Phaidra Blog

Phaidra, CoreWeave and Applied Digital pioneer NVIDIA Max-Q AI factories with agentic liquid cooling

Jim Gao — Mon, 16 Mar 2026 16:00:00 GMT

March 16, 2026 – Phaidra today announced a groundbreaking methodology to drastically improve the thermal stability of liquid-cooled AI data centers. This methodology is outlined in the joint white paper “AI Agents for Liquid-Cooled AI Factories.”

By successfully leveraging AI-driven, feed-forward control systems on production NVIDIA Grace Blackwell platforms, the collaboration is paving the way for the future of "DSX AI factories" — a new operational paradigm where power, cooling, and workload management are unified to maximize efficiency and computational throughput. Phaidra has integrated NVIDIA DSX Max-Q to run GPU clusters as efficiently as possible, so more of the available power can go towards running AI workloads.

The challenge of AI thermal volatility

Modern AI factories are fundamentally different from traditional data centers: defined by massive scale, extreme density, and highly synchronized workloads. Operators of large-scale AI factory campuses, such as Applied Digital, must manage increasingly complex interactions between power infrastructure, liquid cooling systems and rapidly fluctuating GPU workloads for their partners. When massive AI training or inference jobs are dispatched, thousands of networked GPUs ramp up simultaneously, creating "peaky" power profiles that can jump from idle to maximum capacity within seconds.

Traditional liquid cooling relies on Proportional-Integral-Derivative (PID) controllers, which wait for a sensor to register a coolant temperature change before taking action. Because coolant has high thermal inertia, this reactive feedback loop suffers from a 3-to-5-minute delay, resulting in rapid heat spikes that force GPUs to throttle performance to protect themselves. To mitigate this, operators significantly over-cool their facilities to create a safety buffer—a strategy that wastes massive amounts of energy and limits overall compute capacity.

The AI-driven solution

To close this latency gap, Phaidra developed a self-learning reinforcement learning (RL) AI Agent that fundamentally changes how cooling is managed. Instead of reacting after-the-fact to temperature changes, the AI Agent uses real-time rack power data as a leading indicator to predict and prevent thermal spikes. The agent seamlessly sends optimal setpoint commands to the Coolant Distribution Unit (CDU) before the heat fully registers in the fluid, reducing the effective response delay from minutes to under 10 seconds in validated production environments.

Proven results at gigawatt scale

The new methodology underwent rigorous joint A/B testing in live production environments, including an NVIDIA DGX SuperPOD cluster running LLM training workloads and CoreWeave's NVIDIA GB200 NVL72 environments. The results were transformative:

Massive reduction in thermal overshoot: The AI Agent successfully reduced the magnitude of thermal spike overshoots by 75% to 80% compared to optimally-tuned PID baselines during sudden load ramps.
Unprecedented scale: Following this successful validation, Phaidra and CoreWeave are scaling the deployment of these AI agents throughout CoreWeave’s liquid-cooled fleet, bringing AI-driven thermal management to its next generation of data center capacity.

The pathway to Max-Q AI factories

Phaidra has integrated NVIDIA DSX Max-Q to operate the entire AI factory as a single unit of compute, at scale. By deep integration of Information Technology (IT) and Operational Technology (OT), this collaboration bridges the divide between white space compute and facility operations.

With thermal stability secured by Phaidra's AI agents, facilities can safely raise their supply water temperatures, significantly reducing the burden on facility chillers. This provides the foundation for the next phase of the collaboration, where operators dynamically shift stranded power from the cooling system to revenue-generating IT compute. For a baseline 1GW AI factory, raising the coolant temperature safely could unlock billions in additional annual revenue.

"In a world where computational resources are limited by energy availability, every watt that isn’t being used for valuable token generation is a wasted watt," the joint white paper states.

By co-designing power, cooling, and workload management systems, Phaidra, CoreWeave, NVIDIA, and critical infrastructure partner Applied Digital, are setting a new standard for reliability, end-user SLAs, and peak operational efficiency in the age of AI.

Phaidra launches Phaidra Prism, an AI platform to operate AI factories at scale

Chris Vause — Tue, 03 Mar 2026 17:00:00 GMT

At this year’s NVIDIA GTC in San Jose, Phaidra launches its AI platform Phaidra Prism designed to accelerate data center operations and meet the demand of AI workloads. Gigawatt-scale factories require new methods for everything from design to operations, and Phaidra Prism delivers detailed visibility and advanced collaborative intelligence to AI factory operators so that they can maximize tokens per watt while limiting operational expenses.

Phaidra Prism is an AI assistant designed by data center experts for data center operators and technicians and is already integrated with NVIDIA DSX Exchange, which helps AI factories maximize output and token performance via a common data and control layer — unifying Information Technology (IT) and Operational Technology (OT) within AI Factories. Prism users interact with equipment and their data through a customized LLM to understand operating conditions, conduct detailed analyses, and troubleshoot issues before they impact SLAs.

“Traditional methods of operating data centers no longer suffice,” said Phaidra CEO and Co-Founder Jim Gao. “Modern data centers are substantially larger and more complex, coinciding with a significant increase in the cost of downtime. That’s why Phaidra is building new technologies that put data center operations at the forefront. All this mission-critical infrastructure needs new tools to ensure safe and reliable operations, at scale.”

Through Phaidra Prism — the first AI assistant designed specifically for data centers — technicians use the AI’s advanced reasoning capabilities to uncover operational issues, prioritize fixes and ultimately troubleshoot to resolution.

With the tool’s conversational interface, users engage directly with their AI factory infrastructure instead of digging through fragmented documentation or manually combing through BMS data.

As a result, operations teams with Phaidra Prism achieve:

Greater reliability - identify and resolve issues before they cause problems.
Higher uptime - reduce the risk of catastrophic outages.
Increased productivity - troubleshoot complex issues in a fraction of the time with AI assistance.

Phaidra will be demonstrating the power of Phaidra Prism at GTC conference booth #95.

About Phaidra

Phaidra is focusing on enabling gigawatt-scale data center operations to meet the demand of AI workloads. Phaidra combines deep data center knowledge with leading AI expertise and extensive industrial controls specialization. Founded in 2019 by former Google, DeepMind and Trane engineers — Phaidra is ushering in a new age of data center operations with AI agents for AI factories.

Orchestrating AI factory cooling at gigawatt scale

Daniel Fuenffinger — Tue, 06 Jan 2026 17:00:00 GMT

Log Entry: We're not building a data center

I've commissioned and operated a lot of facilities—submarines, data centers, industrial plants. You learn pretty quickly that the engineering challenge is about more than capacity. It's about understanding what kind of machine you're actually building.

This time, we weren't building a data center. We were building a race car.

That's what an AI factory is—a high-performance machine where every system has to coordinate in real time. If one component lags, the whole system skids. And at gigawatt scale, any slide is expensive.

The challenge was building infrastructure that operates as a single unit of compute. It must handle synchronized AI workloads and extreme thermal demands. Most importantly, there’s some room for over-sizing to compensate but every redundant piece of equipment cut into the margin needed to make the facility profitable.

In the following log entries, I’ll cover:

What makes AI factories fundamentally different from traditional data centers
Why thermal control became the bottleneck we didn't see coming
How Phaidra's AI agents changed the way we think about infrastructure management
What it takes to coordinate systems at the speed AI workloads demand

Log Entry: One big computer

Rule #1 of AI factories: This is NOT just a larger data center.

In a traditional data center, loads are distributed. Different tenants, different workloads, different times. You might see a spike on Black Friday or during a product launch, but mostly things ramp gradually and somewhat predictably. This predictability creates flexibility. You can even over-subscribe power (install more IT equipment than the power infrastructure is rated for) because not everything's going to be on at once.

AI factories don't work that way.

When an AI training job kicks off, every rack goes from 20% to 100% almost instantly. Thousands of GPUs light up in sync. And when the job completes, everything drops back down together.

It's like Black Friday every day, except it happens multiple times, with little to no warning, across the entire facility.

This is not a bigger data center. Traditional data centers are digital commercial real estate where compute space is rented. There are multiple tenants or just one, but the compute infrastructure is assigned to do different things at different times, in different ways. Like a flexible commercial office building.

An AI factory is one big computer. Every component must synchronize toward a single purpose. And that changes everything about how you design, operate, and control it.

I'd worked on Google's early AI infrastructure—systems with TPUs and CDUs that behaved a lot like what we were building here. Dense, synchronized, unforgiving. But even those systems didn't operate at this scale or with this level of coordination demand.

The challenge wasn't theoretical anymore. It was right in front of us.

Log Entry: The systems weren't keeping up

Early load testing revealed the problem fast.

We'd bring a training workload online, watch power spike across the racks, and then—lag. The cooling system would react, but not fast enough. By the time chillers ramped up and water started circulating at the right temperature, we'd already seen thermal overshoot.

Not catastrophic. But enough to throttle performance.

Enough to make you nervous.

Traditional control systems wait for temperatures to rise before they respond. That's fine when loads change slowly—comfort cooling, district energy, steady industrial processes. Those systems were engineered for gradual ramps, not synchronized spikes.

But in an AI factory, gradual doesn't exist.

You either keep up, or you fall behind.

The issue wasn't capacity. We had enough chillers, enough pumps, enough cooling infrastructure to handle the load—if we had time to respond. But time was the constraint we couldn't engineer around with more equipment.

The limit wasn't on total cooling available. The limit was on the rate of change.

And that's when it became clear: we needed a different kind of control.

Log Entry: Leading indicators, not lagging ones

There's an analogy I kept coming back to during those early weeks.

When humans learned to hunt, we didn't become the dominant species because we were faster or stronger. We became dominant because we learned to throw rocks then spears where the animal was going to be, not where it was.

That's the difference between lagging indicators and leading indicators.

One reacts. The other predicts.

Traditional control systems are reactive. They wait for a temperature sensor to report that something's getting hot, then they respond. Therefore, they follow lagging indicators.

By that point, the thermal spike is already happening.

What we needed was feed forward control to see the spike coming—workload changes, IT load signals, power ramping, system response times—and act before temperatures ever started to rise.

That's what Phaidra's AI agents do.

They don't wait for the problem. They anticipate it, adjust proactively, and create the needed cooling procedure to get ahead of things, so it won’t need to catch up - it’s already there. It's like traction control in a Formula 1 car… adjusting power delivery before the tires lose grip, not after.

We deployed liquid cooling agents on the CDUs across a 100MW GPU cluster and watched it go to work.

Log Entry: The system started tuning itself

Within hours, thermal behavior tightened.

Thermal overshoot that used to spike 3-5°C dropped to less than 0.5°C. Thermal spikes that would trigger equipment throttling or manual interventions—gone. The system wasn't just responding anymore. It was predicting, adjusting, and learning.

And the best part was that it kept getting better!

Phaidra's agents were trained in simulation first—high-fidelity digital twins of NVIDIA DGX SuperPODs that let the team prototype and test different architectures before deployment. Once the agent hit our live system, it continued learning through reinforcement learning. No manual tuning. No human intervention.

It observed, adapted, and improved.

In one test, we threw a 30-70% load ramp at it—the kind of spike that would've caused major thermal overshoot with traditional controls. The AI agent absorbed it like it was nothing. Smooth, stable, precise.

It reminded me of working on a nuclear submarine, where you had layers of redundancy and safety systems engineered into everything. Except this time, the safety system was learning in real time and getting smarter with every cycle.

The results were clear:

80% reduction in thermal spikes compared to the baseline control system
<0.5°C precision in maintaining target temperatures
Higher TCS temperature setpoint without sacrificing stability
Less cooling infrastructure required overall—no need to over-build just to buffer spikes

We didn't tune the system. The system tuned itself.

And that fundamentally changed how we thought about infrastructure management.

Log Entry: Coordination at gigawatt scale

Here's what I've learned after years of working in mission-critical facilities: systems only perform as well as their weakest coordination point.

You can have the best chillers, the cleanest power delivery, the most advanced GPUs—but if those systems don't talk to each other in real time, you're leaving performance on the table.

AI factories are no different. They're just faster, denser, and less forgiving of lag.

That's why the NVIDIA Omniverse DSX ecosystem exists. It's not just about hardware specs or design plans. It's about coordination—every partner bringing precision to their piece of the stack so the whole system works as one.

Partners like Jacobs design the mechanical and electrical core. Siemens provides digital twin capabilities and automation systems. Schneider Electric handles power distribution. GE Vernova interfaces with the grid. Emerald AI runs predictive simulations. And Phaidra brings the real-time control layer that keeps thermal systems stable under dynamic load conditions.

It's like a Formula 1 team.

Every component matters. Every system has to be synchronized. And when it works, the machine doesn't just run—it out-performs.

