Container-based AI data centre pods.

For thirty years a data centre was a floor space problem. You bought square metres, filled them with racks at 5 to 10 kW each, and cooled them with air.

AI broke that model. A single rack of current accelerators concentrates more power into two square metres than an entire aisle used to draw, and the next generation is specified to draw more again. The constraint is no longer how many racks fit. It is how much heat can be removed from the space you have, and whether the building can deliver the electricity in the first place.

That is why the useful density metric for AI infrastructure is kilowatts per square metre, not rack units per square metre. A conventional 40ft air-cooled module delivers roughly 5 to 8 kW per square metre. A Coral Pod C40 delivers 28.3 kW per square metre of thermal capability.

A Coral Pod is not a container with servers in it. It is seven engineering layers designed together and integrated in a factory: structural enclosure, thermal core, heat rejection, power, network, control and safety, and operations.

The reason this matters is not elegance. It is risk. In a conventional deployment the customer solves power, cooling, water, connectivity, compliance and operations separately, with a different vendor and a different timeline for each, and carries the integration risk alone. That is where modular data centre projects actually fail, and it is the only risk a buyer cannot transfer to a component vendor.

CoralDC moves that integration into a controlled factory environment. Manufacture complexity. Deploy simplicity.

All thermal capability figures are four times the 200 kW rating of the Coral 42U immersion tank.

Thermal capability is what the tanks can remove. Populated load is what a given hardware generation actually draws. They are different numbers and most of this industry quotes only the first.

Calculated at 10.2 kW per 8U HGX H200 class node and 14.3 kW per 8U HGX B200 class node. Based on published NVIDIA GPU thermal design power of 700 W for H200 SXM and approximately 1,000 W for B200.

The headroom is the product. A Coral Pod populated today at roughly 27 percent of its thermal capability does not need replacing when rack density rises through the next two hardware generations. The enclosure, the tanks, the pumps and the heat rejection were sized for where the roadmap is going, not for where it is.

Overhead service clearance. A 42U immersion tank is 1,620 mm tall. An 8U GPU node is 356 mm. A lifting frame and hook adds approximately 450 mm, and handling margin adds 100 mm, so vertical extraction requires approximately 906 mm of clear space above the tank. A standard 40ft container gives 770 mm of internal height above the tank. A 40ft high cube gives 1,080 mm. That is why the C40-HC exists. In a standard 40ft you accept front-access service. In a high cube you get top-loading service with a rail gantry. CoralDC states this on the specification page rather than letting a customer discover it on day one of operation.

Floor loading. A 42U tank has an operating weight of approximately 2,200 kg over a 2.57 square metre footprint, which is 855 kg per square metre, or 8.4 kPa, at every tank position. Every Coral container carries a reinforced floor channel under each tank position as standard. Where a pod is placed on a structure rather than a slab, CoralDC states the loading in writing before survey. A typical data centre slab is rated 1,000 to 1,500 kg per square metre and accepts this without modification. A typical commercial office floor is rated 250 to 500 kg per square metre and requires structural review.

Service aisle. Every Coral container pod has a 1,300 mm clear single-sided service aisle. That is enough for two technicians to work on one node without removing adjacent equipment. It is a deliberate choice to spend container volume on serviceability rather than on tank count, and it is why a C40 carries four tanks rather than a theoretical maximum.

Heat leaves a Coral Pod through four loops. The primary loop is single-phase dielectric fluid in direct contact with the board, with dual redundant pumps, side-stream filtration and continuous condition monitoring. The secondary loop is water and glycol at DN50 and 25 cubic metres per hour. The rejection loop is a roof or skid mounted dry cooler as primary, dew point technology for hot dry climates, and a trim chiller only in extremis. The fourth loop is the control plane, which is what makes the other three worth having.

No facility water connection is required in dry rejection configuration.

The loop temperature question, stated openly. CoralDC pods operate oil at 25-35 C today with a 45 C maximum. That is a conservative design point chosen for component margin.

Current NVIDIA rack-scale reference designs specify coolant supply up to 45 C, leaving the compute at approximately 55 C. Designing the secondary loop to that band is what converts trim chiller hours into dry cooler hours, and converts return heat from thermally worthless into directly sellable to a nearby offtaker.

Raising the Coral loop supply temperature toward that band is a stated engineering direction for the next platform generation, not a current specification. CoralDC publishes both figures because the difference is the single highest-leverage number in the entire thermal design, and a buyer evaluating a ten-year asset is entitled to know which one applies today.

A conventional AI campus is a single large decision: one site, one grid connection, one 24 month build, one moment at which capacity arrives or does not.

A Coral factory is a series of 800 kW decisions. Each pod is independently financeable. Each can frequently be sized to operate inside a building's existing electrical capacity rather than waiting on an interconnection queue. Each earns revenue before the next one is ordered. Pods interconnect through three shared services: a power bus, a liquid loop and the AI fabric.

Durations are typical for factory-built modular infrastructure of this class and are confirmed per site during the architecture review. The structural advantage is that a self-contained pod removes data hall construction entirely, where a conventional build is measured in quarters or years.

CoralDC is a Canadian company. Engineering, systems integration and deployment are based in Vancouver, British Columbia. For Canadian and United States government, defence, healthcare and critical infrastructure procurement, country of engineering is a scored criterion rather than a preference, and CoralDC states its position plainly rather than leaving it to be discovered.

CoralDC publishes its pod general arrangement, transverse section, thermal schematic and scaling topology. Most vendors in this category publish a photograph. A drawing tells a consulting engineer whether the product fits their building. A photograph tells them what it looks like.