Integrated data center design
Nvidia introduced technology that optimizes electricity use in AI data centers to operate more graphics processing units (GPUs) with the same amount of power. In an actual deployment by a cloud developer, the number of GPUs that could be installed increased by 37%, while AI token output rose by 50%. The company also emphasized its strategy of connecting AI chips developed by outside companies to its infrastructure.
Jung Gu-hyung, head of Nvidia Korea's solutions architect team, presented a strategy for treating an entire AI data center as a single computer and integrating the management of power and cooling facilities at Nvidia AI Day Seoul 2026, held at COEX in Seoul's Gangnam district on November 9.
“Nvidia aims to view an entire data center building or site as a single computer,” Jung said. “From this perspective, we are considering how to optimize energy use at data centers.”
Nvidia's DSX platform consists of design guidelines and open-source software needed to build AI data centers. The approach involves designing cooling facilities, power supplies and networks together with GPU servers. “We provide guidance on the cooling design, ceiling heights and building loads needed when deploying Vera Rubin NVL72,” Jung explained.
Dynamic power allocation
Nvidia particularly emphasized DSX MaxLPS, open-source software that adjusts power use in real time. Previously, maximum power limits were fixed for each server rack, but actual AI workloads vary over time, leaving unused power capacity at some racks.
“DSX MaxLPS operates dynamically,” Jung said. “It communicates with the building management system while continuously monitoring each floor's power use and temperature, as well as power consumption by rack.” This allows power to be reallocated between racks with high demand and those with spare capacity.
Nvidia presented its cloud partner Nscale as an actual deployment example. Nscale applied the software to GB300-based racks, increasing the number of GPUs it could install by about 37% within the same power budget. According to Jung, the number of tokens that could be generated with the same amount of power also increased by 50%.
Connecting custom AI chips
Jesse Clayton, Nvidia's AI infrastructure product marketing manager, introduced NVLink Fusion, which connects AI chips from outside companies to Nvidia infrastructure. “The Nvidia platform is vertically integrated but horizontally open,” Clayton explained. “NVLink Fusion extends this openness to custom chips.”
NVLink Fusion allows companies that have developed their own AI chips to use Nvidia's high-speed interconnect technology, server rack designs, networks and supply chain. “Developing a custom chip is only the beginning,” Clayton said. “System integration, networking and supply chain development afterward can require more than $1 billion in additional spending and more than 1 year.”
Original Korean article: Money Today
English brief by Korea Wind Intelligence, koreawindintel.com