Phaidra's role became clear to me: the traction control that keeps the system on-line. Phaidra manages thermal precision so the rest of the infrastructure can do its job without over-building, without waste, and without slipping.

Because at gigawatt scale, small inefficiencies magnify and compound fast.

A single gigawatt AI factory represents a $50 billion investment and a $200 billion revenue opportunity. Every 1% of inefficiency costs ~$2 billion in lost performance.

You can't afford to guess. You can't afford to react too late. And you can't afford systems that don't talk to each other.

Log Entry: A new baseline

We're building smarter facilities—not just bigger ones.

AI factories require infrastructure that thinks ahead—systems that predict, adapt, and coordinate at speeds that exceed human reaction time. That's not optional. It's foundational.

Before Phaidra's agents came online, we were reacting to problems. Now, we're preventing them. The system anticipates thermal spikes before they happen, adjusts proactively, and maintains precision without manual intervention.

That's the difference between lagging and leading. And at gigawatt scale, that difference is everything.

I've seen a lot of control systems in my career. The ones that work best aren't the ones with the most capacity. They're the ones that know how to use what they have.

Phaidra's AI agents do exactly that.

They extend system response time by acting on leading indicators, reducing the need for oversized infrastructure while improving reliability and efficiency.

It's like traction control in a race car. You don't win by having the biggest engine. You win by staying on the racing line—maintaining grip, avoiding drift, and extracting every bit of performance the system has to offer.

That's what intelligent control does. And that's what the future of AI infrastructure looks like.

Final Log Entry: The machine only works when everything synchronizes

An AI factory is a machine built for intelligence generation. But it only works when every system coordinates in real time—power delivery, cooling response, workload scheduling, thermal management. Everything working in harmony to create an infrastructure asset that operates as a single unit of compute.

At the speeds AI workloads demand, I can’t track everything manually anymore. That's why AI agents became essential—not to replace operators like me, but to handle the coordination speed that exceeds what manual oversight can manage alone.

Phaidra's work within the NVIDIA Omniverse DSX ecosystem is defining what this looks like at gigawatt scale. Their AI agents deliver real-time precision that makes it possible with systems stable, efficient, and performing at their peak.

Because in the end, the fastest machine isn't always the one with the most power.

It's the one that can use it all, consistently and reliably.

From context to simulation: How class 3 and class 4 data enables predictive AI

Akshay Jindal — Tue, 02 Dec 2025 17:00:00 GMT

In many industrial systems, the sensors are loud, but the data is vague.

AI can replicate existing control strategies—and in many cases, perform them reliably. But to go beyond matching a manual baseline and actually improve outcomes, the system needs more than surface-level data. It needs to understand why things happen. That’s where AI data classification comes in. At Phaidra, we use this framework to structure data so that AI agents go beyond pattern recognition and into meaningful, predictive control.

Class 2 data shows behavior. Class 3 explains it. Class 4 prepares for what hasn’t happened yet.

If you need a refresher on Class 0, 1 and 2 data, check out our explainer here

We’ll focus on the high end of the classification spectrum that enables both foresight and hindsight.

Here’s what you’ll take away from this breakdown:

How Class 3 and Class 4 data differ in purpose and potential
Why simulated data isn’t hypothetical — it’s essential to system safety
How Phaidra uses these data classes to train agentic AI in real-world environments

Let’s start by exploring why this classification system matters — and what’s at stake when AI is fed the wrong kind of data.

Why AI Data Classification Matters, Especially for Control Systems

Not all data is created equal, especially when you're asking AI to make decisions with it.

AI data classification is a framework for assessing how useful a dataset is for intelligent automation. At the lowest levels, you’re dealing with unstructured or raw data (Class 0), then structured and semantically labeled data (Class 1), and finally cleaned, validated signals with timestamps and context (Class 2). But Class 2 is still just surface-level: the AI sees what happened, but it doesn’t understand why—and that’s a problem when it’s tasked with controlling real-world systems.

It’s the difference between spelling words correctly and understanding the meaning of the sentence.

Without contextually rich data, AI is left guessing—extrapolating from shallow patterns, unable to see the causal relationships that matter most. In systems like industrial cooling or power management, that kind of guesswork isn’t just inefficient… it’s risky. That’s why Phaidra’s AI doesn’t operate on what’s convenient. It operates on what’s complete, enriched, and ready to support autonomous control decisions.

So what exactly makes data “ready” and what sets Class 3 and Class 4 apart?

Class 3 Data — Contextualized and Dynamically Rich

Clean data is nice, but context is what makes it extremely valuable to the organizations utilizing it for decision-making.

Class 3 data is what happens when you stop just recording a system and start understanding it. It goes beyond complete and clean; it’s enriched with variation, context, and external signals that help AI grasp how a system actually behaves under different conditions. This includes things like operational variability — changes in setpoints and the corresponding sensor feedback that together reveal how the system behaves across its full range. It also includes integrating external data, like weather conditions that affect system load and/or performance.

Without this kind of data, the AI is flying blind through anything it hasn’t seen before—and it can’t learn better strategies if it’s only trained on tightly constrained system behavior. That’s why Class 3 data is foundational for smart control.

It gives AI the ability to map cause and effect, learn from corner cases, and prepare for real-world complexity, which is more important than textbook scenarios.

Here’s what Class 3 data enables:

Modeling of edge cases the system hasn’t encountered in daily operation
Backfilling of missing data by learning system dynamics
Feature generation for training simulators and agents

Imagine a data center that’s only ever run at 40% load. That data might be clean, but it’s not very helpful if your AI needs to control for a 65% load spike. Class 3 data enables AI to operate within those ranges safely, intentionally, and with purpose.

So what happens when AI doesn’t just understand how systems behave but can also accurately predict what comes next?

Class 4 Data — Predictive, Generative, and Forward-Looking

Understanding a system is great, but predicting it is next-level.

Class 4 data represents a leap in capability:_ it’s generated by simulators or digital twins that are trained on enriched Class 3 data_. Unlike earlier classes, which come from sensors and historical logs, Class 4 is synthetic; it’s forward-looking data created to simulate how a system might behave under new or extreme conditions. It allows AI to anticipate, besides simply responding.

Where Class 3 is observational, Class 4 is generative.

This is data that can think beyond what it has seen.

It’s the foundation of predictive control. It allows AI to rehearse scenarios before they occur, test actions before they’re taken, and develop strategies for edge cases that haven’t happened yet (and ideally never will). Crucially, these simulations are grounded in physics and real system dynamics.

We’re not inventing fantasy data.

Instead, we’re extending reality with constraint-based foresight.

Examples of what Class 4 data enables:

Forecasting future cooling loads based on external weather predictions
Simulating power outages or equipment failures to train AI agents on recovery paths
Modeling valve behavior under stress to prepare for rapid load changes

Because Class 4 data is generated from accurate Class 3 dynamics, every scenario stays within the bounds of what’s physically possible. It’s not just “what if” data; it’s “what now” intelligence, built from a system’s own rules.

So how does this future-facing data translate into real-world action, especially when paired with AI agents tasked with making decisions?

From Insights to Agentic AI

Seeing a problem is good. Solving it before it happens is better.

At Phaidra, Class 3 and Class 4 data do more than sit in a pipeline. They fuel two distinct but connected capabilities: insights and autonomous agentic control. On the insights side, Class 3 data adds the system context needed to detect anomalies and understand faults in real time. With Class 4, you move from awareness to foresight — simulating failure states that haven’t occurred yet to train models to recognize them when early signals emerge.

Reacting fast is good, but reacting fast with incredible accuracy is better.

That’s why prediction needs purpose.

This is especially useful in high-stakes environments like data center operations, where AI agents must make decisions about thermal management, energy use, and system stability in milliseconds — not minutes. In our work with Cooling Distribution Units (CDUs), for example, agentic AI trained on Class 4 data doesn’t wait for the thermal system to overshoot the supply water temperature. It acts before the spike hits — because it’s been trained on Class 4 data to recognize the early signs, like rising IT load.

It looks like magic, but it’s actually preparation.

How Class 3 and 4 data power Phaidra’s agents:

Insights: Analytics & Observability

Class 3 enables richer fault detection and causal understanding
Class 4 makes predictive maintenance possible by simulating future failure states
Comparing simulated (expected) vs. real (actual) values helps flag early anomalies

Agentic AI for CDUs

Proactive control based on forecasted IT load spikes

Phaidra’s self-learning AI agent is substantially better at reducing thermal spikes than traditional control systems.

Valve and flow behavior simulations train agents on how to respond preemptively
Preventing unsavory futures before they unfold

With simulation data, agents don’t start from zero. They start informed. That means they can reach human-caliber performance quickly, then learn alongside real-world operators. As they improve, operators get time back to focus on higher-value work instead of manual tuning.

So what makes this possible behind the scenes? It’s not just smart code. It’s the coordination and complexity required to generate the right kind of data in the first place.

Challenges in Moving from Class 2 to Class 4

Turning raw data into predictive intelligence is more organizational than it is technical.

New to data classes? Read our breakdown of Class 0–2 here.

Getting to Class 4 data isn’t a matter of adding more sensors or writing more code. It takes purposeful exploration, cross-functional coordination, and the discipline to build systems that reflect how the real world actually works. At Phaidra, moving from Class 2 to Class 4 means going beyond clean data pipelines into rich semantic context, meaningful variation, and physics-grounded simulation, all while making sure the AI is solving the right problem.

In other words, this isn’t plug-and-play.

It’s test, train, verify, align.

Here are the biggest challenges, and how we tackle them:

Problem: Class 2 data doesn’t explore system dynamics

It may be clean, but it’s often shallow—reflecting tightly constrained operations with limited variation. Without controlled experimentation, you don’t get the behavioral range needed for Class 3 enrichment.

Solution:

We work with partners to conduct targeted operational variation—manually adjusting setpoints and operating modes to help AI learn how the system behaves across its full range. That exploration becomes the foundation for modeling and simulation.

Problem: Organizational silos slow down data enrichment

Engineering teams understand the system. Data science teams build the models. Product teams know the use cases. But without coordination, context gets lost — and models miss the mark.

Solution:

We bring these groups together early. Domain experts inform the ontology. Data scientists create the mappings. ML engineers and research teams train and refine the models. The product ensures the right goals are baked into the training. This isn’t a handoff — it’s a collaboration loop.

Because we’ve worked in controls, data centers, and AI, we don’t need to guess how these systems operate. We use that background to move faster and design solutions that fit the real-world constraints from day one. Phaidra’s team brings deep expertise in artificial intelligence applied to mission-critical cooling infrastructure.

Problem: Data alone doesn’t create intelligence

Raw telemetry is just numbers. So it doesn’t carry much meaning on its own. In most building systems, data points are labeled inconsistently—if they're labeled at all. One chiller might show up as chiller1 while another shows as ch101, but neither tells the operator much without a decoder.

The risk isn’t that AI gets it wrong—it’s that it doesn’t know what right looks like.

Here’s what that transformation looks like, taking raw tags to semantically enriched Class 3 data that an AI agent can understand and act on:

Raw industrial data tags vs Class 3 enriched data

Solution:

We semantically map raw data, then build knowledge graphs to define how components interact. This gives the AI more stable, informative, and context-rich inputs — showing what affects what and why. We embed those relationships into simulators, enforcing the physics and rules of real-world operations. That’s how we make sure Class 4 data isn’t just synthetic — it’s believable, usable, and safe.

Scaling from Class 2 to Class 4 isn’t just about readiness. It’s about responsibility — making sure the data, the tools, and the teams are aligned to support AI that can act with confidence.

From Clean to Capable

AI needs more than clean data.

It needs the kind of data that tells a story, reveals patterns, and prepares it for what’s ahead. That’s what Class 3 and Class 4 data make possible.

Here’s what we covered:

Class 3 data provides the context and variability needed to model system behavior
Class 4 data is the result of simulation and enables foresight and predictive decision-making
Phaidra’s AI services use both to safely navigate complex environments with confidence

But enabling this takes exploration, coordination, and a commitment to grounding every insight in reality. The payoff is AI that doesn't blindly guess. It understands, anticipates, and adapts.

This is just the beginning.

As industrial systems generate richer data and AI agents become more capable, the gap between reactive and predictive control will only widen.

So the question isn’t: Do we invest in support for generating and maintaining Class 3 and 4 data?

It’s “How quickly can we get there?”

Creating AI agents for gigawatt-scale AI factories with NVIDIA Omniverse DSX blueprint

Jim Gao — Mon, 03 Nov 2025 17:00:00 GMT

This was a big week for Phaidra and the broader AI factory ecosystem.

