The standard objection to liquid cooling at the server level is water risk. IT teams who have managed air-cooled facilities for years treat any coolant inside the server chassis as a contamination event waiting to happen. ZutaCore's OmniTherm cold plate addresses the objection directly: it achieves two-phase liquid cooling for GPU-class thermal loads without routing water into the server at any point. The system captures heat through a phase-change process, transitioning the working fluid from liquid to vapor at the chip surface and rejecting that heat at a remote condenser, with the water loop staying entirely outside the server enclosure.
The specific target is NVIDIA's RTX PRO 6000 Blackwell Server Edition GPU in a single-slot PCIe configuration. Shahar Belkin, Chief Evangelist at ZutaCore, framed the design priority precisely: the single-slot form factor allows operators to increase accelerator density in standard server architectures while capturing heat into a liquid loop. That combination, higher GPU count per chassis and liquid heat rejection per GPU, addresses both the thermal and the footprint constraint simultaneously. At full Blackwell load, the fan speeds required for air cooling generate noise levels that create operational problems in mixed-use facilities; OmniTherm reduces fan dependency and the acoustic consequence along with it.
The liquid cooling market breaks broadly into single-phase direct-to-chip and two-phase approaches. Single-phase systems circulate water or a water-glycol mixture through cold plates, relying on sensible heat capacity: the fluid absorbs heat and returns to the CDU warmer, carrying that energy out of the rack. Two-phase systems use latent heat instead, which is thermodynamically more efficient per unit volume of fluid moved. The working fluid boils at the chip surface, absorbing heat at constant temperature, and condenses elsewhere in the loop, releasing it. The effective heat transfer coefficient in the boiling regime is substantially higher than single-phase convection.
Immersion cooling also uses two-phase heat transfer in its vapor phase variant, but the deployment model is completely different: the entire server is submerged. OmniTherm retains standard server architecture. The chassis, the PCIe form factor, the memory configuration, the NVMe storage, all remain in place. Only the GPU cooling path changes. That compatibility with standard enterprise server infrastructure is what ZutaCore is positioning as the differentiator against both immersion and against the NVL-class chassis designs that integrate liquid cooling at the OEM level.
Hyperscale operators building ground-up AI clusters have different options than enterprise operators retrofitting existing facilities for AI inference workloads. The hyperscale path involves OEM-designed systems with integrated liquid cooling from the chassis up, like the NVL72 architecture, where NVIDIA's watt roadmap is already deciding the cooling architecture. Enterprise operators running standard rack infrastructure face a harder trade-off: accept the thermal and acoustic penalties of air cooling at Blackwell TDP, or invest in a liquid cooling retrofit that preserves standard server form factors. OmniTherm is aimed at the second group. The question for those operators is whether the density and efficiency gains justify the transition from a facility infrastructure perspective, specifically whether their existing chilled water or CDU loop can serve the OmniTherm condensers without additional plant investment.