The next stage of AI infrastructure expansion is increasingly a problem of moving electricity in and heat out. A group of October 12 announcements around the OCP Global Summit shows suppliers working at different points along that chain: Enteligent with its PowerCore 800VDC-to-chip conversion stack, DG Matrix with Interport power infrastructure, Cooler Master with a broader liquid-cooling portfolio, KIOXIA with liquid-cooling support for enterprise SSDs, and Accelsius with expanded testing capacity in Austin. Taken together, the announcements illustrate an engineering problem that reaches from facility power delivery to individual components inside a server.
Enteligent’s announcement focuses on converting DC power toward the voltages computing components require, while DG Matrix’s Interport 360 announcement positions power infrastructure around NVIDIA MGX-based AI factories. Their shared relevance is the need to manage power delivery as an architecture, rather than a sequence of isolated purchases. Higher-voltage distribution can reduce the current required to deliver a given amount of power. The benefit in a real installation, however, depends on conversion efficiency, protection, cabling, service arrangements and compatibility across the full chain. A higher voltage is a design choice with consequences, not a standalone efficiency guarantee.
Power conversion and cooling are closely connected. Losses in the electrical path become heat, while computing components turn most of their consumed electricity into heat that the facility must remove. Improving one part of the system can therefore influence the demands on another. But those improvements do not automatically translate into lower total consumption. If an operator uses improved infrastructure to install more computing equipment, the facility’s overall load can still rise. Efficiency per unit of useful work and total electricity demand need to be assessed separately.
Cooler Master’s portfolio is anchored by a 2,500 kW liquid-to-liquid coolant distribution unit, with mass production targeted for the fourth quarter of 2026. The company also describes rack and server cooling equipment and tools used for filling and commissioning. The inclusion of those deployment tools is revealing. Liquid cooling requires more than an effective cold plate: the system has to be assembled, prepared, checked, maintained and serviced reliably. The plumbing around the electronics becomes part of the computing platform’s operational performance.
KIOXIA is extending that thermal approach into storage. Its new CM9 direct-liquid-cooling models use the E3.S form factor and support cold-plate-enabled server designs, while also accommodating traditional air-cooled environments. The company says CM9 drives are sampling to selected customers. Storage cooling may attract less attention than accelerator cooling, yet its inclusion makes the broader point: increasing server density requires designers to consider the thermal behavior of the whole machine. Removing heat from the largest component does not remove every other temperature constraint.
Accelsius’s announcement of multi-megawatt testing capacity at its Austin laboratory highlights another requirement: proving a cooling system under demanding conditions. Laboratory capability does not demonstrate the reliability or economics of every customer installation. It does create an opportunity to examine behavior at a more representative scale. For prospective operators, evidence about stability, failure handling and maintenance can be as important as peak heat-removal performance. A system that performs well only under ideal conditions is a difficult foundation for a facility expected to run continuously.
Existing data centers face a different problem from new sites. Their electrical distribution, floor layout, cooling equipment and maintenance practices may have been chosen for earlier generations of servers. Introducing dense liquid-cooled systems can require changes that extend far beyond replacing a rack. New facilities can begin with a more integrated plan, but they still have to coordinate equipment, commissioning and operations. In either case, a component’s advertised capability needs to be evaluated within the limitations of the actual building.
The business implication is a wider set of potential infrastructure suppliers and a more complicated purchasing decision. Operators will need to compare lifecycle cost, delivered performance and reliability across interconnected systems. Product announcements establish possible building blocks; customer qualifications and deployments establish whether those blocks work together at scale. The next useful evidence will be installation experience, service requirements and sustained operation, rather than headline capacities alone.
AI deployment ultimately depends on a chain of physical tasks being completed well. Electricity must reach the equipment at usable voltages, computing must proceed within thermal limits, and the resulting heat must leave the system. The October announcements show several suppliers addressing that chain from different directions. Their collective significance is that infrastructure design is becoming harder to separate into neat categories. The practical unit of progress is a working facility, with power and cooling matched to the work it is expected to perform.