During his keynote address at GTC DC, NVIDIA founder and CEO Jensen Huang unveiled the NVIDIA Omniverse DSX Blueprint — a comprehensive and open blueprint for designing and operating gigawatt-scale AI factories.

Phaidra is proud to build the future of AI factories (i.e. data centers designed specifically for massive AI workloads) alongside NVIDIA and the other DSX ecosystem members. We envisage a future where AI factories operate continuously at peak performance — aided by AI agents that vigilantly manage the complex infrastructure on a 24/7 basis.

Jensen announcing the NVIDIA Omniverse DSX Blueprint with Phaidra and others at GTC DC

Why now? Two major forcing functions are driving the need for the Omniverse DSX Blueprint:

Inefficiencies get magnified at gigawatt-scale: a single GW AI factory represents both a $50Bn investment and $200Bn revenue opportunity. Every 1% of inefficiency therefore represents $2Bn in lost revenue.
Extreme performance requires extreme co-design: AI factories today are so large, complex, and interconnected that they must operate as a single integrated machine rather than a collection of loosely-orchestrated components. This is the pathway towards step function tokens/watt improvements.

By openly sharing a reference architecture for the entire industry to build upon, we make it easier for everyone to collaboratively and iteratively improve energy efficiency, time-to-value, and reliability at scale.

Phaidra’s unique contributions to the Omniverse DSX Blueprint are:

Helping define the open data exchange standards and communications protocols that enable the various data center control planes (i.e. BMS, PMS, workload manager, etc.) to freely and openly communicate with each other.
Ensuring readiness for the new generation of agentic AI companies like Phaidra, Emerald AI, and others to readily integrate with the physical and digital twin parts of the AI factory — in pursuit of extreme performance.
Developing and sharing OpenUSD SimReady assets, and utilizing existing SimReady assets within the NVIDIA Omniverse digital twin, to rapidly train and evaluate our AI agents. This makes it easier for everyone to leverage robust simulation capabilities to improve AI factory performance.

The benefits of a collaborative, simulation-first approach are clear. Within the GTC DC keynote demo, an animation of Phaidra’s liquid cooling AI agent was shown in action (reposted below).

Phaidra’s self-learning AI agent is substantially better at reducing thermal spikes than traditional control systems.

This AI agent eliminates the thermal spikes resulting from massively synchronized AI workloads, which in turn enables the AI factory to run at higher TCS temperatures. The end result is substantially less power and capital required for cooling, and more power available for revenue-generating GPUs.

Our AI agents were developed and trained in simulation before they were deployed into real-world production systems. Phaidra collaborated closely with NVIDIA’s data center engineering team and used an operational digital twin of an NVIDIA DGX SuperPOD to rapidly prototype various AI agent architectures under varying environmental conditions. The optimal AI agent was trained further in simulation (i.e. bootstrapped on synthetic data) before testing on production NVIDIA DGX GB200 systems with live AI workloads.

Interestingly, we discovered that the AI agent trained in simulation dramatically improved the performance of the existing liquid cooling control system that had already been fine-tuned by human experts.

This is shown in the graph below, which illustrates the before/after performance of our AI agents at providing precision thermal control in liquid cooling systems. The red dots correspond to the traditional control system’s performance given a 30-70% load ramp. The blue dots correspond to an AI agent that had been trained in simulation. Finally, the green dots correspond to an AI agent after several hours of live learning on the production system (i.e. the reinforcement learning-based AI agent taught itself to get better without human intervention).

Phaidra’s AI agent teaching itself to get better at managing AI factory infrastructure without human intervention.

This mini-case study illustrates two important points:

Design and iteration in simulation can greatly accelerate your product development. What would have taken multiple quarters of iteration time (because this is mission-critical infrastructure) instead took weeks in simulation.
AI agents trained in simulation can already outperform existing systems — provided the simulation environment is of sufficiently high fidelity.

The Phaidra team is proud to work alongside NVIDIA and the team of industry-leaders to define the future of gigawatt-scale AI factories. Our goal is to openly share our learnings and know-how through the Omniverse DSX Blueprint initiative to ensure that AI factories can readily leverage agentic AI technologies to achieve extreme energy efficiency, time-to-value, and reliability.

Is your data ready for AI? A guide to data readiness

Chris Vause — Sun, 05 Oct 2025 16:00:00 GMT

A lot of attempts to use agentic AI fail because of something no dashboard ever flags: inaccurate or bad input data.

Sensor data is limited. Operators make mistakes. Sensors drift out of calibration or are moved and mislabeled. Sometimes sensors and machines stop communicating altogether. And sometimes, data just… misleads. The result is a system that appears legible but often tells a story that isn’t quite true. In the absence of complete truth, operators rely on tools—like failure detection mechanisms and intelligent automation—to reconstruct what’s real.

The most advanced AI tools can only do their job if the data is accurate.

In this article, you will learn:

The 4 initial data classes and why they matter for operational AI applications
How imputation and data quality influence analytics, alerts, and prescriptive guidance
A practical framework for assessing and upgrading your data readiness

Let’s begin by looking at the fragile foundation that underpins every operational decision: data.

Why Data Quality Is the Unseen Bottleneck in Agentic AI

Most AI failures are caused by broken data, not necessarily a broken algorithm.

Operators work with sensor feeds that were never designed for machine learning. These streams are noisy, incomplete, or completely lacking semantic context—especially when human error, sensor drift, or outdated systems go unchecked. And yet, AI models are expected to make intelligent decisions based on these flawed inputs.

This disconnect between raw data and AI readiness is what makes data classification essential.

Without a clear framework for identifying and improving data quality, even the most sophisticated systems can fail in surprising—and costly—ways:

Anomaly detection flags normal behavior. Poor labeling or corrupted time series data can make healthy operations appear as outliers, triggering false alarms or unnecessary interventions.

AI analysis misinterprets the system state. When data lacks coherence, tools designed to surface trends and recommendations assume faulty inputs are real conditions, leading to misinformed decisions.

Model training reinforces failure patterns. Low-quality or unstructured data—if used without cleanup—can encode operational mistakes directly into a model’s logic.

Model training fails to capture real system dynamics. When datasets lack edge cases or exploratory behavior, models can’t generalize or adapt.

Sensor drift goes unnoticed. Without a mechanism to catch slow changes in sensor behavior, long-term control and performance degrade silently.

Teams waste time chasing phantom problems. When data is untrustworthy, operators spend hours trying to fix issues that aren’t real—or miss the ones that are.

If you want AI to work, the data needs to be cleaned first.

And that starts by understanding what kind of data you're actually working with—and how to level it up.

The Four Data Classes (And Why They Matter)

Not all data is created equal, and most of it isn’t ready for machine learning.

What’s often called “sensor data” is actually a hierarchy of quality, usability, and meaning. Raw data (Class 0) may be the most immediate, but it’s also the least reliable. As data is structured, validated, and cleaned, it becomes increasingly valuable for automation, analytics, and machine learning.

This progression is captured in four distinct data classes: Class 0 through Class 3.

Here’s what each class looks like—and why it matters:

--> Class 0: Raw and Unstructured Semantic labeling is missing, limited, or too esoteric for external use. These are triplet values—timestamp, point name, and value—with no validation and frequent contradictions. Useful for troubleshooting or audits, but too chaotic for automation or AI to derive much value from it.

--> Class 1: Structured but Unverified This is tagged data with meaningful names, hierarchical metadata, and system context. It’s understandable across teams and usable for checking system limits or running manual analysis—but it may still contain physically impossible or misleading values.

--> Class 2: Clean, Coherent and Reliable This is the minimum viable dataset for trustworthy AI outputs. Invalid values have been corrected or removed via imputation. It’s physically consistent and trusted for tools that provide clean data insights. It must also be consistently monitored and maintained as the equipment providing the measurement can drift or fall out of calibration.

--> Class 3: Exploratory and ML-Optimized Built from intentional system manipulation, Class 3 captures dynamic edge cases and rare regimes. It supports robust machine learning and adaptive control—but only in safe, explorable environments. It requires effort, risk management, and strategic foresight.

Beyond simple validation, Class 3 data is enriched and contextualized. It includes semantic metadata, equipment hierarchies, and dynamic system relationships that give the system both measurements and meaning. External context, like weather or occupancy data, can be layered in, along with feature engineering to expose hidden patterns and internal states.

The result is data that fuels predictive control, adaptive optimization, and advanced fault detection.

But it’s costly to build, demands deep domain expertise, and depends on a solid foundation beneath it.

Because before you can unlock Class 3, your facility has to make the leap from Class 1 to Class 2.

And that leap depends entirely on one thing: imputation

How Imputations Keep Your ML Systems on Track

Data imputation—the process of replacing missing or invalid data with estimated values—is essential for clean, reliable datasets.

Even with semantic tags and structured inputs, operational data often contains errors that will mislead a model. Imputation helps repair those errors by removing anomalies, filling gaps, and approximating reality in ways that machines can trust. At Phaidra, our Insights agent helps surface anomalies in structured Class 1 data, making it easier for teams to identify gaps, inconsistencies, or invalid values that require imputation.

These imputations transform Class 1 data into the clean, coherent foundation needed for Class 2.

Here’s what effective imputation does—and why it matters in real-world operations:

--> Protects AI from training on bad data By correcting or excluding invalid time slices, imputations prevent ML models from internalizing faulty operational behavior or sensor anomalies.

--> Preserves performance when sensors fail When a sensor drifts or drops out, a well-placed imputation can keep the system running until a fix is implemented—without compromising model integrity.

--> Flags the need for deeper investigation Imputations are not permanent fixes. Each one should prompt teams to investigate the root cause, document the failure mode, and work toward a long-term solution.

--> Enables accurate real-time guidance Because effective agentic AI systems require Class 2 data, imputations ensure that bad inputs don’t crash its logic—or worse, lead it to take unsafe or suboptimal actions.

--> Requires human vigilance and documentation Imputations must be logical, rigorous, and clearly tracked. Without good hygiene, you risk masking real operational issues under temporary patches.

Sensor-aware imputation techniques are being actively explored in time-series machine learning research, offering promising tools for reducing bias and drift in operational data. The better your imputations, the more reliable any AI application outputs become.

But first, you need to know where you stand.

How to Assess (and Improve) Your Current Data Class

Most teams don’t know what class of data they’re working with until something breaks.

Without a shared vocabulary for data quality, it’s easy to assume your inputs are “good enough” just because they show up. But what seems usable for humans can still be disastrous for automation. That’s why having a clear classification—and knowing where your facility stands—is critical for building trust in any AI system.

Upgrading your data class starts with an honest assessment.

Here are a few questions every operations team should ask:

Is our data semantically labeled in a way others can understand? If not, you’re likely operating in Class 0—even if the values appear technically correct.
Do we have meaningful tags and structure across systems? If tags are intuitive, consistently named, and logically grouped, you may be at Class 1.
Are we correcting or flagging physically impossible data? If you’ve built a layer of imputations or exclusion rules, you’re likely entering Class 2.
Do we explore system boundaries intentionally to generate learning data? If your facility is capable of controlled experiments that create new regimes, you’re moving toward Class 3.
Do we have documentation for our corrections and approximations? Imputation hygiene is what separates one-time fixes from sustainable, trustworthy Class 2 data.

Tools that prepare your facility for AI can help teams visualize data maturity and pinpoint areas for improvement—before small problems become systemic risks.

Knowing your current data class is just the beginning.

The real opportunity lies in building toward what comes next.

Beyond Class 3 — The Future of AI-Ready Data

Class 3 is the beginning of what’s possible.

When facilities generate exploratory data intentionally, they enable machine learning models that are not just reactive but adaptive. But the next evolution in AI-readiness will demand more than structured data. It will require systems that can self-correct, self-label, and even generate synthetic training scenarios. Class 4 hasn’t been formally defined yet, but its outlines are already visible, especially when you look at what’s emerging in agentic AI systems and real-time adaptive control.

The teams preparing for AI factories are the ones investing in data maturity today.

Here’s what the future of AI-ready data could look like, and how to start getting there:

Self-healing data infrastructure Systems that detect and correct sensor failures or semantic errors in real time, before they degrade performance.

Synthetic data for model pre-training Safe-to-fail simulations and AI-generated sensor streams that accelerate model readiness without real-world risk.

Embedded validation at the edge On-device intelligence that filters and flags bad data before it ever enters your control system or data lake.

Continuous labeling and feedback loops Instead of static data tags, future systems will evolve their own schemas based on model performance and operator corrections.

Cross-functional collaboration from day one Data maturity isn’t just an engineering problem—it requires alignment between operators, data scientists, and domain experts from the ground up.

Emerging research into generative models for industrial time-series data is already shaping what the next class of AI-ready data could look like. The cleaner and more contextual your data becomes, the closer your facility moves toward AI-assisted control.

