Table of Contents
Power distribution optimization is driving 800Vdc solutions that will move operators beyond standard side-car deployments
Sébastian Cruz-Mermy, vice president of Data Center Innovation and Tech Explore League, has spent 18 years incubating new business models, scaling start-up divisions, and driving technical exploration of AI-ready data centers and digital services. In today’s RCR AI TechTalk, he discusses high-density AI workloads, a phased approach toward 800Vdc sidecar architecture, and a future in which AI factories foster a decarbonized energy ecosystem.
AI workloads and higher rack densities
“The industry is facing challenges fueled by the level of growth it has to tackle. The need to scale at speed has never been so important,” says Sébastian Cruz-Mermy, who sees two primary challenges for data center operators: the new profile of the workload and the connections and access to the grid. He believes transitioning toward 800 High-Voltage Direct Current (Vdc) power distribution systems and sidecar configurations will be the most practical way to incrementally meet growing MW-scale rack demands.
“This new AI workload and the way the profile of the workload is changing means the operation of the data center and especially the power distribution systems and the thermal cooling systems have to adapt,” notes Cruz-Mermy, referencing the need to evolve from traditional 48/54V in-rack power distribution toward 800Vdc systems. As AI cluster workloads push compute requirements past 100 kW and rapidly toward 1 megawatt (MW) per rack, Cruz-Mermy believes that traditional low-voltage distribution methods have reached their physical and thermal limits.
“The focus is ‘right density,’ meaning pooling more and more GPUs together to enable the right compute.” In order to get the “right compute,” engineers must pack GPUs as tightly as possible to communicate continuously at ultra-high speeds. According to Cruz-Mermy, “data center operators adjusted to a first threshold, where high-density compute pushed rack power demands above 40 kW per rack. And now, the new reality is that densities are forcing the next threshold of 100kW and eventually as much as 500 kW per rack,” which he contends is rapidly pushing physical limits. In AC-to-DC power conversion, for example, the rack footprint is already cramped with power shelves, PSUs, and batteries. The goal is to move to smaller, lightweight cables and connectors, and to poach from already well-established supply chains (such as that 400 and 800 high-voltage components used for fast-charging EVs).
Because operators don’t want to rip out their existing AC electrical infrastructure, Cruz-Mermy says some operators first isolate the transition to the data center’s “whitespace” (the server floor), leaving the facility’s core upstream power architecture completely intact by using “sidecars” next to high-density GPU compute racks. Through a phased technical roadmap, the operators work to reclaim space and scale power with AC-to-DC conversion gear pulled out of individual compute racks and consolidated into “sidecar” cabinets at the end of a server row. After that, they can work toward native DC configurations in which AC-DC conversion is pushed upstream to the main data center substation or power plant backbone.
Cruz-Mermy acknowledges there are some limitations to such short-term temporary patches. While the row-level hybrid model solves the immediate space problem inside the server rack, it can still suffer inefficiencies that can only be fixed by moving the DC conversion step upstream in the power system. “To expand this DC distribution higher on the distribution chain of the power system and to have more centralized DC architecture is the dynamic that we are supporting and where a lot of innovations are happening to enable the industry to move in that direction,” explains Cruz-Mermy.
Liquid cooling and evaporative coolers
To solve the extreme thermal challenges of high-density GPU clusters, cooling has to capture intense heat directly from the silicon chips inside the server rack, and evaporative coolers have to safely reject that collected heat out into the atmosphere at the facility perimeter. Mermy frames liquid cooling as a baseline operational requirement for next-gen data centers, emphasizing that liquid cooling delivers radical energy efficiency gains and can drastically reduce overall water consumption.
“Liquid cooling drives an efficiency that is about 70% more energy efficient than air cooling, so we can drive potentially higher temperature of operation as well, and optimize heat ejection upstream of the data centers by leveraging technologies that consume far less water,” says Cruz-Mermy, who looks to a future of energy-efficient dry coolers and true zero-water consumption for a large portion of the year.
Because liquid cooling has a thermal conductivity that is roughly 25 times higher and a specific heat capacity 4 times greater than air, and because liquid is highly efficient at transferring heat away from the chips, the facility water feeding into the data center doesn’t need to be nearly as cold as the air used in a standard raised-floor server room. As data centers lessen their dependence on energy-intensive evaporative chillers, they can push the heat out to external closed-loop dry coolers that will rely on ambient outside air to reject heat without evaporating any water—achieving true net-zero water consumption for a vast portion of the year, depending on the location and climate of the facility.
Data Center Innovation and Tech Explorer League
To gauge the impact on power, water, and carbon footprints, Cruz-Mermy leads the Data Center Innovation and Tech Explore League at Schneider Electric – an internal and collaborative R&D incubator tasked with anticipating the long-term infrastructure disruptions of computing roadmaps (from the likes of Nvidia, AMD, and Cerebras) and identifying, testing, and validating next-gen electrical and thermal solutions . “The purpose is to go beyond and to look at the data center of the future – the next generation, and the next generation after that, to consider which new technology will disrupt and change the equations for automation, control, power, and cooling distribution,” says Cruz-Mermy.
Currently, the League is working with NVIDIA to deploy integrated reference architectures spanning energy, chips, infrastructure, models, and applications, as well as the DSX Blueprint, which allows Schneider to drop physical infrastructure assets directly into a digital twin for advanced simulation about how ultra-high-density rack spaces will behave thermally and electrically.
“We have a strong understanding of the automation and energy parts, so we can really combine the subsystem together to bring much more intelligence, leveraging AI to accelerate energy intelligence optimization for the data centers,” says Cruz-Mermy. This vision unifies two historically distinct business pillars: power delivery (energy) and cooling/building controls (automation). Merging the electrical distribution subsystem with automated components—like pumps, valves, and liquid flow controls—under AI-driven management software allows for more seamless facility orchestration.
“We continue to explore new fields and new technologies for high density, evaluating which levers can be applicable and innovate on the power generation and automation side,” says Cruz-Mermy, adding that “the data center has to become an active player of the ecosystem, to act as a stabilizer of the grid, and eventually to even help decarbonize the ecosystem, for instance, by enabling more renewable energy to be injected into the grid for the benefit of everyone.”
Renewable energy sources like wind and solar are dynamic and intermittent, so Schneider Electric is looking at how data centers can absorb some of the instability for utility companies through intelligent throttling and backup batteries that can push stored electricity to the grid to smooth out drops in renewable energy production.
“That is the innovation we are working on to enable the dynamic management of different energy sources,” points out Cruz-Mermy, who believes the data center industry will continue to be a pioneer in efficiency and in net-zero goals. If AI facilities ultimately seamlessly blend traditional grid power, on-site solar arrays, wind energy, local battery storage, and alternative backup generation (like hydrogen fuel cells) it’s possible that they could dynamically adjust to grid stress, carbon intensity, and fluctuating utility pricing.
He contends that the data center industry is committed to net-zero goals, and that efficiency innovation will accelerate: “If you look at the energy consumption over the past 15 years, we have been exponentially growing the number of data generated by data centers, but the energy consumption has remained steady. Now with this growth, we have to help the industry decouple energy consumption and energy efficiency from growing data generation.”
Moving to high-efficiency power delivery systems, such as 800VDC and shifting from air cooling toward direct-to-chip liquid cooling are steps toward high-efficiency data centers that will run more sustainably and inject flexibility back into the grid, “with very low carbon intensity and an ability to decarbonize the grid so everyone can move toward lower carbon intensity,” according to Cruz-Mermy.
