The AI infrastructure boom is colliding with the physical realities of power, cooling, and utility capacity. As operators deploy larger AI clusters and increase rack densities, long-standing assumptions about where workloads run and how infrastructure is designed are changing.
Amid these sweeping changes, utility interconnection timelines, power distribution equipment, and facility cooling have become crucial planning variables. Communities are also scrutinizing water consumption tied to new developments. At the same time, organizations must show that AI investments deliver business value, not just higher resource use.
Those pressures are reshaping customer behavior, according to Andrew DesRochers, principal technologist for sustainable transformation at HPE. In a conversation at HPE Discover 2026, he described a shift from sustainability-focused discussions to operational efficiency, arguing that energy availability, cooling infrastructure, resource utilization, and utility constraints are increasingly steering IT decisions as AI moves from experimentation to production.
Data Center Knowledge spoke with DesRochers about changing customer priorities, AI infrastructure constraints, cooling challenges, and what operators are learning as they deploy AI at scale.
The following interview has been lightly edited for clarity and length.
Data Center Knowledge: What has changed most about customer conversations over the last two years?
Andrew DesRochers: One of the biggest changes is that IT operators are starting to see how critical energy efficiency is within their own operations. Historically, many IT teams assumed energy would be available whenever they needed to deploy infrastructure and that facilities teams would handle the grid connection and power requirements.
Today, with rising energy costs and constraints around power availability in some regions, energy has become a major factor in infrastructure planning. Energy is in the conversation now. Customers are asking how they can consume less energy and get the maximum amount of IT work from the energy they deploy.
DCK: Has the conversation shifted from sustainability to operational concerns?
AD: Absolutely. A few years ago, sustainability was the term customers used. Today, execution is the focus. We had a customer tell us they haven’t heard the word sustainability in years.
The challenge is that efficiency still matters. If AI workloads consume excessive energy or resources, organizations will struggle to demonstrate a positive return on investment. Efficiency has to be embedded into an AI strategy because it’s directly connected to business outcomes.
DCK: Are customers running into power constraints, cooling constraints, or utility constraints first?
AD: It depends on the region. What we’re seeing broadly is increased scrutiny around data center resource consumption. Energy remains the primary concern today, but water is emerging as the next major issue. In some regions, we’re seeing community pushback and greater attention on water consumption associated with new data center developments.
One challenge is helping people understand where resources are actually being used. Direct liquid cooling, for example, is often assumed to be the primary source of water consumption, but those systems are typically self-contained. Operators need to look beyond individual technologies and evaluate the full system, including facility operations and impacts beyond the walls of the data center.
DCK: How often are utility timelines shaping infrastructure decisions?
AD: More and more. Customers are increasingly looking at what they can realistically deploy within utility timelines and within the availability of infrastructure. We’re even seeing discussions emerge around technologies like high-voltage DC as organizations evaluate every available option to meet future power requirements.
One interesting development is that power conversion equipment and supporting electrical infrastructure are becoming significant factors in deployment planning. Those timelines are changing rapidly.
DCK: Are there assumptions about AI-driven power demand that have turned out to be wrong?
AD: It’s important to separate large-scale AI training facilities from typical enterprise AI deployments.
Most enterprise customers are not training foundation models from scratch. They’re deploying inference workloads and using smaller or distilled models wherever possible. Those workloads have very different infrastructure requirements than the massive training clusters that tend to dominate public discussions.
Not all AI workloads should be treated the same. Training environments and enterprise inference environments have very different power and efficiency profiles.
DCK: What surprises customers most after deployment?
AD: One recurring theme is the importance of measurement and analytics. Organizations deploying the latest hardware need visibility into power consumption, cooling requirements, and overall efficiency. Without that data, it’s difficult to optimize operations.
We’re also seeing opportunities through relatively simple changes. In some cases, moving from older equipment to newer systems with larger, more efficient fans can significantly reduce cooling-related energy consumption. The key is having the data needed to identify those opportunities.
DCK: GPU efficiency and utilization continue to be major topics. What are you seeing?
AD: Utilization matters tremendously. We’ve even seen this internally. In one project, we discovered systems were running in performance modes that weren’t necessary for the workload. Simply moving to more energy-efficient operating modes significantly reduced energy consumption without affecting outcomes.
The lesson is straightforward: organizations need to ensure they’re getting the most useful work from the infrastructure they’re deploying. Efficiency starts with understanding how systems are actually being used.
DCK: Are operators underestimating any cooling challenges?
AD: Many organizations are taking a broader look at facility-level cooling infrastructure. Customers increasingly recognize that cooling efficiency isn’t only about the server itself. It’s also about the building, facility systems, and how everything works together. That creates opportunities to reduce energy consumption beyond the IT equipment.
DCK: What rack densities are becoming common enough that they are no longer edge cases?
AD: Industry averages continue to move upward.
The discussion increasingly centers around when liquid cooling becomes necessary. While many organizations still want to maximize the life of air-cooled infrastructure, higher-density deployments continue to push operators toward liquid-cooling solutions.
It’s also important to remember that cooling isn’t limited to compute. Networking and storage increasingly factor into these discussions as well.
DCK: Is water availability becoming a gating factor?
AD: Water consumption is definitely receiving more scrutiny. At the same time, we’re seeing some organizations become more flexible about where workloads run. In certain AI use cases, users are willing to accept slightly higher latency if it allows operators to take advantage of cooler climates or more efficient facilities.
That creates opportunities to think differently about site selection and resource optimization.
DCK: Are customers changing site-selection strategies because of water concerns?
AD: Water is becoming a larger consideration, particularly in regions where resources are already constrained.
We’re also seeing growing interest in waterless cooling technologies. Historically, those approaches carried a significant cost premium, but the economics have improved considerably. As communities and regulators pay closer attention to water consumption, those alternatives become more attractive.
DCK: What operational behavior is changing most among AI infrastructure operators?
AD: Customers are becoming more deliberate. Early on, there was a tendency to pursue generative AI because everyone else was. Today, organizations are stepping back to ask whether AI is the right tool for a specific problem and which type of AI deployment best fits their business objectives.
We’re seeing more focus on practical applications and measurable outcomes.
DCK: Is there a design decision customers regret from the last few years?
AD: I wouldn’t characterize it as regret. Many organizations experimented with AI and learned valuable lessons. Now they’re reassessing strategy, focusing on adoption, and thinking carefully about how efficiency supports business value.
One challenge is ensuring employees understand how AI fits into their work. Adoption ultimately plays a major role in determining whether organizations achieve meaningful returns from AI investments.
DCK: When we revisit this conversation in two years, what will have changed?
AD: I hope efficiency becomes something organizations simply do by default. My hope is that energy, cooling, and resource efficiency become standard design considerations rather than topics that require constant education. There’s still a lot of work to do, but that’s where we would like the industry to be.
Efficiency should become a normal part of how organizations design, deploy, and operate infrastructure.