It’s not enough to clean data once.

Treat data quality as a living system.

Conclusion

Most agentic AI efforts don’t fail because of bad models. They fail because of bad data.

And as you’ve seen, that failure usually starts long before anyone notices.

By classifying your data, correcting its flaws, and preparing for the next level of AI readiness, you create the conditions for long-term performance, safety, and insight.

Take it from us — to deliver our AI control services, we need high-quality data. That’s why we include data classification improvement from the start and maintain it throughout.

Continuous improvement isn’t optional. It’s built into every deployment we deliver.

That’s how you lay the foundation for actually applying agentic AI at scale.

AI for data center operations: From reactive to proactive

Giuseppe Pinto — Mon, 11 Aug 2025 16:00:00 GMT

I used to sleep with my phone next to my head, volume all the way up.

Because when you're the one they call at 10:17 p.m. on a Sunday night… or 4:43 a.m. on a holiday… you learn to live in a kind of alert purgatory. Not because systems are failing constantly, but because no one upstream ever had the time or space to fix what wasn’t yet broken. It wasn’t sustainable. It wasn’t smart. It just was.

Now, for the first time, I’m seeing what it looks like to get ahead of the chaos.

The following log entries explore how to finally get your life back (and save on your investments).

This is a fictional account based on real conversations and operator insights.

In this article, you will learn:

Why most operators are stuck in reactive mode—and how to break out
How AI-powered insights built for predictive maintenance shift from “putting out fires” to preventing them
What it feels like when machines work with—not instead of—human expertise

Log Entry: The Hum of the Racks

Some rooms hum so loudly you forget what silence sounds like.

Data centers are like that. They’re dense, humming nerve centers of our digital infrastructure where uptime is king, and failure is a costly catastrophe. When you sit in a room like that long enough, you start to hear the problems before they even begin.

It’s no wonder the people who work there live in a kind of permanent readiness.

I used to think the goal was full industrial automation—that if we could just throw enough sensors, scripts, and algorithms at a system, we could finally step away. But the truth is, AI doesn’t replace our intuition; it amplifies it. And without guardrails, context, and judgment, that amplification can become dangerous.

Which is why expertise still matters more than ever.

The mistake isn’t that companies rely on AI. It’s that they expect AI could do everything alone.

This false assumption leads to a set of recurring problems in mission-critical environments:

False confidence in model fidelity A model trained in a clean, idealized setting often breaks down in the wild.
Lack of contextual data AI can’t always tell you why something’s wrong—just that something is.
Underutilized human feedback loops Insights from operators often go unlogged or unanalyzed, even when they catch what the problem is.

The danger isn’t the tech. It’s the belief that the tech is infallible.

Reactive culture builds around that belief. People stop asking questions. Systems fail silently. And the operator becomes the scapegoat when everything goes sideways.

But there has to be a better way than constantly living in react-mode.

What happens when the model fails quietly?

Log Entry: When the Model Fails Quietly

What happens when a system doesn’t know what it doesn't know?

AI won’t break anything. Instead, it follows the rules it’s given. In physical systems, those rules often leave out critical, undocumented steps that some operators may take for granted—like shutting off auxiliary equipment under edge conditions or overriding a sequence during seasonal transitions.

Practices that aren’t in the standard operating procedures but do happen regularly.

Early on, the AI might flag these manual changes as violations and disengage because it needs to understand why they’re happening. So it opts out.

The first time we saw this in action, we thought it was a bug. It was so subtle.

The system kept pausing under what seemed like normal conditions. At first, we thought it was a logic flaw. But when we traced the behavior, we realized our ops team had been manually disabling certain equipment during low-load conditions for months. It’s such a familiar auto-response that no one thought to mention it.

The habit was brought up to our team. They adjusted the constraints to account for the nuance. And the AI adapted.

This shows why we shouldn't expect perfection from Day 1. Because on Day 1, the AI runs within the limits of its training data. Something we knew but didn’t apply.

But almost a year in, the AI started to adjust to the real-world logic at our site.

I learned that these systems don’t run on code alone. They need lived experience and guidance from the people closest to the equipment—like a mentor-mentee relationship.

We supervise the AI while it’s running not because it can’t be trusted, but because it needs some time to learn why we do what we do.

Before that learning happens, expectations and reality often drift apart.

Leadership expects perfect optimization from day one. But without embedded knowledge from operators, the system can’t deliver no matter how polished the dashboard looks.

Here’s what we’ve seen happen when that knowledge isn’t baked in:

Operators carry the cost of gaps they didn’t create. Because if something fails, the system doesn’t take the heat. The operators (people) are the ones who get the 3 A.M. call.
Automation disengaged during edge conditions. If a manual override violates an unknown constraint, the AI just opts out.
Small inconsistencies snowball. What looks like a flicker in a sensor might point to deeper issues. But without proper context, the system can’t tell the signal from noise.
Trust in automation stalls. Especially when alerts fire off for problems no one on-site recognizes… or when the data doesn’t line up with what we’re seeing on the floor.

We took a different approach.

By embedding domain expertise into the system gradually, the system learned undocumented logic behind every operator’s intervention.

Since integrating this new system, things have sharpened fast. We catch sensor drift early, flag corrupted inputs, and filter out the noise. Once the system understood what “normal” really looks like, it started acting a whole lot smarter.

Patterns that used to take weeks to notice now surfaced in hours.

When you treat data like a living system, you invest in its health, and everything downstream becomes smarter.

One thing is sure, intelligence is earned, not assumed.

Log Entry: Smarter Doesn’t Mean Wiser

Why do even the smartest systems still need human judgment?

Because intelligence isn’t just recognizing patterns—it’s knowing which ones matter. It’s understanding when something looks off, even if it’s technically within spec. And it’s having the context to act on signals machines don’t fully understand.

That’s why I trust the systems we’ve helped shape.

Instead of flagging every flicker in the data as a crisis, the system highlights what actually needs attention. Such as what’s urgent, what can wait, and what’s just background noise. That’s where our judgment sharpens the edge of the tool.

You’d be surprised how much more effective that makes both of us.

We had a case where an efficiency dip triggered a medium-priority alert.

Not usually the kind of thing that sends you running. But this one came with confidence scores, historical trends, and a contextual narrative. We traced it to a valve on the verge of failure, and fixed it before it took anything else down with it.

This is the kind of AI I’ll stand behind.

When it works with operators, not around them, you get:

Prioritized alerts that focus attention where it’s needed most
Context-rich insights that reduce second-guessing
More time for strategic work, less wasted on filtering noise

Smarter tools are great. But wiser tools—the kind that let you lead, not chase—are what make the difference.

Log Entry: Revisiting the Bitter Lesson

The first time I read Richard Sutton’s “The Bitter Lesson,” I felt like I’d found scripture.

The essay makes a striking claim: over the long term, AI systems built on simple algorithms and massive computational power will outperform human domain expertise in nearly every field. It’s a clean, convincing argument—one that aligns with how breakthroughs in chess, image recognition, and video games have all unfolded.

Back then, I was so convinced, I summarized his entire book and called it “the Bible.”

But over time, my experience—first in research, then in industry—started to chip away at that faith. Because in physical systems, the lesson plays out differently.

In academic labs or digital simulations, you can assume nearly infinite, accurate data.

You can run millions of training iterations and tweak variables in sterile environments. But in the real world labels are messy, sensors drift and greenfield projects give you nothing to start from.

The environments I worked in—from energy systems to building automation—were anything but ideal for training a model. And that gap between theory and application becomes glaringly obvious once you’re the one responsible for real-world outcomes.

Let me put it this way: even if the algorithm could eventually outperform every human expert, “eventually” doesn’t pay salaries or keep systems running today. And even if you had a perfect model, you’d still need a domain expert to validate it.

That’s not a bug—it’s part of the process.

Because the moment you try to simulate physical phenomena with enough fidelity, you’re building a world where domain knowledge (physics) is embedded into the very structure of the model itself.

This isn’t just a theoretical stance—it’s operational reality.

NVIDIA, for example, has started to model this with Omniverse, where they’re building high-fidelity digital twins to train AI inside real-world constraints.

I've seen cases where AI systems optimized for metrics that didn’t reflect actual system health. I’ve spent months calibrating models only to realize that the data we had wasn’t enough—and might never be.

And every time an algorithm suggested a course of action that looked right on paper but would have risked damage in practice, it was a human who intervened usually because we had access to information the system didn’t. The model lacked context—like a sensor that had been moved from a pipe to a cabinet, so it was still reporting data, but now measuring something completely different. Or an undocumented workaround the team had used for years. Or simply a piece of logic that the data didn’t capture at all.

The lesson I’ve learned isn’t bitter—it’s sobering: AI needs us.

Not “just for now.” Likely for a long time.

It’s tempting to idolize the elegance of computation. But in our world, it’s like tuning a violin in a vacuum. The engineers who succeed tomorrow aren’t the ones who try to outpace AI—they’re the ones who shape it, validate it, and push it to serve real-world constraints.

So no, I don’t think Sutton got it wrong. I think he got it right for the wrong domain.

And that makes all the difference.

Log Entry: Why Pure AI Struggles in Physical Domains

Not everything you can measure is useful, and not everything useful can be measured.

That’s the first thing I remind myself when an algorithm spits out something that “makes sense” but feels wrong. In industrial systems, lacking accurate data and holistic context at all moments is why most AI tools fail.

They’re not broken.

They’re just blind.

A lot of models are trained on sanitized datasets—neatly labeled, low-noise environments that look nothing like the actual facilities we operate in. That’s a problem. Because real-world systems are noisy, unpredictable, and weird.

We deal with edge cases while engaging with them.

We had a system once that flagged efficiency issues three weeks after we’d already diagnosed and resolved the actual fault manually. Why? Because the AI was still waiting for a specific pattern that only showed up in the simulation—not the real data.

That’s how we lose time, trust, and tolerance.

It’s not just our operations team who suffer for it.

When AI misses the mark:

Engineers waste cycles chasing artifacts instead of solving problems.
Leadership thinks the site is under control—until it’s not.
Operators become numb to alerts that don’t reflect reality.

So Phaidra keeps the operator in the loop—where they belong. These AI-powered insights built for predictive maintenance are a gamechanger. It learns from automated decisions. It observes what we prioritize, how we validate issues, and what we do with the information next.

It’s AI, yes. But AI that finally speaks operator.

Log Entry: What AI‑Insights Actually Do

You can’t automate what you can’t measure correctly.

That’s a truth most people miss when they talk about AI. Fancy dashboards and alerts mean nothing if they’re built on noisy data, mislabeled trends, or metrics no one actually uses. And in my experience, the best systems don’t just analyze—they guide, validate, improve, and diagnose.

That’s exactly what our system was designed to do.

AI readiness is helpful, but this is how Phaidra actually helps you do your job better:

--> It guides maintenance proactively

Instead of treating every sensor blip as a potential emergency (or something to completely ignore), it identifies what truly requires action… and what doesn’t. It’s like using a magnet to pull the needle from the haystack.

Fewer false flags, contextualized alarms and more meaningful direction.

--> It validates maintenance and proves ROI

I don’t have to guess if a fix worked or if a new component was worth it. With every action tracked and measured, our team can show the cost-benefit of every intervention. Recently, our system flagged performance issues with a piece of equipment that had been underperforming for months. After it was serviced, we saw massive improvements in days.

--> It improves performance over time

Catching small drops in performance adds up. Whether it’s a lazy valve or a misbehaving sensor that triggers excess equipment use, nudging systems back into alignment helps recover performance that would’ve been lost silently.

--> It accelerates diagnostics and root-cause analysis

Finally! No more generic “something went wrong” alerts. Phaidra added context using confidence scores, historical comparisons, and correlated metrics to help our team solve issues fast and move on.

This is how you gain trust—not by removing the operators, but by giving us the tools we actually need.

Log Entry: The Future Isn’t Frictionless

The first time Insights flagged an issue before I did, I was actually impressed.

At first, it was a gut check. I’ve spent years listening for problems before they show up on a screen. So when a machine caught a failing valve before I even heard it hiss, I didn’t shrug it off. Instead, I paid attention.

Because maybe that was the point. Not to prove who’s smarter, but to prove what’s possible when we work side by side.

That’s when it hit me: I’m not just maintaining systems anymore. I’m helping shape them.

Work looks different now.

I still walk the floor, trusting my senses. But now, I’m calibrating models, validating feedback loops, and translating edge cases that would’ve fallen through the cracks. And I’m not the only one. Across the site, we find ourselves less in reactive mode and more proactive.

The game has changed.

How much? We’re still figuring out.

