At the Dell Technologies World keynote, Dell and NVIDIA unveiled a sweeping expansion of their enterprise AI infrastructure. The new AI server "Dell PowerEdge XE9812" is built on NVIDIA's new "Vera Rubin NVL72" architecture and is said to lower cost-per-token by up to 10x compared with the previous Blackwell generation. The announcements reflect a shift in which generative AI is moving past the pilot stage toward running autonomous "agentic AI" at scale[1].

Enterprise AI Demand Enters the "Useful AI" Era

Speaking at the keynote, Dell Chairman and CEO Michael Dell projected that worldwide AI infrastructure spending could reach 3 to 4 trillion USD (about 480 to 640 trillion yen) by 2030, with token consumption growing as much as 3,400 percent over the same period[1]. ※1 USD = 159 JPY (as of May 29, 2026)

NVIDIA founder and CEO Jensen Huang explained that demand is surging because we have "arrived at the era of useful AI." Tasks that once took months now take weeks, what took weeks now takes days, and the gains in productivity come with a corresponding leap in the amount of computation required[1].

Already, 5,000 enterprises such as Lilly, SAMSUNG, and Honeywell are running AI workloads in production on the "Dell AI Factory" with NVIDIA, signaling that corporate AI use is moving from proof-of-concept to full-scale deployment[1].

New Vera Rubin Servers Compress Cost-Per-Token

The core of the announcement is a refresh of servers for accelerated computing. The flagship Dell PowerEdge XE9812 is built on Vera Rubin NVL72 and is described as cutting cost-per-token by up to 10x versus Blackwell for large-scale agentic AI inference[1].

It is joined by the "PowerEdge XE9880L," "XE9885L," and "XE9882L" — the first Dell systems built on NVIDIA HGX Rubin NVL8. They support up to 144 GPUs per rack with 100 percent direct liquid-cooled compute nodes, aiming for up to 10x the performance of HGX B200[1].

On the networking side, a new "Dell PowerSwitch" lineup with NVIDIA Quantum-X800 InfiniBand arrives alongside NVIDIA Spectrum-6 Ethernet. Dell also introduced "Dell PowerRack," a fully integrated system that engineers compute, networking, and storage as one, removing the overhead of assembling components individually[1].

A Refresh for the Vera CPU and Data Platform

On the CPU side, "Dell PowerEdge M9822" and "R9822" servers bring NVIDIA's new "Vera" CPU into the enterprise AI factory. Purpose-built for the sequential steps of agentic AI such as data pipelines and analytics, Vera offers 1.2 TB/s of memory bandwidth and is said to complete agentic workloads 50 percent faster than x86 processors[1].

Huang stated that "Vera CPU has the highest single-threaded performance of any CPU in the world," noting that its 3x memory bandwidth translates into faster database processing. "Starburst," a new data engine added to the Dell AI Data Platform, is said to deliver 3x faster query throughput for large-scale SQL analytics on the Vera CPU[1].

Frontier Models and AI Agents Protected On Premises

Dell's own survey, cited from the keynote stage, found that 67 percent of AI workloads now run outside the cloud — on premises, on device, at the edge, or in colocation — and that 88 percent of respondents run at least one AI workload on premises[1].

Positioned as the key to meeting this demand is "NVIDIA Confidential Computing," which processes sensitive data and models while keeping them protected. Delivered with partners including Fortanix, Google, and Red Hat, it is designed to let enterprises run frontier models securely inside their own perimeter without exposing model intellectual property or data[1].

Specifically, Google Distributed Cloud with Gemini 3.0 is offered in preview on the NVIDIA Blackwell-powered "PowerEdge XE9780," while open models such as NVIDIA Nemotron run on the Dell AI Factory. For agents, OpenAI's coding assistant "Codex" will connect with the Dell AI Data Platform, with a plan to link it to internal codebases and business systems[1].

Closer to the individual workspace, deskside agentic AI using NVIDIA Nemotron open models runs on the "Dell Pro Max with GB10/GB300," powered by the NVIDIA Grace Blackwell architecture. Supporting these are NVIDIA Nemotron for building agents, agent orchestration, and an open-source runtime for security, all provided across the entire Dell AI Factory[1].

Summary

These announcements show that corporate AI use is moving from proof-of-concept to production, entering a stage where autonomous agentic AI can run securely and at scale on premises. Compressed cost-per-token from Vera Rubin servers, a refreshed Vera CPU and data platform, and mechanisms to protect sensitive data were presented as a single package. Deeper sessions are scheduled for the second day of Dell Technologies World and beyond, with the conversation carrying over to GTC Taipei at COMPUTEX, held from June 1 to 4.

出典:https://blogs.nvidia.com/blog/dell-technologies-agent-enterprise-ai