.png){kind=logo}
The AI infrastructure race gets measured in GPUs. Operators compete for Nvidia accelerator allocations, developers encounter waitlists for high-end clusters, and customers compare hourly rental rates across cloud providers. However, buying accelerators only solves one part of the deployment puzzle. Dense GPU racks require substantial electrical capacity, cooling equipment sized for sustained heat loads, high-capacity networking, and a site capable of supporting all of those systems at once.
Hardware also moves faster than the infrastructure around it. Accelerators can lose their relative performance advantage within a few product generations; substations, fiber routes, cooling systems, and purpose-built data halls can support successive fleets of machines.
Grid connections often move on an even slower timetable, creating a metric that dictates the industry: time to power. The wait between requesting grid capacity and receiving it can stretch from two to 10 years across the EU, according to the International Energy Agency. When grid connection takes this long, a site that secures deployable power in 18 months decisively outranks the financial appeal of a cheaper, delayed parcel.
Securing thousands of accelerators has limited value when deployable electrical capacity arrives years after the hardware. Cooling presents another power problem. Every watt of energy spent moving heat is a watt stripped from compute, meaning local ambient climate directly governs how much usable AI capacity a given megawatt actually delivers.
For Clichmont, an ecosystem designed for scalable AI infrastructure, those physical constraints shape the operating model. The company owns and operates its data centers and GPU infrastructure, which places power, cooling, hardware configuration, and deployment under the same organization rather than separating facility capacity from the compute running inside it.
Ownership also shifts the planning horizon. Site upgrades and GPU deployments can be coordinated against projected customer demand, while the surrounding electrical and mechanical infrastructure can continue serving newer accelerator generations as hardware changes.
Clichmont ranks available power as its leading site-selection criterion, followed by the time required to deliver that capacity. Cooling conditions and fiber connectivity come next; land cost follows all four. A low-cost parcel loses much of its economic advantage when energization trails construction by several years.
Major European data center hubs illustrate the problem. Connection queues in Frankfurt, London, Amsterdam, Paris, and Dublin average seven to 10 years. For projects that require large blocks of power, a credible energization date can matter as much as the nominal capacity available on the grid.
Alicante puts this site-first philosophy into practice. Here, a high ambient temperature requires robust mechanical cooling, but strong solar yield allows for 1.8 MWp of on-site generation. Construction is scheduled to begin May 1, 2027.
Clichmont expects the array to produce approximately 2.9-3.1 GWh per year. Against an initial planned IT load of about 331.5 kW, or roughly 398 kW at site level with a projected PUE of 1.2, annual solar output would equal around 83-89% of the facility’s electricity consumption on an energy basis.
However, because solar output is intermittent, Clichmont plans to combine this generation with battery storage and grid supply, shifting surplus power into non-solar hours to reduce overall dependence on utility delivery schedules.
Bodø, Norway, provides the exact counter-example. A naturally cool ambient climate and regional energy profile lower the cooling load per unit of compute, allowing for a different density ceiling and a distinct architectural approach.
For Bodø’s next 500-square-meter phase, construction, power infrastructure, cooling, connectivity, security, and fire-protection systems must be completed before any new hardware is installed. Clichmont has not yet published final megawatt or rack-capacity figures because the engineering configuration is still under development.
The contrast between the two sites points to a hard operational reality. The same four criteria, applied honestly, produce entirely different engineering answers. This is what designing around a site actually means, and it dictates the hardware timeline. Procurement follows site readiness, instead of leading it.
The company already operates GPU infrastructure at Billingstad and Bodø Harbor in Norway. Billingstad houses 12 deployed GPUs, while Bodø Harbor has 20, with another 40 RTX 5090 units supplied by Cybertek in the deployment pipeline. The larger purpose-built Bodø facility of roughly 500 square meters targets Q4 2027 for operation.
AI infrastructure operators compete for grid capacity, suitable sites, cooling capability, and network access alongside accelerator allocations. Those constraints function on different schedules, and the longest lead time can determine when a cluster starts generating usable compute.
Clichmont’s approach puts power procurement and site engineering inside the same planning model as GPU deployment. The company intends to track utilization of deployed compute as its primary operating metric, alongside contracted capacity and cost per unit of compute, which shifts attention from announced hardware toward how much infrastructure customers actually use.
Accelerator availability will keep changing with each hardware cycle. Grid connections, substations, cooling plants, and fiber routes move more slowly. Infrastructure built around deliverable power can therefore outlast several generations of GPUs, making time to power one of the numbers that ultimately determines where AI compute gets built.