Conclusion Log: From Chaos to Clarity

We’ve all felt the weight of reactive mode—responding instead of preventing, guessing instead of knowing, hoping systems won’t fail while dreading the moment they do.

But it doesn’t have to be that way anymore.

What we’ve built here with Phaidra won’t replace the operator. It reinforces them. Because when AI works with us—not around us—we improve stability and reliability, thus increasing overall efficiency and clarity.

And clarity, especially, changes everything:

Fewer false alarms, more targeted action
Less second-guessing, more confidence and speed in every fix
A shift from chaos to calm, from burnout to balance

Data center cybersecurity: Threats and best practices

Chadwick Thresher — Wed, 11 Dec 2024 17:00:00 GMT

In 2024, significant cybersecurity challenges, like the National Public Data (NPD) breach, exposed the sensitive information of millions, including Social Security numbers. This event emphasizes the growing complexity of cyber threats and the critical need for data center operators to remain vigilant.

With these increasingly intricate threats, proactive measures are essential to safeguard critical infrastructure. This involves understanding emerging risks, crafting comprehensive defense strategies, and carefully evaluating the role of cloud-based service providers in security planning.

This blog will equip you with the insights needed to:

Identify and navigate emerging cybersecurity threats.
Implement practical strategies to reduce risk.
Weigh the pros and cons of outsourcing to cloud-based service providers.

By staying informed and adopting a proactive stance, your data center can remain secure, resilient, and ready for the future.

Understanding the Threat Landscape

With ever-evolving cybersecurity, data centers are integral to the digital infrastructure for supporting businesses, governments, and day-to-day life. Housing vast amounts of sensitive information and enabling critical operations, they’re prime targets for malicious actors seeking power, profit, or disruption.

Why the IT/OT Boundary is the New Frontline

Traditionally, IT (Information Technology) and OT (Operational Technology) systems in industrial environments operated as isolated entities, with little interaction or overlap. IT systems manage data, communications, and administrative functions, while OT systems control physical processes and machinery. However, the push to interconnect these systems has grown in pursuit of operational efficiency and innovation.

This interconnectivity, while advantageous, introduces risks. The boundary between IT and OT has become a prime target for cyber attackers. At this critical juncture, attackers often infiltrate IT systems and pivot into OT environments—systems essential to industrial operations and infrastructure stability. These incursions are not random or hasty. They’re calculated, long-term operations designed to establish a foothold for future exploitation.

Threat actors, particularly those with significant resources, target this boundary for several reasons. IT systems often serve as an easier point of entry due to their connectivity, making them a springboard into more sensitive OT systems. The motivation isn’t always immediate disruption. In many cases, these attackers are silently pre-staging their access to prepare for larger, more impactful events.

For data centers, the integration of IT and OT environments offers a practical yet advanced approach to improving operations. The industry, however, has been cautious to adopt it. Traditionally, these systems are kept separate to reduce perceived risks, with OT services, like building management systems, isolated from IT networks. Yet, this separation can limit oversight and the advanced management and data processing capabilities provided by cloud and other IT-hosted services.

Interconnecting these systems brings clear advantages, such as streamlined management and improved functionality. Still, concerns about increased vulnerabilities persist, as a compromised IT system could serve as a launchpad for deeper infiltration into OT networks, disrupting critical operations. By prioritizing robust security measures, organizations can address these risks effectively, making the IT/OT boundary a secure and practical connection point for innovation.

Hacking for Profit or Power

Cybercrime has developed into a sophisticated marketplace, where attackers no longer need deep technical skills to launch devastating attacks. Instead, they can outsource nearly every step of an operation through "cybercrime-as-a-service" models.

Services such as ransomware marketplaces and initial access brokers enable even non-technical individuals to orchestrate complex breaches. For example, attackers can hire specialists to handle reconnaissance, exploitation, or even money laundering, often with service-level agreements (SLAs) and credits for “failed” attempts.

This commoditization of hacking has drastically lowered the barrier to entry, increasing both opportunistic and deliberate attacks. Opportunistic attackers cast a wide net, scanning for exposed vulnerabilities or misconfigured systems, while deliberate actors invest time and resources to penetrate specific targets.

Artificial Intelligence (AI) adds a tool for cybercriminals. Large Language Models (LLMs) can generate flawless phishing emails, mimicking natural language patterns, in any language native to the targets, to bypass traditional red flags like grammar errors. Beyond phishing, AI can automate reconnaissance, build exploit code, and even adapt in real time to a defender’s countermeasures. Attackers are deploying their own AI agents and models with focused tasks and defined lines of management. This ensemble method of deployment enables greater attack system "resiliency", adapting to each frontline agents’ experience. Increased effects on the end objective results as manager and planner agents select the most appropriate tools for current and future operations.

The growing access to sophisticated tools and services creates a critical challenge for data centers. Whether financial gain or long-term disruption motivates attackers, their ability to operate at scale means no organization is immune.

Protecting your infrastructure requires not only robust defenses but also a clear understanding of the evolving tactics used by these adversaries.

The Top Threats for Data Centers in 2025

As cyber threats grow more sophisticated, data centers face a unique set of risks that threaten both their operations and their customers’ trust. Understanding these threats is the first step to building a resilient defense.

1. Credential Misuse

Credential misuse remains the top attack vector, with compromised passwords acting as the gateway for most breaches. Attackers use techniques like credential stuffing or phishing to gain access, then “live off the land” by leveraging existing tools and permissions within the system. This tactic reduces their detectability and maximizes the damage they can inflict.

According to a recent Uptime Institute survey, 75% of data center operators reported experiencing cybersecurity incidents in the last three years, with credential misuse consistently identified as a primary attack vector.

2. Lite-Phishing and OAuth Trojans

Traditional phishing attacks have evolved into “lite-phishing,” which are like stealthy exploits. Lite-phishing omits suspicious links and attachments, instead relying on well-crafted messages to bait responses.

Similarly, OAuth Trojans disguise themselves as legitimate applications, tricking users into granting dangerous permissions. Once they’re authorized, these malicious apps can impersonate users, allow access to critical systems, and enable lateral movement throughout and across environments, posing a significant threat with broad integrations.

3. Supply Chain Risks

The growing interconnectedness of supply chains creates new entry points for attackers. By targeting smaller, less-secured vendors, cybercriminals can infiltrate larger organizations through backdoor vulnerabilities. For example, attackers might exploit a vendor’s poorly secured identity or workstation to gain access to a data center’s infrastructure, cascading risk throughout the network.

With 43% of data centers allowing third-party access to their networks, as highlighted in the security survey, the risks posed by supply chain vulnerabilities are increasingly pressing. A single weak link in the vendor chain can expose an entire operation to exploitation.

A Data Center Nightmare

Picture this: A data center partners with a third-party vendor to handle routine system maintenance. One day, an engineer at the vendor receives an AI-crafted, error-free phishing email disguised as a routine service request. The email lacks suspicious links or attachments, making it indistinguishable from legitimate correspondence. Believing it’s genuine, the engineer responds and is connected to a very helpful support engineer who passes a link to a new support web application that’s required. This “new tool” is opened, and the initial login and acceptance windows are quickly dispatched.

What happens next? The attacker uses this OAuth Trojan and the permissions that were willingly granted to impersonate the engineer. The stolen identity is used to gain initial access to the vendor’s network. From there, they move laterally, uncovering a direct pathway into the data center’s operational systems. Within days, the attacker manipulates critical cooling systems, overheating servers and causing widespread outages.

The fallout results in hardware failures, customer disruptions, and weeks of operational downtime. Worse, the incident severely impacts customer trust and exposes gaps in vendor oversight.

Practical Steps to Reduce Risk

Recognizing threats is one thing. Building a defense to match them is another. For data center operators, a reactive stance isn’t enough, so proactive measures must be woven into every layer of the operation. The goal is resilience, not perfection.

Here’s how you can tackle the most pressing issues with practical strategies.

Segmentation

Think of your network like a ship. Without bulkheads to contain flooding, a breach in one area can sink the entire vessel. Segmentation is the bulkhead of your cybersecurity strategy—it isolates critical systems, making it harder for attackers to move laterally and wreak havoc.

Traditional network segmentation remains valuable, creating detonation zones where breaches can be monitored and contained. The Purdue Model offers a structured approach to controlling traffic between levels (North-South) and within them (East-West).

But today, we can take it further with micro-segmentation. Imagine putting each device into its own private VLAN. If one device is compromised, it’s like locking the door to a single room rather than the entire building.

Identity segmentation plays a complementary role. If one user’s identity spans both high-risk activities, like accessing operational systems, and low-risk ones, like checking email, a breach of that identity becomes catastrophic. Assigning distinct identities for different functions limits the blast radius of any compromise.

Identity Management

“Identity is the new perimeter” is more than a catchy phrase—it’s the reality of modern cybersecurity. Attackers don’t break in. They log in. So, the strength of your authentication measures often determines the strength of your defenses.

Multi-factor authentication (MFA) is the backbone of identity protection, but it’s not one-size-fits-all. For systems where MFA isn’t an option, deploying a bastion—a secure checkpoint where access is monitored and enforced—adds a critical layer of security. Think of it as a toll booth that checks every vehicle before it gets on the highway.

The Consumer Authentication Strength Maturity Model (CASMM) offers a helpful framework for gauging and improving your authentication practices. Moving from shared passwords (the weakest layer) to hardware tokens like YubiKeys (the strongest) is more than an upgrade. It’s a necessity since attackers outsource phishing like it’s a gig economy.

Traffic Monitoring

Traffic monitoring is your early warning system. It detects anomalies before they escalate into full-scale breaches. For data centers, where the stakes are high, proactive monitoring is critical to identifying threats that traditional defenses may miss.

Every network has a rhythm with consistent traffic patterns that reflect normal operations. Establishing this baseline is the foundation of effective traffic monitoring. Deviations from this norm, such as unexpected spikes in data transfers or irregular access attempts, often signal malicious activity.

A recent security survey revealed that 62% of data center operators struggle with identifying and responding to anomalous traffic patterns due to insufficient resources or outdated tools.

Beyond simple anomaly detection, protocol-aware tools analyze traffic patterns and content to uncover subtle signs of malicious behavior. These tools can differentiate between routine activities and potential threats, such as unusual data flows indicative of lateral movement or exfiltration attempts.

Detection is only half the battle. Real-time response mechanisms ensure that once an anomaly is flagged, appropriate actions are taken immediately. This might include isolating affected systems, notifying security teams, or triggering automated countermeasures.

Operators using real-time traffic monitoring reduced response times by 40% compared to those relying solely on periodic manual checks. This highlights the value of continuous monitoring for mitigating threats.

Effective monitoring also involves understanding the context of anomalies. Without the tools to analyze these nuances, organizations risk overreacting or, worse, missing genuine threats. Investing in AI-driven monitoring systems capable of contextualizing anomalies would detect deviations and provide actionable insights into their cause, allowing for targeted and effective responses.

Vulnerability Prioritization

Fixing every vulnerability isn’t feasible, nor is it wise. The Common Vulnerability Scoring System (CVSS) is often treated as gospel, but it doesn’t account for the specific context of your environment. A vulnerability with a high CVSS score might pose little risk if it’s buried deep in a system with no external exposure.

Threat-informed models, like the Exploit Prediction Scoring System (EPSS), offer a more nuanced approach. By analyzing hundreds of data points, these tools predict the likelihood of exploitation, helping you prioritize what matters most. And when a patch isn’t an option? Compensating controls—like restricting network access or tightening permissions—can buy you the time needed to implement a more permanent solution.

Threat Modeling

Threat modeling is where strategy meets execution. By mapping out how your systems interact, what data flows between them, and where vulnerabilities lie, you can anticipate attacks before they happen. Frameworks like STRIDE or its expanded version, STRIDE-LM, provide a structured way to identify threats like spoofing, tampering, or lateral movement.

This isn’t a one-time exercise. Threat modeling should evolve with your systems, offering an up-to-date blueprint for addressing the most pressing risks. It’s the difference between guessing what an attack may target and knowing what systems would impact you most, how those impacts could materialize, and strategically reinforcing your defensive posture.

A Smarter Foundation

Reducing risk is about building layers of defense that reinforce each other. Segmentation isolates attacks. Identity management limits their scope. Traffic monitoring alerts you to anomalies, while vulnerability prioritization and threat modeling keep your efforts focused where they matter most.

This makes your data center secure and resilient—ready to withstand the inevitable.

Pros and Cons of Remote Cloud-Based Service Providers

As data center operators face increasing pressure to adapt to evolving cyber threats, many turn to remote cloud-based service providers to offload certain responsibilities. These partnerships can bring significant advantages, but they also introduce unique risks. Understanding both sides of the equation is critical to making informed decisions that align with your operational goals.

