5 AI Principles for Data Center Efficiency

AI is driving new demand for energy in data centers. But it would be short-sighted to think of AI solely in terms of new demand. It can also be part of the sustainability solution for data center operators.

At Equinix, we’re using adaptive AI controls to optimize operational efficiency in our data centers. We started with a targeted deployment at several of our facilities, and we’re evaluating a broader rollout to come. We’ve achieved promising results so far:

• In our Equinix FR6 data center in Frankfurt, we reduced energy consumption by 900 MWh per year.
• We lowered power usage effectiveness (PUE) at FR6 to around 1.2, well below the industry average of 1.54.[1]
• In Singapore, we achieved annual PUE improvements of 10% at our Equinix SG4 site.

While these results are encouraging, we’re equally excited about sharing the key learnings from this experience. Read on to learn five take-aways:

1. Starting small can make the biggest impact
2. Guardrails are essential
3. Dashboards are dead
4. Agentic is the next frontier
5. It all starts with data

These five principles can be used to guide future AI projects, especially when they involve sustainability.

1.    Starting small can make the biggest impact

There’s practically no limit to what we can do with AI, which is part of what makes it so challenging to get started. For instance, we could have taken a scattershot approach to optimizing every single aspect of our data center operations, but then we’d have no way of knowing what changes were really moving the needle.

That’s why we intentionally kept our AI inputs as simple as possible. A simple approach also reduces costs and complexity, making it quicker and easier to get started. We sought to strike the right balance between what we want to invest and what we hope to achieve. We decided that incrementally optimizing the temperature and humidity of our airflow was the best way to do that.

We implemented a machine learning model to guide our physical air-conditioning systems. The model makes small tweaks to our airflow and then observes the results. It works on the principles of reinforcement learning: If a particular tweak leads to lower energy consumption, then the model knows to make similar tweaks in the future.

Starting with a small, targeted AI initiative has helped us achieve the best return on our investment. We’re already demonstrating how doing the little things right can add up to a big impact over time.

2.    Guardrails are essential

Using a machine learning model is a lot like training a dog. There needs to be positive reinforcement for desirable behaviors, but also negative reinforcement when the model steps out of line.

When the model makes a recommendation, it must pass through an AI gateway:

• The gateway is responsible for translating recommendations into a format that local systems can understand.
• It’s also responsible for applying guardrails. It detects if the recommendations would take our data centers outside of safe, reliable operational parameters. If so, it does not pass those recommendations on to local systems.
• If a recommendation is rejected, the AI gateway provides feedback to the model, so that the model knows not to make similar recommendations in the future.

AI guardrails help us balance the tension between AI innovation and AI safety. This ensures that any optimizations we make will not have unexpected consequences for our customers.

3.    Dashboards are dead

One question that inevitably arises is how AI-driven optimization will impact human data center technicians. Adaptive AI controls are intended to optimize efficiency, not replace jobs. In fact, the data center industry is already facing a massive talent shortage, so our goal is to support and empower the workers we do have.

Adaptive AI controls are impactful because they’re completely automated. In the past, data center optimization revolved around dashboards: Human technicians would have easy access to all the data they need, but they’d still be responsible for interpreting that data and reacting accordingly. Like the “check engine” light on a car, dashboards can identify when there is a problem, but they can’t fix the problem for you.

Now, it’s become common knowledge that dashboards are dead. Workers don’t have time to look at them, so it doesn’t matter how good the insights are that they provide. What workers need are solutions that give them time back. Our adaptive AI controls are one example of this, because optimization happens without the need for direct human intervention. This frees up our people for more strategic work.

4.    Agentic is the next frontier

We’re already beginning to see what AI agents can achieve in the world of software. But optimizing physical systems with AI agents is an entirely different challenge than optimizing virtual systems.

There’s a lot of talk in the industry about what this might look like, but very few companies are actually doing it today. We hope that our initial forays into machine learning for data center optimization will pave the way to using agentic AI for the same purpose.

We particularly hope to implement agent-to-agent communications with the vendors and suppliers that help our data centers run. Many of these partners have expressed interest in pursuing this with us, but there are challenges we must overcome first.

Just like enterprises once went from siloed applications to connected API gateways, we’ll need to establish an ecosystem of agents from different organizations that are trusted to connect and collaborate with one another. These agent-to-agent communications will drive exponential value and create further time-savings benefits for human technicians.

5.    It all starts with data

Like any other machine learning use case, adaptive controls for data center operations rely on access to high-quality datasets. The principle of “garbage in, garbage out” applies here. Inaccurate or obsolete datasets make it impossible for the model to formulate helpful recommendations. Building the right data pipeline is the first step an organization should take before implementing an AI initiative like ours.

All of the recommendations that our machine learning model makes are based on real-time data that’s collected from sensors throughout our facilities and then shared with the model via APIs. With this kind of visibility into our operational data, the model can predict thermal spikes and adjust proactively.

What this means for technical leaders

We’re proud of the results our AI adaptive controls have achieved so far, but we’re equally proud of what it represents: an example of what’s possible when we apply the power of AI for good.

There are several implications that stand out for technical leaders who are managing the trade-offs between AI and sustainability:

• You don’t have to choose: Increasing demand for energy is a valid concern, but sustainability versus AI is a false choice. As we’ve demonstrated within our own data centers, AI can be part of the sustainability solution by helping businesses operate IT infrastructure more efficiently.
• The benefits can scale: With adaptive AI controls, there’s no built-in limit to what we can achieve. As reinforcement learning continues, our operations will grow progressively more efficient. We also plan to expand our impact by adding AI adaptive controls in more data centers.
• Data + strategic thinking = results: At a time when businesses are using AI to tackle any and every issue they might face, we‘ve demonstrated why a simple, focused approach is best. If you start with a clear vision of what you want to achieve and then apply the right data, you can get the intended outcome while avoiding waste.

Equinix puts efficiency innovations to work for our customers

Improving data center efficiency requires both consistent investments and innovative thinking. At Equinix, we know this as well as anyone, and it shows:

• PUE improvements: In 2025, our global annualized average PUE was 1.37. This represents an efficiency improvement of more than 5% compared with 2024, and an improvement of about 31% from our 2019 baseline.
• Focused investments: Our recent efficiency improvements came after investing $36 million to replace infrastructure across 79 sites with more efficient technology.
• A multifaceted approach: In addition to AI adaptive controls and airflow management, we’ve implemented advanced liquid cooling to remove heat more efficiently and support high-density AI workloads. We also help customers avoid waste by sizing infrastructure appropriately for their needs.

In 2025, we made progress toward our sustainability goals, from maintaining high levels of renewable energy coverage to using water more responsibly. Read our 2025 sustainability report for a closer look.

[1] Uptime Institute Global Data Center Survey 2025, Uptime Intelligence, July 2025.