NVIDIA outlines infrastructure direction for agentic AI workloads
DataNews Editorial Team
In an interview with Data Center Richness, NVIDIA Senior Director Dion Harris described an infrastructure approach for agentic AI built around specialised CPUs, multiple GPUs, advanced networking, liquid cooling, higher-density racks and new power architectures. The discussion sets out NVIDIA’s technology direction rather than announcing a specific data-center deployment.
- NVIDIA frames agentic AI infrastructure as an integrated CPU, GPU, networking, cooling and power challenge.
- NVIDIA says Blackwell and Vera Rubin require liquid cooling at high compute densities.
- DSX Flex is presented as a tool for controlled power reduction during grid peak events.
- The company identifies 800V DC distribution as a direction for high-density AI clusters.
- No specific data-center campus, capacity, investment, construction timeline or operational deployment was announced.
NVIDIA is positioning agentic AI as an infrastructure problem that extends beyond accelerator availability. In an interview with Data Center Richness, Dion Harris, Senior Director of HPC and AI Hyperscale Infrastructure Solutions at NVIDIA, said these workloads require tighter integration across specialised CPUs, networking and multiple GPUs, alongside higher-density rack designs and revised cooling and power systems.
According to NVIDIA, its Blackwell and Vera Rubin platforms require liquid cooling at high compute densities. This places cooling architecture alongside IT design as a core consideration for operators deploying dense AI clusters, rather than as a downstream facility engineering decision.
NVIDIA also described DSX Flex as a grid-interactive tool intended to reduce power consumption safely during grid peak events. For data-center operators, such controlled load reduction could become increasingly relevant where available capacity and grid interconnection are constraints.
The company said the industry is preparing for 800-volt DC power distribution in high-density rack clusters, with the stated objective of improving energy efficiency and delivering more tokens per megawatt. This should be viewed as a development direction, not evidence of broad commercial deployment.
NVIDIA further claimed that its Vera CPU delivers twice the performance of traditional x86 architecture. The comparison was presented without methodology or independent verification in the source material.
The interview also referenced NVIDIA’s entry into the Land, Power and Shell market for data-center campuses. However, it provided no details of a specific campus, including capacity, investment, construction schedule or operational status.
If NVIDIA’s proposed direction is implemented widely, competition in AI infrastructure could increasingly depend on the ability to deliver integrated, high-density systems combining liquid cooling, DC electrical distribution, network fabrics and managed load-curtailment capabilities. That could accelerate the transition from conventional server halls toward purpose-built AI factories and increase the strategic value of available power and grid connections.
The approach points to a shift from assembling individual server components toward integrated AI-factory infrastructure. If adopted broadly, it could raise the importance of liquid cooling, DC power distribution, network fabrics, grid interconnection and controllable load flexibility in high-density AI data-center design.
- — / October 16, 2026 · EnglishSource →
This article was prepared by the DataNews editorial team based on the sources listed above.