AI is changing the data center at a pace few infrastructure transitions have matched.
Much of the industry conversation has understandably focused on GPUs, power availability, higher rack densities and liquid cooling. But at Datacloud USA, our panel, Operating in the Age of AI: Can Data Center Operations Keep Up?, focused on an equally important question: once this infrastructure is deployed, are our operating models ready for it?
The discussion brought together perspectives from across the data center ecosystem and different geographic markets. What became clear is that AI infrastructure is not simply creating a new engineering challenge. It is changing the responsibilities, skills, processes and performance expectations required to operate mission-critical facilities.
For Salute, the conclusion is straightforward: AI infrastructure requires an AI-ready operating model.
Higher density changes more than cooling
The progression toward increasingly dense AI infrastructure is well understood. What receives less attention is what happens operationally when liquid cooling becomes an integral part of the production environment.
Direct-to-chip liquid cooling introduces systems and operating conditions that traditional data center teams have not historically had to manage at this scale.
Operators now need to understand coolant distribution units, facility water systems, filtration, coolant and waterchemistry, flow rates, differential pressure, temperature and leak detection alongside the established disciplines of electrical and mechanical infrastructure management. Even seemingly routine activities can become significant operational processes.
During the Datacloud discussion, one example highlighted the work required simply to prepare liquid cooling loops for production. Flushing, filtration and achieving the required cleanliness can require sustained operational attention before equipment is ready to run.
This is an important distinction. Deploying liquid cooling is an engineering project. Operating liquid cooling reliably is a lifecycle capability.
Organizations planning AI infrastructure therefore need to think beyond whether their facility can technically support the equipment. They need to determine whether their people, procedures, monitoring systems, safety protocols and escalation models are prepared to operate it continuously.
The new operational question: Where does responsibility begin and end?
One of the most engaging areas of the Datacloud discussion concerned operational demarcation. As AI infrastructure becomes more complex, the traditional line between the data center operator, customer and technology provider is shifting.
In a liquid-cooled environment, that creates immediate operational questions. Who owns and operates the coolant distribution unit? Who manages filtration and coolant chemistry? Who responds to a leak? Who owns the connection between the facility water system and the technology loop? And, critically, how far should the data center operations team go when an incident reaches the cabinet? These cannot be questions answered for the first time during an incident.
As John Shultz, Chief Product and Learning Officer at Salute, highlighted during the discussion, Salute has mapped operational demarcation models through work and interviews with more than 25 organizations. What emerges is not one universal model, but three broad levels of operational responsibility.
- Type 1 places the demarcation at the CDU. The data center provider manages the facility loop to the CDU connection, while the customer retains responsibility for the technology loop, CDUs, chemistry, racks and servers. This limits the operator’s responsibility but requires the customer to maintain the operational capability and 24×7 support necessary to manage the technology side.
- Type 2 extends data center responsibility to the cabinet connector. The operator manages the facility and technology loops, including CDUs and coolant chemistry, while the customer remains responsible for the cabinets, racks and servers. It reduces the operational burden on the customer, but requires greater staffing, training and technical capability from the data center operations team.
- Type 3 goes further, enabling end-to-end management of the liquid infrastructure and operational response up to the cabinet, while stopping short of server intervention. In a leak event, for example, the data center team can isolate the affected cabinet, shutting down flow and power where required, without opening or servicing customer IT equipment. John described precisely this distinction during the panel preparation: the operator can isolate the cabinet for the customer, while intervention inside OEM equipment remains outside its responsibility.
The significance of these models goes beyond liquid cooling. They illustrate how AI is changing the operating contract between infrastructure providers and their customers. There may never be a single demarcation model for every AI deployment. Facility design, customer requirements, technology architecture, OEM requirements and contractual obligations will continue to determine where the line sits. Other panelists reinforced that point, noting that the demarcation can change according to the individual customer contract and who chooses to own and manage the CDU.
What must become standard is clarity of accountability. A demarcation point should not simply define where one company’s equipment ends and another’s begins. It should establish who monitors, who maintains, who responds, who has authority to isolate equipment and what happens when an incident crosses that boundary. This is critical not only for clear responsibilities but also for ensuring uptime, mitigating serious operational risks and protecting the safety of everyone in the data center.
For AI infrastructure, that operational clarity is becoming as important as the physical design itself.