The Double-Edged Sword of Outsourcing

When you partner with a remote cloud provider, you’re essentially inheriting their risk posture. If their defenses are robust, your systems gain an additional layer of security. But if their approach is lax, that risk becomes your problem. The challenge is visibility—how much insight do you have into their processes, tools, and response capabilities?

For instance, questionnaires and compliance audits are often used to assess vendors, but let’s be honest: these methods are frequently check-the-box exercises. True due diligence requires deeper conversations, clear service-level agreements (SLAs), and ongoing evaluations to ensure alignment.

The Risks of Outsourcing

Inherited Risk: Your provider’s vulnerabilities become your vulnerabilities. If they fall short of securing their systems, the ripple effects can directly impact your operations.
Supply Chain Abstraction: Cloud providers often rely on their suppliers, creating additional layers of risk you may not see or control.
Core Competency Erosion: Outsourcing essential tasks can lead to a gradual loss of in-house expertise, making you more dependent on external partners over time.
Misaligned Priorities: Your goals and those of your provider might not always align. While you focus on reliability and security, they may prioritize cost efficiencies or scalability, potentially leaving gaps in critical areas.

The Advantages of a Trusted Partnership

When done right, outsourcing can free your team to focus on core operations and strategic goals. It’s more than saving time. It’s about accessing specialized expertise and avoiding the burden of maintaining services that aren’t central to your business.

Efficiency Gains: By offloading non-core tasks, your team can direct its energy toward higher-value activities.
Outcome-Focused Collaboration: A strong partnership can provide thought leadership and insights, helping you innovate and stay ahead of emerging challenges.
Reduced Overhead: Maintaining ancillary services in-house often leads to creeping costs and risks, especially when budgets tighten. Outsourcing can mitigate these challenges, ensuring consistent quality without straining internal resources.

In a 2024 survey, nearly half of operators reported measurable operational improvements after partnering with providers who aligned with their security goals.

Key Considerations for Choosing a Provider

Alignment with Your Goals: Look for providers whose priorities align with your own, particularly in terms of security and long-term reliability.
Transparency and Accountability: Ensure clear SLAs that define not only performance metrics but also their responsibility in the event of a breach.
Collaborative Approach: The best partnerships go beyond transactions—they act as extensions of your team, with a shared commitment to achieving your goals.

Use a decision-making matrix to evaluate providers, balancing their strengths against the specific needs of your operation.

Outsourcing isn’t a silver bullet—it’s a strategic decision that requires careful consideration. For data center operators, the choice comes down to whether the benefits of freeing up resources and accessing external expertise outweigh the risks of dependency and reduced control.

By thoroughly vetting providers and fostering strong partnerships, you can make outsourcing a tool, not a liability.

Conclusion

Cybersecurity is an infinite game. Threats evolve, tools become more sophisticated, and attackers adapt in ways we can’t always predict. But by understanding the landscape, taking proactive measures, and carefully evaluating partnerships, we position ourselves not just to survive but to thrive in this ever-shifting terrain.

Data centers are the backbone of modern infrastructure, and their security is non-negotiable. Whether it’s reinforcing the IT/OT boundary, implementing segmentation, or making strategic outsourcing decisions, every action contributes to a stronger, more resilient foundation. The key is preparedness.

As you move forward, remember that sophisticated security requires creating layers of defense that work in harmony. Ask the right questions, work with the right partners, and stay ahead of the curve. Because in the end, the goal isn’t only to protect what you have—it’s ensuring your operations remain reliable, your reputation intact, and your future secure.

Stay vigilant. Stay resilient. And keep playing the game.

Future-proofing data centers for AI & high-density compute

Dan Bishop — Mon, 18 Nov 2024 17:00:00 GMT

If you're a data center operator today (or really any industrial controls operation), you're feeling the pressure. The landscape is changing fast, expectations are increasingly high. With the rise of AI compute loads and high-performance/high-density computing, expectations of performance will continue to grow at an incredible rate. While there is an arms race to secure new land and power for future data center builds, there is an incredible amount of currently existing infrastructure in need of an update to handle the demand in advance of new builds becoming operational. It's not just about keeping the lights on anymore—it's about making sure your powertrain, cooling systems, and infrastructure are ready to handle what’s coming.

Whether it’s preparing for AI integration, determining how to unlock stranded power, or ensuring effective communication between your tenants and your facility teams, future-proofing your data center means planning well beyond immediate demands. These challenges aren't something you can afford to ignore if you want to stay ahead.

In this article, you will learn:

Why assessing your powertrain is the first step to AI optimization.
Tips to move towards unlocking stranded power and avoid overloading your system.
The critical role of proactive communication between tenants and operators.

Push the Mundane to the Machine

We’re beginning to see artificial intelligence applications revolutionize data center operations, bringing new efficiencies and capabilities but also new challenges. AI compute loads, especially those used in advanced applications like Chat-GPT that have incredibly vast datasets and equally immense compute demand, push data centers to their limits. As these demands continue to grow, operators need to rethink how their infrastructure is designed to handle these high performance demands.

One of the clearest benefits of AI optimization is improved Power Usage Effectiveness (PUE). AI systems can monitor and adjust energy consumption in real-time, ensuring that cooling is distributed efficiently across the facility, reducing the energy intensity related to a more broad, conservative cooling approach that is typical in the industry. This dynamic control approach significantly reduces wasted energy, helping data centers maintain optimal performance while keeping costs down. Additionally, AI-driven monitoring is key to elongating the life of equipment. By continuously analyzing the performance of critical systems, AI can detect early warning signs and enable predictive maintenance, reducing downtime and extending that lifespan. And preparation is important.

Leadership needs to assess the entire facility powertrain and determine if it is engineered, designed, or capable of high compute densities. Being proactive is essential as compute loads, and energy demands, increase.

Managing IT loads efficiently is also critical as AI workloads surge. AI-powered systems can dynamically allocate resources, ensuring that power and cooling are directed where they’re needed most. This prevents overloading and maximizes the efficiency of the entire data center. AI compute loads, such as those used in GPT, will significantly affect operations, making predictive maintenance and resource allocation all the more crucial.

Hence, preparing for AI in facilities requires a thorough assessment of the powertrain. Assessing the powertrain from the utility grid all the way down to the rack PDUs is the first step in AI optimization. Operators need to ensure their systems can handle current workloads efficiently while also scaling to meet future demands.

Powertrain Assessment and Unlocking Stranded Power

As mentioned, one of the most critical aspects of future-proofing data centers is ensuring that the powertrain—from the utility grid to the rack PDUs—is designed to handle the growing demands of AI and high-performance compute. When AI workloads increase, operators need to evaluate whether their infrastructure can support these higher densities without straining capacity or risking downtime.

A common challenge is determining whether the current infrastructure can handle future demands. Let’s say, for example, that a data hall was designed for a 5-megawatt load. You need to gauge if the high-performance compute is going to push load capacity beyond threshold limits. So operators need to continuously monitor and compare their design load intent to their actual IT load, avoiding potential overloads that could compromise performance and safety.

Another key opportunity for data center optimization lies in stranded power, which typically resides in the battery plant of the Uninterruptible Power Supply (UPS) system. These UPS batteries often sit unused, serving as backup in case of emergencies or during monthly maintenance cycles. However, they represent untapped potential. By efficiently managing and intelligently utilizing this stranded power, operators can reduce unnecessary grid consumption and improve overall power usage efficiency. PUE records the total power consumed by the site so the energy stored in a battery is already counted, therefore distributing it at strategic moments can help reduce overall PUE.

Stranded power is in the batteries of the UPS system. Batteries are constantly trickle-charging, and most of the time, the energy just sits there. Finding ways to purposefully use this power would not only add a layer of efficiency but also support better energy management, especially during peak loads or unexpected surges in demand.

By addressing these key areas, data centers can unlock stranded power and ensure their powertrain is equipped for the demands of AI and high-performance computing.

Strategies for Unlocking Stranded Power

Assess the Entire Powertrain: Start by examining your powertrain from the grid to the rack PDUs. This holistic approach ensures that every component is functioning as intended and has the capacity to handle future AI-driven loads.
Increase Power Capacity: As high-compute workloads grow, data centers need to ensure they can scale. Operators should plan for increased power capacity by upgrading UPS systems, busways, and other power distribution components.
Battery Management: Optimizing UPS battery management is essential for unlocking stranded power. Intelligent systems can monitor battery health and usage, ensuring they are efficiently utilized and not simply sitting idle for long periods.
Planning for Future Demands: Future-proofing requires forward-thinking strategies. As AI compute loads continue to grow, it’s not enough to design for today’s needs. Operators must consider what the next 5 to 10 years will bring in terms of IT load growth and power demand. This means building in the flexibility to scale and ensuring that your power infrastructure can adapt.

Data Management

Effective data management is not just crucial, it can be a competitive advantage. As AI and high-performance computing increase demands on infrastructure, optimizing resource allocation, enhancing system performance, and ensuring quick disaster recovery allow operators to stay ahead of any challenges while maintaining efficiency and minimizing risks.

The only way to stay on top of all these challenges is to have an adequate robust and accurate operational telemetry dataset to analyze. One of the key aspects of managing data center operations is ensuring your cooling systems are keeping up with the actual, changing demands on the floor.

It’s not enough to assume an even distribution of IT loads across the data hall. Certain areas heat up faster as workloads ramp up in specific racks. So if a tenant loads the data hall unevenly—say, concentrating servers in one corner—then you’re dealing with heat buildup in that one spot. Meanwhile, the cooling is still distributed evenly across the hall, providing unnecessary cooling and not properly handling the heat spike in that corner.

To effectively manage cooling, you need systems that can adjust dynamically based on real-time load data, rather than cooling the entire hall uniformly. This targeted approach prevents overstraining the cooling infrastructure and helps avoid inefficiencies and potential failures caused by uneven heating.

Staying proactive with cooling adjustments, especially as high-performance compute and AI loads keep growing, ensures you’re meeting the facility’s cooling needs where they’re most critical.

Continuous monitoring allows operators to avoid a mismatch between the designed capacity and actual usage, which is critical for avoiding bottlenecks and unnecessary strain on resources.

Leveraging AI to optimize data flow and cooling systems is another powerful strategy for future-proofing data centers. Industrial AI systems can dynamically adjust cooling and resource usage based on real-time data, ensuring that performance is maximized without wasting energy.

This kind of data-driven approach helps maintain peak efficiency, especially as workloads fluctuate. AI’s role in resource allocation becomes even more essential as compute loads increase, helping operators ensure that both power and cooling systems are optimized for performance.

Additionally, a comprehensive data management strategy also includes disaster recovery planning. Integrating data systems can enable faster responses to unexpected failures and reduce downtime. This involves having backup systems in place, ensuring that data is properly replicated, and automating failover processes—a backup mode that switches to a redundant system when a primary system fails. The goal of failover is to minimize downtime and data loss and ensure business continuity from a secondary disaster recovery (DR) site.

With real-time monitoring and integrated recovery systems, data centers can respond quickly to disruptions and maintain continuity of service, even in the face of system-wide issues.

Incorporating these strategies optimizes daily operations and lays the groundwork for future scalability. Effective data management hinges on continuous monitoring and proactive adjustments, allowing operators to maintain the balance between IT loads and design capacity, improve system performance, and be prepared for unexpected disruptions.

Building Management Systems (BMS)

Building Management Systems (BMS) are at the core of running a data center. They’re what you rely on to keep track of all the key infrastructure—whether it’s power, cooling, or security. As AI workloads ramp up and high-performance computing pushes demands even higher, it's crucial to make sure your BMS integrates with other systems. But just as important is improving communication between tenants and operators to stay ahead of any changes or challenges coming your way.

One of the biggest advantages of an advanced BMS is automation that anticipates and plans accordingly, especially when it comes to managing power and cooling. The fewer manual inputs, the better. Reducing these inputs frees operators to focus on valuable improvements across the facility rather than routine, reactive tasks.

AI integration also offers dynamic response capabilities, adjusting power and cooling levels based on real-time data and load shifts. However, AI needs effective communication between tenants and colocation providers to function optimally. Even a slight heads-up on anticipated load changes allows the AI to prepare accordingly, minimizing human error and enabling smoother, more efficient responses to unexpected demands.

Cooling is a big deal, especially with AI workloads generating more heat than ever. That’s where Phaidra’s AI control service can really shine as a complement to a BMS. While a BMS provides basic monitoring, our AI Agent uses real-time data to adjust the cooling based on what's happening at the moment and what’s expected to happen in the near-term future, not just some pre-set rules. The key is to maximize efficiency by operating near the hot aisle limit temperature without crossing it. This approach conserves energy by fine-tuning cooling to precise levels, ensuring that systems stay within safe limits without unnecessary cooling. Keeping an eye on cooling becomes even more important as high-compute loads keep growing.

