Sovereign AI and settlement workloads run hot and dense. We match the cooling method to rack density and local climate, so a nation's compute stays inside its thermal envelope and inside its borders.
The right method depends on how much heat a rack rejects, and dense AI training racks demand a fundamentally different approach than general compute.
General-purpose and settlement compute at moderate density is served efficiently by optimized air cooling. Hot-aisle containment and tuned airflow handle these loads without the plumbing complexity of liquid.
High-density GPU and accelerator racks reject heat that air cannot economically remove. Cold plates carry coolant directly to the processors, capturing heat at the source where it is densest and hottest.
The densest training clusters can be submerged in dielectric fluid that contacts every component. Immersion removes fans and air handling entirely for the racks it serves, addressing heat flux that other methods cannot reach.
A single facility mixes methods — air for balance-of-plant compute, liquid for training, immersion for the extreme tier. Each hall is provisioned for the workload it hosts rather than forcing one method across the building.
In-nation residency means building where the sovereign is, so heat rejection is engineered around that site's ambient conditions and water reality.
Where ambient conditions allow, outside air and economizers reject heat with minimal mechanical work for much of the year. Chillers become a backstop rather than the primary path, cutting cooling energy substantially.
In arid regions, closed-loop dry and adiabatic systems minimize water draw. Cooling design respects local water stress so a data center does not compete with a population for a scarce resource.
In tropical conditions, liquid and immersion methods become more attractive because they decouple heat rejection from air handling. Design accounts for wet-bulb limits that constrain evaporative approaches.
Water usage effectiveness is treated as a first-class metric alongside power. Closed-loop coolant circuits, recovery, and treatment reduce consumption and keep the facility defensible to the community and regulator hosting it.
Bringing coolant to the chip introduces new failure modes that must be contained with the same rigor applied to the electrical path.
CDUs isolate the facility water loop from the technology loop feeding the racks, controlling temperature, flow, and pressure. This isolation keeps facility-side water chemistry away from sensitive server-side plumbing.
Rack-level sensing detects loss of coolant before it reaches live electronics, and drip trays and containment channel any escape. Quick-disconnect fittings let a node be serviced without draining the loop.
Pumps and heat exchangers are configured so a single failure does not starve a hall of coolant. Flow paths are designed to be maintainable while the racks stay cooled, mirroring concurrently maintainable power design.
Coolants and dielectric fluids are chosen for thermal performance, material compatibility, and safe handling. Fluid inventories and top-up logistics are planned so in-nation operation is not dependent on fragile external supply.
Liquid cooling concentrates waste heat into a usable form, which higher-temperature loops make practical to export.
Running the technology loop warmer raises the grade of recovered heat, making it useful downstream. Warmer coolant also widens the window for free cooling, compounding the efficiency benefit.
Where a site connects to district heating or a neighboring industrial process, captured heat can offset fossil demand elsewhere. Reuse turns a cost of computation into local value for the host community.
Temperatures, flow, and heat capture are metered continuously so cooling performance is observable and auditable. The same telemetry discipline that governs power governs thermal behavior.
Controls hold safe chip temperatures even during a cooling disturbance, throttling or shifting load before hardware or ledger integrity is at risk. Thermal protection is coordinated with the priority load classes it must defend.