CoralDC's Coral Edge, the 12U pod, and Coral OnPrem, the 42U pod, deploy single-phase immersion cooling today: sealed dielectric fluid, dual redundant pumps, no fans in the thermal path. This is the current, shipping architecture, not a roadmap claim. CoralDC is also engineering a hybrid direct-to-chip-plus-immersion platform as its next generation, for a specific and disclosed reason set out below.
Current NVIDIA rack-scale AI systems, the generation after the hardware CoralDC's pods are built around today, ship one hundred percent liquid cooled at the compute tray, with microchannel cold plates fitted at manufacture and no air-cooled configuration offered. Coolant supply is specified up to 45 C, leaving the compute at approximately 55 C. NVIDIA has published no immersion reference architecture for this generation.
The practical consequence: on that next-generation rack-scale hardware, the cold plates are already there, bonded through a thermal interface stack onto the package. Removing that assembly to submerge a bare die is not an upgrade. It is a warranty question, an electrical safety case nobody has published for a busbar of that rating, and an unresolved compatibility question for co-packaged optical components, in exchange for no thermal benefit the fitted cold plates are not already delivering.
This is why CoralDC's next platform is hybrid, combining direct-to-chip liquid cooling for that class of hardware with immersion retained for PCIe and HGX class hardware, and for board-level and sealed-environment heat, where it remains the stronger answer. This is a direction under active engineering, not a current product, and this page says so rather than describing it as available today.
CoralDC states both of these things openly, the strength of immersion today and its limit on next-generation hardware, because a buyer who is told immersion is the universal answer, and later discovers otherwise, has been given a reason to distrust everything else they were told.
Immersion is CoralDC's current product, not a fallback.
Power delivery components, voltage regulation and the residual load that cold plates do not reach are exactly the class of heat a dielectric bath handles well. In a hybrid architecture this is where the bath earns its place.
A large share of institutional AI, in hospitals, universities, government departments and industrial sites, does not run on flagship rack-scale systems. That hardware is immersion-viable and it is where Coral's fluid engineering remains a genuine advantage.
Dust, humidity, vibration, salt, acoustic limits, no room, no air handling, no on-site technician. Air cooling is not merely less efficient in these conditions, it is frequently not possible. This is the strongest case for immersion anywhere, and it is the Coral Edge case.
Once the Gen 2 hybrid platform is available, five inputs will determine which architecture a deployment uses. Today, every Coral Edge and Coral OnPrem pod ships single-phase immersion, and the inputs below describe the future selection logic rather than a current menu of options.
What the customer is actually deploying determines what is thermally possible before anything else is considered.
Rack power for current generation rack-scale systems is planned at 220 kW sustained with provision to 250 kW. Published figures range widely, from around 120 kW to over 250 kW for nominally the same system, and CoralDC plans against the high end deliberately. A pod sized for 130 kW that meets a 200 kW rack is not a product with a performance shortfall. It is a product that cannot be sold.
Whether facility water is available, permitted and politically secure at the site over the asset's life.
Rack-scale systems have been engineered so a tray service event takes minutes. Fluid handling changes that materially, and for some customers the service model is the deciding constraint rather than the thermal performance.
Floor loading, access, acoustics, ambient conditions, staffing.
A 45 C supply loop is the first generation where chiller-free dry or dew-point heat rejection is reachable across most of the climates CoralDC deploys into. That is the design decision that makes near-zero operational water consumption achievable rather than aspirational.
For context on the scale of the difference: evaporatively cooled facilities have been characterised at approximately 2.6 million US gallons per MW per year, against near zero for a closed-loop reference design. CoralDC models its own deployments at PUE 1.12 for single-phase immersion with dry cooling, and 1.07 for a hybrid architecture at a 45 C loop with automated thermal control. These are planning figures, deliberately more conservative than published vendor best cases, and CoralDC will not present a vendor best case as its own performance.
A 55 C return is hot enough to be useful to someone else. Most cooling architectures mix that grade down before rejecting it, which destroys the option. Coral designs the loop to preserve the grade, because it is the difference between waste heat as a sustainability talking point and waste heat as a contracted supply. Whether the economics work is site-specific and is treated honestly on the Heat Reuse research page.
A structured technical conversation about power, thermal capacity, density, jurisdiction and deployment sequencing for a specific site. If the answer is that CoralDC is not the right fit, it says so.
CoralDC's Coral Edge and Coral OnPrem pods use single-phase immersion cooling today, shipping as the current product. CoralDC is engineering a hybrid direct-to-chip-plus-immersion platform as its next generation, specifically to support next-generation rack-scale hardware that ships with cold plates fitted at manufacture. The two are not blended in current product marketing; this site states which is which.
Not at the complete-system level. The generation of hardware after what CoralDC's pods are built around today ships with microchannel cold plates fitted at manufacture, has no air-cooled configuration, and has no published immersion reference architecture. This is the reason CoralDC's next-generation platform is hybrid rather than immersion-only, and it is not a statement about the immersion pods CoralDC sells now.
CoralDC plans current-generation rack-scale thermal envelopes at 220 kW sustained with provision to 250 kW. Published figures vary widely, and the cost of over-sizing a secondary loop is a small percentage of pod capital cost while the cost of under-sizing is an unusable product.
A closed-loop deployment with dry heat rejection consumes effectively no water in operation. Whether dry rejection is viable across a full year depends on climate, and is confirmed per site during the architecture review rather than claimed generally.