Future Proofing with Communication Systems

Now, it’s not just about technology—it’s also about communication. There’s a responsibility both ways for change management. The data hall tenant needs to be responsible for giving the facility management personnel a change management process as well. If the tenant and the facility team aren’t on the same page, things can go sideways fast. Setting up proper communication channels is critical so that any load changes are flagged early and both sides can act quickly.

Future-proofing is more than just having the best hardware or AI-driven systems. It needs improved communication at every level. Communication systems between tenant and operator need some signal that goes out in advance. Data center operations run most smoothly when tenants and facility teams are proactively communicating. Integrating automated, real-time notifications can help make that happen. These systems reduce the strain on operators, allowing for faster responses to load changes and avoiding potential crises.

AI systems can help with automating some of these notifications, but you can’t remove humans from the process. Someone has to initiate the system, and the tenant needs to provide accurate information to avoid surprises. It involves working together to ensure that load changes are managed smoothly with everyone involved in the process from the start.

Data Center Infrastructure Management (DCIM) and Integration

DCIM, or Data Center Infrastructure Management, is an essential tool in modern data centers. It’s how you keep a close eye on everything—tracking your powertrain, monitoring your IT load, and making sure all your systems are running smoothly. But DCIM does a lot more than just basic monitoring. It helps you manage energy, predict future demands, and even streamline maintenance when integrated with other systems.

One of the biggest advantages of DCIM is the ability to track everything in real time. Whether it’s the power your servers are pulling or how much cooling is being used, you can spot problems before they become big issues. You get a live view of your IT load and powertrain performance, which means you can make adjustments on the fly to prevent overloads and keep things running efficiently.

DCIM systems, especially when paired with AI, can help optimize energy usage. These systems can make automatic adjustments to keep your power usage in check, which reduces waste and cuts costs. And since Power Usage Effectiveness (PUE) is such a key metric, DCIM gives you the data to track it closely and make sure your energy management is dialed in.

Another big win with DCIM is its predictive capabilities. By looking at historical data, it can help you forecast future IT loads, power demands, and cooling needs. That’s crucial as AI-driven compute loads keep growing. The ability to predict and plan for future infrastructure needs is what’s going to keep your data center from running into problems down the line.

One of the most important things DCIM can do is** integrate with CMMS (Computerized Maintenance Management Systems)**. There should be some capabilities of integration between your building automation system and your computerized maintenance management system. When you connect DCIM with your maintenance systems, you get a ton of useful telemetry data. This lets you plan your maintenance activities more efficiently and stay ahead of potential issues, meaning fewer unexpected breakdowns and smoother operations overall.

Integrating CMMS and BMS doesn’t just generate more data—it creates enriched data that gives AI valuable context for decision-making. For instance, if a chiller is offline, an unconnected BMS might only register it as unavailable. But when BMS and CMMS systems are linked, an AI learns that it’s out for scheduled maintenance, not due to an operational issue - thus providing important context for the situation. This enriched data allows AI to interpret and respond to real-time conditions with greater accuracy, minimizing human intervention and supporting more efficient operations.

Right now, a lot of data centers still run with siloed systems. You’ve got one system for monitoring power, another for maintenance, and a whole bunch of manual inputs in between. That’s where things can get messy. What we really need is more integration—tying DCIM, BMS, and CMMS together so everything works in sync. That way, you get the full picture of what’s happening in your data center, and you can make smarter decisions based on real-time data.

In the end, integrating DCIM with other systems like BMS and CMMS is the way forward. Whether it’s predicting future loads or optimizing maintenance schedules, these tools help you stay ahead of the curve and keep your data center running efficiently. That’s what future-proofing is all about.

Liquid Cooling as a Future Strategy

As AI compute loads continue to grow and high-performance computing becomes more demanding, data centers will need to look beyond traditional airside cooling. While airside cooling has worked well up to now, it only gets you so far. Eventually, it won’t be enough to keep up with the heat generated by these heavy workloads. That’s when liquid cooling starts becoming an attractive option.

People are going to start considering how to start converting their data hall to liquid cooling because airside cooling only works great to a certain point. Liquid cooling can handle those high-performance computing loads better, but it’s not something you jump into lightly. There are a few things to keep in mind before you make the switch.

Cost and Labor: Converting a data hall to liquid cooling isn’t cheap. Installing liquid cooling systems takes time and expertise that not every facility is equipped for. You’ll need to make sure your team knows how to handle the new systems, so there’s a training aspect, too. You can’t just put liquid cooling in and expect it to run itself.
Structural Integrity: Next comes the question of whether your facility can handle the extra weight. Liquid cooling systems can add a significant load to your floors and racks, so you need to be sure your data hall is up to the task.

It’s like with electric vehicles and bridges—you’ve got to ask, are the structures built to handle that extra weight? The same goes for data centers. You have to make sure the structural integrity is there before you start adding all that extra weight to the pipes and water.

Moreover, there’s the risk of liquid cooling leaking that needs to be considered. Leaks can wreak havoc on an entire data center which can lead to equipment shorting out. This is detrimental to operations and adds cost to equipment needing to be replaced or repaired.
Long-Term Efficiency: Once it’s up and running, though, liquid cooling can be more efficient than traditional methods. Unlike airside cooling, where you’ve got filters to replace and regular maintenance to keep it going, liquid cooling systems can run with fewer interruptions. That said, it’s important to train your facility management staff on how to maintain these systems properly since they’re not used to dealing with water cooling technology.

Liquid cooling might become a more common strategy for data centers that need to handle massive AI workloads and high-performance computing. But like anything else, you have to weigh the benefits against the costs and risks.

Now what?

When it comes to future-proofing data centers, it’s not just about keeping the lights on—it’s about making sure your infrastructure can handle what’s coming down the road. With AI workloads increasing and high-performance computing pushing your systems harder than ever, you’ve got to be ready. That starts with being proactive.

AI isn’t just a buzzword—it’s a tool that can make your data center more efficient, from energy use to predictive maintenance. But it only works if your infrastructure is prepared. There’s untapped power sitting idle in your UPS systems that could be put to better use. By assessing your entire powertrain and managing UPS batteries smarter, you can scale up and avoid hitting capacity limits.

Moreover, running everything in silos just doesn’t cut it anymore. Integrating BMS, DCIM, and maintenance systems gives you a real-time, full view of your operations so you can make informed decisions when it matters most. And let’s not forget the importance of communication. Tenants and operators need to stay on the same page—especially with load changes. Setting up proactive systems helps you stay ahead of problems, instead of scrambling to fix them.

In the end, adaptability and forward-thinking are what will keep your data center ahead of the curve. The operators who assess, optimize, and integrate their systems now are the ones who’ll be ready for whatever’s next.

Hyperscalers: Sustainability goals vs. operational reality

Mike Sweeney — Tue, 08 Oct 2024 16:00:00 GMT

Hyperscalers—think Amazon, Google, Meta and Microsoft—are the giants behind the digital world. Their vast data centers keep our cloud-based service running smoothly, keeping our connected lives moving. As their influence grows, so does the scrutiny of their environmental impact. With the spotlight on sustainability, these tech titans have made bold promises.

They’ve set ambitious goals like net-zero emissions, running entirely on renewable energy, and even Microsoft’s unique pledge to remove all the carbon it has emitted since its founding by 2050. As impressive as these pledges sound, there’s often a gap between what’s promised and the realities on the ground. Though these promised goals are excellent and well-progressed, the rising demand for artificial intelligence applications and the enormous computing power required make these sustainability ambitions increasingly difficult to achieve.

In this article, you will learn:

How hyperscalers' sustainability commitments compare to their actual practices.
The innovative strategies hyperscalers use to improve energy efficiency.
The regulatory and public pressures urging hyperscalers to align ambition with reality.

Let’s take a closer look at the real story behind the promise and the practice.

Understanding Hyperscalers

What Are Hyperscalers?

Hyperscalers are the powerhouse companies behind today’s digital infrastructure. These giants have built a global network of data centers that power everything from your email to streaming services. What makes them “hyper” is their ability to scale quickly, handle massive workloads, and operate with remarkable efficiency, all on a global level.

These companies continue to grow larger each year and are committing massive amounts of capital expense to securing new land, power and equipment for larger and more complex data center builds. Hyperscalers stand apart not only because of the sheer size of the new data center capacity they have planned but are also investing in the latest cutting-edge technology to allow them to meet global demand in real-time with their fleets of existing data centers around the world. In fact, hyperscalers account for more than half of all cloud services consumed globally.

But they don’t operate alone. Colocation data center providers play a crucial role in supporting hyperscalers in meeting explosive demand. Colocation companies will either rent IT space that’s available within currently built data centers they own or will ‘build-to-suit’ vast new data centers that are pre-leased by a single hyperscale customer. Like hyperscalers, colocation providers design, build and manage their own physical infrastructure—space, power, and cooling.

Companies like AWS, Google Cloud, and Microsoft Azure are setting industry standards for scalability and efficiency, continuously pushing the boundaries of what’s possible in data center operations.

Hyperscalers vs. Colocation and Enterprise Data Centers

Data centers have long since evolved from the “shipping containers with servers” starting point. They come in different forms—hyperscalers, colocation providers, and enterprises—and each provides critical infrastructure.

Enterprise data centers started as a few servers in something as small as a utility closet. Today they may take up full floors of a building or operate in some built-to-suit space, but they are typically smaller and serve the specific needs of a single organization, like a bank or telecommunication company. These organizations must still manage some of their critical IT infrastructure in-house. Due to their legacy status and smaller scale, enterprise data centers tend to be reactive. They rely on human operators to manually adjust systems and respond to changes in demand, making them less adaptable to rapid scaling. Built in an earlier tech era, they are numerous but unable to be dynamic, operating within the constraints of their original design.

Colocation providers, also sometimes known as “retail” data centers, lease space for IT servers to either various companies in a shared facility or a single tenant. These providers own and manage the power, cooling, security and various other facility functions. Having been around for nearly 20 years, colocation providers also offer "build-to-suit" services, designing, building, and operating entire facilities for a single tenant’s needs—often hyperscalers. This model allows companies to expand their infrastructure without the expense and complexity of constructing and managing additional data centers.

Hyperscalers, on the other hand, are globally distributed networks that use advanced software, predictive algorithms, and the latest tech to manage everything from energy consumption to data storage. They both design, build, and own (or operate) their own data centers and contract with colocation providers for additional capacity. This modern, proactive approach combines elements of both enterprise and colocation data centers but scaled to a global level. As a result, hyperscalers can optimize resources and handle demand spikes more efficiently, achieving a level of efficiency and flexibility that’s difficult for other data centers to match.

What Makes Them Different?

Besides their massive scale, hyperscalers typically excel at or at least attempt automation, rapid scalability, and global reach. Unlike enterprise data centers, which typically serve a single region or purpose, hyperscalers can ramp up their operations across the globe in seconds.

Need more storage, processing power, or bandwidth? Hyperscalers can deliver it on demand with precision and speed, thanks to sophisticated algorithms and automation.

Efficiency is also a top priority. By operating on economies of scale that enterprise data centers can’t match, hyperscalers can offer services that maintain high levels of performance and reliability.

Why Hyperscalers Matter

Hyperscalers set the standard for global digital infrastructure. They’re essential to advancements in artificial intelligence, machine learning, edge computing, and 5G.

By providing the infrastructure that handles the immense flow of the world’s data, hyperscalers have the power to drive global standards for energy efficiency, sustainability, and operational best practices. Their products and services are among the most widely used in the world, meaning their data centers process enormous amounts of information. As a result, their operational choices don’t just impact their own businesses—they set benchmarks for the entire tech ecosystem, influencing colocation providers, suppliers and even governments.

In short, hyperscalers have the ability—and the responsibility—to lead the charge toward more sustainable, efficient, and innovative data center operations worldwide.

The Sustainability Commitments of Hyperscalers

Hyperscalers like Amazon, Microsoft, and Google have made ambitious sustainability promises. Amazon aims for net-zero carbon by 2040, Microsoft by 2030, and Google intends to run on 24/7 carbon-free energy by 2030. These commitments show the industry's shift toward sustainable operations.

While the goals are bold, turning them into reality is more complex. Hyperscalers are investing in renewable energy, energy-efficient hardware, and advanced cooling technologies such as liquid and immersion cooling.

However, despite these innovations, many data centers still rely on static control systems that don’t fully leverage AI’s potential for optimizing energy use. These static systems create inefficiencies by reacting too slowly to demand, leaving much of the promised energy savings unrealized.

