Executive Summary
- For the past two years, gigawatt-scale campuses have dominated the AI infrastructure story, with the biggest power deals and construction programmes dominating the headlines.
- But, the growth of inference is beginning to challenge the assumption that bigger is always better.
- Here, Uri Zilberman, VP Technology Partnerships, ZutaCore explores and explains how we measure the next phase of AI infrastructure.
OpenAI and Anthropic have actively pursued interest in 20–30MW capacity deals across the US, UK and Nordics, alongside their larger investments, is an early indication of what that shift could look like. These discussions, which are understood to concern inference workloads, suggest AI infrastructure will increasingly need to scale out as well as up.
Training large AI models favours concentrated compute clusters, whereas inference serves live requests, making geography more important for some workloads. Regional capacity can reduce latency where response times are critical and help meet data residency requirements.
A wider network can improve resilience, provided workloads can be redirected when disruption occurs. As AI infrastructure is increasingly being treated as critical infrastructure and concentrated compute sites are viewed as potential strategic targets, that resilience is becoming more important. Realising these benefits, however, depends on how quickly regional facilities can accommodate high-density computing hardware.
The cooling constraints of regional deployment
High-density AI racks generate substantial heat whether they sit in a gigawatt campus or a smaller regional facility. Purpose-built sites can incorporate cooling infrastructure from the outset, whereas existing colocation facilities may have limited room for additional equipment and tighter power allocations. A thermal design developed for a flagship campus cannot always be transferred without modification.
Power and water constraints also vary between locations. In September 2026, California signed seven data centre bills addressing water-use disclosures and electricity costs, illustrating how local requirements can affect deployment plans. Other facilities may lack the heat rejection capacity needed for high-density AI racks. If cooling has to be redesigned at every site, much of the speed advantage of regional capacity disappears.
Every kilowatt counts
Under a fixed power allocation, every kilowatt spent on cooling is a kilowatt unavailable to computing. As rack densities rise, cooling efficiency therefore becomes a direct factor in how much AI compute an operator can deploy from a given facility. Large mechanical installations can also occupy space that could otherwise accommodate racks, limiting the GPU capacity an operator can deliver from each site.
A common rack-level cooling architecture can help address these constraints. Standardised components and connections reduce the engineering required at individual sites, allowing operators to adapt the system to local infrastructure without redesigning the thermal solution from scratch.
Cooling infrastructure built to repeat
Waterless, two-phase direct-to-chip cooling offers one way to achieve this consistency. A dielectric fluid absorbs heat and changes phase within cold plates mounted on processors, before condensing and circulating through a sealed loop. This process transfers substantial thermal loads using relatively small fluid volumes, reducing pumping requirements and the space occupied by supporting equipment.
The same core architecture can support deployments of different sizes without fundamentally changing how heat is removed from the chips. Operators can establish common performance specifications and rack configurations, making successive installations more predictable while retaining flexibility over how each facility rejects heat.
Waterless cooling at chip level removes the need for water within the server cooling loop, providing a consistent thermal architecture that can be deployed across different locations. Heat rejection can then be adapted to local conditions, including approaches that reduce or eliminate water use at facility level where required. This separation gives operators greater flexibility while maintaining a common cooling design.
Performance on repeat
Repeatability must extend beyond thermal performance to day-to-day maintenance. Regional facilities may have fewer specialist engineers on-site, making clear servicing procedures and reliable fault management essential. Features such as non-conductive dielectric fluids can help reduce operational risk by limiting the impact of accidental leaks on IT hardware. Equipment that requires extensive modifications or specialist intervention at every location can quickly undermine the advantages of standardisation.
Consistent multi-site deployment therefore relies as much on partnerships as on engineering. Operators need a repeatable cooling architecture that can be deployed across sites with minimal redesign, from the cold plate through to heat rejection. Server manufacturers need to validate cooling systems within standard rack designs, while colocation operators must establish clear requirements for connecting those systems to existing infrastructure. Regional integrators need the expertise to commission and maintain installations to common specifications.
But, make it bespoke
Establishing these relationships early reduces repeated qualification work and gives operators greater confidence in deployment schedules. Even a proven cooling architecture can become a bespoke engineering project if the partners responsible for installation are working to different specifications.
Gigawatt-scale campuses will remain essential to AI development but distributed inference is creating a parallel requirement for smaller facilities that can be brought online quickly. Operators increasingly need to scale out as well as up, without restarting the design process every time capacity is added in a new location.
The next phase of AI infrastructure will not be measured only by how big operators can build, but by how quickly they can turn regional capacity into reliable compute.