AI infrastructure requires a different workforce model
Technology is only one side of operational readiness. The other is whether the industry can build a workforce at the speed AI infrastructure is being deployed.
The Datacloud discussion made clear that this is not simply a question of filling today’s vacancies. Higher-density and liquid-cooled environments are fundamentally changing the skills required inside the facility at the same time as the industry needs to attract an entirely new generation of talent.
That creates two related challenges: upskilling the workforce we have and building the workforce we will need.
The first requires structured technical development. Operators increasingly need people who understand liquid cooling, coolant distribution units, coolant and water chemistry, filtration, leak response and the interaction between facility and technology infrastructure. As operational responsibility extends further towards the cabinet, the workforce implications become greater. The demarcation model and workforce model are therefore directly connected.
But the panel also challenged the industry to think further ahead.
Phillip Marangella of EdgeConneX spoke about the importance of reaching young people before they enter the workforce and raising awareness of the opportunities created by data centers and AI infrastructure. His work to establish an AI Academy and educate students is a practical example of how the industry can begin building that future talent pipeline rather than simply competing for the people already working within it.
That distinction matters. The skills challenge cannot be solved through recruitment alone.
Many students interact with AI and digital services every day without necessarily understanding the physical infrastructure behind them, or the breadth of engineering, operations, technology and facility management careers available within the data center sector. Creating educational pathways and giving students direct exposure to the industry can help close that awareness gap.
For Salute, this is particularly important. The answer to the AI skills challenge cannot be to compete repeatedly for the same limited pool of experienced data center professionals. We need multiple pathways into data center operations, combined with training models that turn potential into operational capability.
At the same time, SOPs, MOPs and EOPs must evolve alongside the infrastructure. Operational knowledge needs to be documented, repeatable and transferable so that new and existing employees can develop the competencies required to operate increasingly complex AI environments safely and consistently.
Global scale will not mean identical operations
Another important theme from the Datacloud panel was geography. AI infrastructure may be global, but data centers operate within local power markets, climates, water environments, regulatory structures, labor markets and communities.
The operating model therefore needs enough standardization to create consistency while retaining enough flexibility to respond to local conditions. The discussion highlighted how priorities can differ between markets such as North America and South America, particularly around renewable energy, cooling ecosystems and community expectations.
This balance between global standards and local execution will become increasingly important as AI capacity expands into new markets. Operators need repeatable processes. But repeatability should not be confused with uniformity.
Community must be part of the infrastructure strategy
The panel also reinforced that operational readiness cannot be considered independently of the communities in which data centers operate.
Alessandro Lombardi of Elea Data Centers brought an important Brazilian perspective to the discussion. Developing data center infrastructure in a market such as Brazil, and particularly in areas associated with the Amazon region, creates a very different set of environmental and community considerations from those encountered in many established North American data center markets.
The discussion highlighted the importance of renewable energy, cooling, sustainability and, critically, bringing local communities into the conversation around new infrastructure development. In environmentally sensitive regions, operators and developers need to be able to explain not only what they are building, but how it will use local resources and what responsible development looks like in that specific location.
That is an increasingly important lesson for a global data center industry. There is no universal community engagement model. A project in Houston, Texas or northern Brazil exists within a different environmental, economic and social context. Successful development requires operators to understand those differences and engage accordingly.
Community engagement therefore cannot begin once a facility has been designed or when approval is required. It needs to form part of the infrastructure strategy from the outset. As AI drives larger facilities and greater demand for power, water, land and skilled people, the industry’s license to grow will increasingly depend on its ability to demonstrate that growth can create value for the communities hosting that infrastructure.
As for the supply chain and the entire ecosystem – who is at risk, who is ready?
The data center industry has become highly effective at solving difficult infrastructure problems.
AI is asking it to do so again, at unprecedented speed. Power will remain a constraint. Supply chains will remain a constraint. Cooling technology will continue to evolve. But another constraint deserves equal attention: our ability to operate what we are building.
AI-ready data centers require more than AI-ready infrastructure. They require clear operational demarcation, new performance metrics, documented procedures, trained teams and operating models capable of evolving as rapidly as the technology itself. The organizations that address those requirements during design and deployment, rather than after commissioning, will be better positioned to scale AI infrastructure safely and reliably. The question coming out of Datacloud is therefore not simply whether data center operations can keep up with AI. It is whether we are prepared to redesign operations for the infrastructure that comes next.