AI could be a game-changer in bridging the gap between ambition and execution. With AI-driven controls, hyperscalers could manage energy use dynamically, preventing inefficiencies before they occur. Yet, many hyperscalers are not using AI to its full potential, continuing to rely on outdated systems that limit real-time optimization.

PUE – A Useful, But Incomplete Metric

Power Usage Effectiveness (PUE) has long been the gold standard for measuring data center energy efficiency. Introduced in the early 2000s, PUE calculates the ratio of total energy consumed by a data center to the energy used by its IT equipment. The closer the PUE value is to 1.0, the more energy-efficient the facility.

Figure 1: A chart explaining how PUE, Power Usage Effectiveness, is measured by comparing a data center's total energy consumption to the IT equipment energy consumption

As an example, if a data center is promoting a 20MW IT load capacity and runs the facility at 1.5 PUE, then the total maximum power this data center consumes is 30MW. This means 10MW of the energy is used for cooling, lighting, and other infrastructure needs when the data center is operating at 100% IT load capacity.

PUE offers a simple, comparable measure of performance across facilities, making it an important baseline for the industry. For years, it has provided a comparable measure of performance which urged data centers to find more efficient ways to optimize energy use.

However, while PUE is an essential starting point, it doesn’t tell the whole story. PUE captures only the energy dynamics inside the data center, primarily focusing on IT and mechanical power. As data centers adapt to industrial changes, there’s a growing recognition that factors, like water usage and carbon intensity, play a significant role in sustainability.

A Need for Additional Metrics Beyond PUE

PUE does account for innovations that directly impact energy consumption, such as AI-driven systems that optimize energy usage in real-time and liquid cooling that significantly reduces cooling energy. Yet, it doesn’t consider:

Water Usage: Some data centers have come under scrutiny on their water usage because they are in what is considered to be “water-scarce areas” but this will become a more important metric even in non-water-scarce areas as the amount of evaporative cooling needed for larger and more dense data centers continues to grow.
Energy Source: PUE doesn’t differentiate between renewable and non-renewable energy sources, which is critical for assessing carbon footprints.
Embedded Carbon: The building materials, equipment and transport needed to go from an empty plot of land to a fully functioning data center are not counted in the carbon footprint of the industry in any meaningful way just yet.

This focus on energy efficiency often overlooks these broader elements of a data center’s environmental impact. Additional metrics give a complete understanding with comprehensive measures. These metrics are:

Water Usage Efficiency (WUE): With the rise of liquid cooling systems and larger evaporative cooling data centers, measuring water consumption alongside energy efficiency is crucial. WUE helps track how much water a facility uses per unit of energy, reflecting its water sustainability practices.
Operational Lifecycle Costs: Metrics that factor in equipment lifespan, maintenance, and energy use provide a more holistic view of a data center's environmental and economic sustainability.
Carbon Intensity: This metric assesses carbon emissions related to the upstream supply chain for all materials, equipment and transportation used to construct the data center.

PUE Is A Strong Foundation, Not the Whole Solution

PUE serves the industry well and continues to be a valuable benchmark. Yet, as data centers grow more complex and sustainability goals become more ambitious, it’s clear that we need a broader set of metrics. By tracking water usage, lifecycle costs, and carbon intensity alongside PUE, hyperscalers can gain a more complete view of their environmental impact.

In short, PUE is a great starting point, but the industry must go further, leveraging additional metrics to truly measure and improve sustainability.

Discrepancies Between Sustainability Goals and Operational Reality

Hyperscalers have made bold sustainability commitments, but operational realities often tell a different story. While their goals for net-zero emissions and renewable energy adoption are ambitious, achieving them is complicated by inefficiencies and inconsistencies.

A major challenge is their reliance on static control systems, which react after demand spikes instead of adjusting in real-time. This creates inefficiencies, especially when hyperscalers hold colocation providers to stricter standards than they apply to their own facilities. For example, colocation partners are often required to meet stringent Power Usage Effectiveness (PUE) caps that hyperscalers themselves may not achieve.

Scope 3 emissions—those indirect emissions that occur in a hyperscaler’s supply chain—pose another issue. While hyperscalers aren’t directly responsible for the energy consumption of colocation providers, they have committed to reducing these emissions. This means they need their colocation partners to be more efficient. Achieving these reductions and closing the gap between sustainability goals and actual performance becomes increasingly challenging without dynamic, AI-driven controls, and better coordination.

Although hyperscalers like Google and Amazon have made strides in sustainability, such as upgrading to highly efficient equipment, recommissioning control programming to further optimize and investing heavily in renewable energy, misalignments with colocation providers persist.

Hyperscalers often enforce strict PUE caps on colocation partners, requiring them to maintain a PUE of 1.40 or lower to secure contracts. However, many colos may actually operate at a PUE closer to 1.45 or higher.

In these cases, colocation providers absorb the extra energy costs (and the related emissions) to meet the hyperscaler's requirements. This allows hyperscalers to claim they have a lower global PUE across their value chain without fully reflecting the reality of their colocation providers.

Moreover, hyperscalers often demand more conservative temperature controls in colocation facilities. A hyperscaler might run its own data halls at a maximum hot aisle temperature of 84°F but contract that colocation partners maintain a cooler 75°F. This demand forces colocation providers to consume more energy for cooling, undermining overall efficiency and sustainability goals.

This dual approach leads to fragmented sustainability efforts. Hyperscalers cap the PUE they’re willing to pay for while mandating more energy-intensive practices in colocation facilities. This contradicts their promotion of stringent efficiency standards, as the practices they mandate often drive inefficiencies elsewhere.

Additionally, there is little to no communication when IT load is expected to increase at a particular facility. This leaves collocation providers reacting to IT load spikes rather than proactively preparing for them—a typically more efficient way of handling the cooling needs.

For example, a colocation provider may be contractually required to hit a PUE of 1.35, which the hyperscaler then includes in its reported sustainability metrics. In reality, the colocation provider may operate at a higher PUE, such as 1.45, absorbing the additional costs to avoid breaching the contract. This practice allows hyperscalers to base their efficiency claims on contracted targets, ignoring the true energy consumption within their supply chain.

This lack of transparency creates friction across the data center ecosystem. Static systems prevent hyperscalers from fully realizing energy savings, while colocation partners bear the burden of unrealistic demands. This hampers sustainability progress and risks damaging hyperscalers' credibility as environmental leaders.

Why It Matters

Hyperscalers influence the global data center industry, and when they fail to meet their own standards, it weakens the broader sustainability movement. Every inefficiency adds to energy waste, emissions, and missed opportunities for improvement.

To maintain public trust and industry leadership, hyperscalers must align their operational practices with their sustainability promises. As regulatory scrutiny tightens and public expectations grow, bridging this gap is critical for their long-term success.

Innovations in Energy Efficiency and Sustainability

Hyperscalers are at the forefront of energy efficiency through advanced cooling systems, AI-powered automation, and renewable energy integration, closing the gap between sustainability goals and real-world operations.

Cooling is a major challenge for hyperscalers, but immersion and liquid cooling are helping to significantly reduce energy use.

Immersion cooling submerges hardware in liquids to dissipate heat more efficiently than traditional air-cooling methods.
Liquid cooling circulates coolant directly around the hottest components, minimizing wasted energy.

AI-empowered anticipatory cooling takes things further. It anticipates and adjusts cooling needs before temperature spikes occur. This ensures cooling systems activate only when needed, reducing unnecessary energy consumption. Moreover, AI can target location-specific cooling, focusing on particular sections of the data hall where IT load is ramping up, rather than cooling the entire space. This precision reduces unnecessary energy consumption and optimizes overall efficiency.

AI is becoming a game-changer in data center management. By dynamically adjusting cooling, balancing workloads, and anticipating energy spikes, AI optimizes data center performance in real-time.

Think of it like "cruise control" for data centers. AI would automatically make adjustments based on demand before problems arise. Like a smart thermostat, it senses when a room's temperature is rising and takes corrective action in a precise place before discomfort even registers.

While data centers are reaping the benefits of the AI boom, many have yet to fully embrace AI for enhancing their operational efficiency. This gap leaves significant room for improvement in energy optimization and sustainability.

Hyperscalers, however, are investing in energy-efficient hardware to cut down on power consumption. Their servers and modular data centers are designed to perform at high levels while being light on energy usage, helping to reduce energy footprints while boosting performance.

Perhaps one of the most transformative steps hyperscalers are taking is their commitment to renewable energy. Amazon and Google are among the largest corporate buyers of renewable energy, with Google matching 100% of its energy usage to renewables since 2017. Some hyperscalers are beginning to design in on-site energy storage or flexibility to use renewable energy more effectively. They also are contracting even more new production of solar and wind farms, near their data centers. This helps reduce their reliance on the grid while offering long-term energy cost control.

By taking control of their own green energy production, hyperscalers are setting new sustainability benchmarks and working toward meeting their ambitious goals as energy demands continue to rise.

Environmental Impact and Mitigation Efforts

Hyperscalers and their data centers come with a heavy environmental cost. In 2022, data centers consumed an estimated 460 terawatt-hours (TWh) of electricity, accounting for nearly 2% of global electricity use, a figure set to double by 2026. Despite significant renewable energy commitments, hyperscalers still face challenges in reducing their carbon footprints, especially when it comes to managing fluctuating energy demands.

So, the question remains: Is it enough? While renewable energy helps offset some emissions, the sheer power demands of these data centers often outpace available green resources. Simply purchasing renewable energy credits doesn’t fully address the immediate impact of their operations, especially in regions where non-renewable energy remains dominant.

Hyperscalers have made strides, but inefficiencies persist. Many data centers still rely on static control systems, missing out on potential energy savings. Additionally, Scope 3 emissions—indirect emissions from their supply chain—are difficult to track accurately due to contracting issues and therefore hard to mitigate, complicating sustainability efforts further.

The Room for Improvement

AI has the potential to forecast energy spikes, adjust cooling systems before problems arise, and optimize server workloads for greater energy efficiency. But the adoption has been slow and uneven. Many hyperscalers are still relying on outdated systems that aren’t designed to optimize energy use dynamically.

Real-time dynamic control could be the key to unlocking the full potential of their sustainability initiatives.

The gap between hyperscalers’ sustainability goals and their current operations undermines their green pledges and erodes public trust. As regulatory scrutiny grows, hyperscalers must prove their progress to avoid tighter regulations and potential penalties. Leading the charge with AI-driven systems and transparent communication could set a new sustainability standard across the industry.

The Influence of Regulatory and Public Pressure

As hyperscalers expand, they face growing regulatory scrutiny, especially in regions like the European Union (EU). The EU's new reporting requirements mandate data centers to disclose energy efficiency, renewable energy usage, and water consumption, aiming to cut energy use by 11.7% by 2030. These regulations push hyperscalers toward greater transparency, especially around complex environmental issues.

The EU often sets the standard for global regulations, meaning companies must adapt as similar demands emerge worldwide. Failure to meet these expectations could result in operational challenges and reputational damage.

As public awareness of data centers’ environmental impact grows, so too does the demand for accountability. Both regulators and consumers are seeking real proof of sustainable operations—not just promises.

In addition to regulatory pressure, public expectations are becoming a driving force. While companies are quick to market their eco-friendly credentials, this positioning carries risks. Over-promising and under-delivering on sustainability goals leaves them vulnerable to accusations of greenwashing.

Public trust is fragile, and increasingly more important as data center colocation providers and hyperscalers are looking for more land and more power and must deal with more and more community stakeholders concerned about their presence in or near their community. Any perception that overstate green efforts could result in lost business and damaged reputations.

Despite these challenges, hyperscalers are in a unique position to drive sustainability standards across the data center industry. With their scale, they influence colocation partners and suppliers, pushing for best practices. However, they must ensure that their internal operations meet the same high standards they demand from others. As regulatory pressure and public scrutiny increase, hyperscalers have a chance to lead by example and drive meaningful sustainability progress.

What Comes Next?

Hyperscalers are undeniably shaping the future of the digital world. Their bold sustainability commitments, from net-zero emissions to renewable energy targets, lead the charge in how data centers should operate. While innovations in AI, advanced cooling, and renewable energy are helping to bridge the gap, static control systems and simplified metrics like PUE limit the full potential of these efforts.

The future of sustainability in data centers depends on adopting more comprehensive metrics that go beyond energy efficiency. AI and automation will be pivotal in driving real-time optimization, allowing hyperscalers to dynamically manage their resources, reduce carbon footprints, and meet their ambitious goals.

As hyperscalers continue to grow, holding them accountable to their green promises is critical. By embracing emerging technologies and more holistic approaches, they can truly deliver on their sustainability goals and set a new standard for the industry.