Executive Summary

• Alon Tal, VP Corporate Development and Investor Relations at ZutaCore pens an article on how power pressure is reshaping AI infrastructure decisions; AI infrastructure is already in a power-constrained environment, but with the recent geopolitical tension, it has sharpened the grid access challenge.
• With electricity consumption from data centres set to double by 2030, energy planning has become central to AI infrastructure strategy.
• In a market this power-constrained, every wasted watt carries a cost; the next phase of buildout will depend on how much useful compute operators can draw from the energy available to them.

Access to power is now one of the main constraints shaping AI infrastructure. Demand for compute has not softened and pressure to add capacity remains intense, but power is becoming harder to secure.

That pressure was already evident before the latest escalation in the Middle East. AI infrastructure was operating in a power-constrained environment, with grid access influencing where capacity could be built and how quickly. Recent geopolitical tension has sharpened that challenge, adding uncertainty to energy markets and reinforcing how closely power availability is tied to AI buildout plans.

At the same time, AI workloads are changing the demands placed on the facility around the server. Higher-density chips use more power and generate more heat, increasing pressure on cooling infrastructure. Power delivery and heat removal now need to be planned together, with weakness in either area slowing deployment or weakening the commercial case.

Power is setting the pace

Data centre growth was once framed around land, fibre and connectivity, but power now leads the conversation. A site that cannot secure enough megawatts may never move beyond the planning stage. And even when power is available, excess overhead can erode the economics AI customers expect.

Projected demand explains why this issue has become so urgent. The International Energy Agency has said electricity consumption from data centres is set to double by 2030, with power use from AI-focused data centres expected to rise even faster. Goldman Sachs Research has forecast that data centre power demand could increase by 165% by the end of the decade compared with 2023.

Together, those figures point to a market where energy planning has become central to AI infrastructure strategy. Cost and energy input remain core considerations, but recent events have made the issue more immediate. Increasingly, the question is how much compute a facility can deliver for the money and power it consumes.

Power usage effectiveness (PUE) remains a useful measure of that performance because it shows how much power reaches IT equipment and how much is lost to supporting systems. The closer a facility gets to using its energy for compute rather than overhead, the stronger its operating position becomes. When energy prices move quickly, that efficiency directly affects total cost of ownership and long-term resilience.

Cooling belongs in the power equation

This is why cooling now belongs in the same planning conversation as grid access. Securing power is only part of the challenge; a data centre also has to remove the heat high-density chips create. If the cooling approach adds too much energy overhead or limits rack density, it reduces the useful compute the site can support.

That challenge becomes harder as chip densities rise. Heat inside AI servers is increasingly difficult to remove with conventional air-based systems alone, and for the highest-powered silicon, direct-to-chip liquid cooling is no longer simply an option but a requirement. By removing heat closer to the processor, it can support denser configurations and reduce energy overhead. Its value is commercial as much as technical, because better thermal management helps power go further and improves total cost of ownership.

Once liquid cooling becomes part of the answer, the choice of architecture matters. Water-based systems can handle higher heat loads than air, but they also raise concerns about leakage, corrosion and upkeep, as well as local water use. In AI environments where uptime and hardware protection are central to the business case, those trade-offs are becoming harder to ignore.

For the highest-powered silicon, waterless, two-phase, direct-to-chip cooling is the architecture that can meet the thermal and operational challenge at scale. Using a dielectric fluid at chip level allows heat to be removed efficiently while reducing risk near sensitive electronics. It also supports higher-density AI infrastructure, helping operators make better use of constrained power and reduce dependence on local water resources.

Resilience is also becoming part of the same calculation. Location decisions have always involved some assessment of regional stability, but recent geopolitical events have made that concern more immediate. As a result, supplier diversity and site strategy are likely to carry more weight, especially for facilities exposed to disruption or energy market volatility.

None of this means demand for AI infrastructure will fall away. The need for compute is too strong and the strategic importance of AI capacity is too high. Customers will still focus on the quality of the solution and the reliability of the path to scale, but they will do so with a sharper view of operational risk.

That shift will also change how technology providers are assessed. Buyers will look for evidence that systems can be delivered and scaled in conditions where power availability and supply chain risk are under pressure. In high-density AI environments, capacity promised on paper has little value unless the infrastructure can support live workloads and turn constrained power into usable compute.

AI infrastructure will continue to expand because demand has not gone away. The real question is whether facilities can secure enough power and use it efficiently enough to scale through uncertainty. In a power-constrained market, every wasted watt carries a cost, and the next phase of buildout will depend on how much useful compute operators can draw from the energy available.