify hot spots and adjust fan speeds, vent positions, and blanking panels dynamically
- Predictive maintenance — AI detects early signs of equipment degradation that would reduce efficiency, flagging components for replacement before they fail
- Workload scheduling — shifting non-urgent AI training jobs to times when renewable energy is abundant or electricity prices are lowest
This isn’t theoretical. These systems are running in production facilities today. And the energy savings are substantial — typically 15-30% reduction in cooling costs, sometimes more.
Ireland’s Unique Position
Ireland sits at an interesting crossroads in this conversation. The country has become one of Europe’s largest data centre markets, hosting major facilities for most of the big tech companies. That brings investment, jobs, and infrastructure.
It also brings strain. EirGrid has flagged concerns about data centre demand outpacing grid capacity. There’s been a moratorium on new data centre connections in parts of Dublin. And public debate about whether the country’s electricity infrastructure can support continued growth while meeting climate targets.
The honest answer? It can, but only if energy management improves dramatically. Building more data centres with PUE of 1.5 isn’t sustainable. Building them at 1.1 with waste heat recovery and renewable power integration? That changes the maths entirely.
Ireland’s climate actually helps here. Cool, mild temperatures mean free-air cooling works for much of the year. Some Irish facilities achieve mechanical cooling for fewer than 200 hours annually. That’s a natural advantage that good design can exploit.
What Businesses Should Take Away from This
You might be reading this thinking “I don’t run a data centre, why should I care?” Fair question. Here’s why.
If your business uses cloud services — and in 2026, almost every business does — you’re indirectly a data centre customer. The energy efficiency of your cloud provider affects your Scope 3 carbon emissions. Increasingly, that matters for ESG reporting, customer expectations, and regulatory compliance.
- Ask your cloud provider about their PUE. If they can’t tell you, that’s telling.
- Look for renewable energy commitments. Microsoft, Google, and Amazon all publish sustainability reports. Smaller providers should too.
- Consider data residency. Running your AI workloads in a well-managed Irish facility with low PUE and renewable power is better than hosting in a region with coal-heavy grids.
- Factor energy costs into AI adoption decisions. Running your own GPU servers on-premises has very different energy implications than using a cloud API. Understand the trade-offs.
The Bottom Line
AI is brilliant. I genuinely believe it’s going to improve how most businesses operate and how most people live. But it runs on electricity, and the amount of electricity it needs is growing fast.
Energy management in AI data centres isn’t a niche technical concern. It’s the infrastructure challenge that determines whether AI’s growth is sustainable or self-defeating. The good news is that the tools exist — intelligent energy management systems, liquid cooling, waste heat recovery, renewable integration, AI-optimised operations. The question is whether the industry deploys them fast enough.
For Irish businesses, the proximity of this issue makes it especially real. Those data centres humming away in west Dublin and Kildare aren’t abstract. They’re on the same grid as your home and your office. How well they manage their energy is, in a very direct sense, everyone’s business.