Dell proposes new AI-infrastructure metrics to measure agentic multi-agent workloads

Dell Technologies used its annual conference in Las Vegas on May 18 to make a specific argument about how enterprises should buy and measure AI infrastructure. The company said the old yardsticks — GPU counts, model benchmarks and broad cloud-versus-on-premises comparisons — do not capture the economics of agentic AI systems that call multiple models, tools and data sources to complete a task. Lopez Research, which attended the event, said Dell pushed “time to token” and “cost per token” as more useful measures for enterprise buyers. Dell tied that framing to a broader set of product announcements around local agentic AI, multi-agent orchestration, storage and rack-scale infrastructure. ### Which metrics is Dell trying to put at the center of the discussion? Lopez Research reported on May 19 that Dell Chairman and CEO Michael Dell told enterprise technology leaders that infrastructure metrics are changing, with “time to token” and “cost per token” becoming more relevant for AI deployments than simple hardware counts. The firm said the point was to measure how quickly a system begins producing useful output and what each unit of output costs. (lopezresearch.com) Dell’s own event materials did not publish a standalone manifesto on those metrics, but the company repeatedly described its new agentic AI products as “cost-predictable,” “local” and designed to move customers from “AI ambition to realized outcomes.” In a separate conference blog, Dell Chief Operating Officer Jeff Clarke said the company was focused on the economics of running AI at enterprise scale, including lowering token costs through on-premises systems. (lopezresearch.com) ### Why does Dell say agentic AI changes the math? Lopez Research wrote that agentic AI systems can require 100 to 1,000 times more computation than a simple query because they break work into multiple reasoning, retrieval and tool-use steps. That means infrastructure spending is no longer tied only to the cost of answering one prompt, but to the cost of completing a workflow. (dell.com) Forbes, citing Clarke’s remarks, reported on May 19 that token consumption for AI reasoning can rise sharply in agentic systems, adding pressure to cloud economics and data-center design. That account is consistent with Dell’s conference messaging that enterprises need infrastructure sized for sustained inference and orchestration, not only model training or chatbot-style interactions. (lopezresearch.com) ### What products did Dell use to support that pitch? Dell announced on May 18 that it was adding “Dell Deskside Agentic AI” to the Dell AI Factory with NVIDIA. The company said the offering uses Dell workstations, NVIDIA NemoClaw software and Dell services to run models ranging from 30 billion to 1 trillion parameters locally, with OpenShell integration to build, test and govern agents from the desk to the data center. (forbes.com) The same day, Dell said it was expanding the Dell AI Factory with NVIDIA across data orchestration, storage, networking and rack-scale systems. Those additions included faster file indexing, claims of up to six-times-faster SQL query performance, new Omniverse integration, and PowerRack infrastructure meant to package compute, storage, networking and cooling into a single system. (dell.com) ### Where do multi-agent workflows and governance fit in? Dell said on May 18 that support for NVIDIA AI-Q 2.0 blueprints would speed deployment of multi-agent workflows on Dell AI Factory infrastructure. The company also said OpenShell would provide a sandboxed runtime with security and privacy controls for agent development and governance. (dell.com) Zawya, citing Dell’s announcements, said the company was extending production-ready multi-agent workflows and making enterprise data more “AI-ready.” Dell also said it was adding an ecosystem program and partnerships with companies including Google, Hugging Face, OpenAI, Palantir, Reflection and ServiceNow to broaden deployment options on infrastructure customers control. (dell.com) ### What is the practical change in how buyers may compare systems? Lopez Research said Dell’s presentation shifted the conversation toward unit economics: how much infrastructure is required to deliver a usable business outcome, and at what cost. That does not eliminate traditional performance measures, but it puts more weight on throughput, latency, orchestration overhead, governance and data locality in production systems. This is an inference from Dell’s product framing and Lopez Research’s account of the conference. (zawya.com) Dell’s public materials point to the next phase of that effort. The company said the new agentic AI, data-platform and rack-scale products were introduced at Dell Technologies World on May 18 in Las Vegas, and several of them are tied to the Dell AI Factory with NVIDIA and partner software that customers can now evaluate for enterprise deployments. (dell.com) (lopezresearch.com)