1. To make a wise purchase decision, you need real sizing data

   Agentic AI consumes much more computing power than a standard chatbot. Instead of a single inference call, an agentic task can spawn hundreds of calls—all running in parallel, with tool execution and agent coordination between every step. We tested the Dell PowerEdge XE7740 on real-world agentic workloads, giving you real, fact-based sizing data.

2. More GPUs mean more agents

   On our financial services workload, the 8-GPU configuration supported up to 74 simultaneous AI agents while keeping task-completion latency under 30 seconds. On the manufacturing workload, doubling GPU count nearly doubled supported agents. While you should choose the right GPU count for your needs, it’s clear that more GPUs means more agentic AI opportunity.

3. Throughput scales with your GPU investment

   More throughput means agents complete tasks faster. While all configurations performed well, the 8-GPU configuration we tested more than doubled the throughput of the 4-GPU configuration on the manufacturing workload while staying within our latency threshold.

4. On-premises ownership can give you a cost advantage

   Cloud platforms charge per token. With an on-prem solution, however, hardware is a one-time expense. With high utilization, every PowerEdge XE7740 configuration we tested had a significantly lower cost per million tokens ($/Mtok) than Amazon Bedrock for the same model, an advantage that compounds for workloads running around the clock.

5. Agentic AI demands CPU as well as GPU power

   Agentic AI requires both powerful CPUs and GPUs. In our testing, CPU utilization ranged up to 99% within a single run, peaking during task assignment and agent orchestration. GPU count alone doesn’t tell the full sizing story.

Why agentic AI matters today

Over the past several years, AI has become ubiquitous. The hottest frontier now is agentic AI: systems in which an AI agent receives a goal, rather than a single prompt, and works through the steps to accomplish that goal on its own. The agent plans, takes actions, evaluates the results, and decides what to do next. Along the way it can browse websites, query databases, write and run code, send emails, and invoke APIs and software tools, among other activities. Where a chatbot answers, an agent acts.

With more powerful and accurate models behind them, AI agents can handle increasingly complex tasks. That translates into cost savings for organizations, enabling staff to hand off their repetitive and manual tasks to AI agents and focus their time on more valuable, higher-order strategy. Businesses across industries are looking to take advantage of these potential benefits and move from demo to production. According to a recent Deloitte study of IT leaders, 74% of IT and business leaders anticipate that their companies will be using agentic AI “moderately” or more by 2027.¹

But agentic AI requires a great deal of computing power—which falls on the entire platform, not just the GPU. A single chatbot exchange (one user message and its reply) typically maps to one inference call. In contrast, a single agent task can involve many, many more inference calls running in sequence and in parallel. A predefined workflow may run 5 to 20 model calls per task, while an open-ended agent that directs its own steps can run into the hundreds. In some workloads, agents can even spawn sub-agents, which multiply the workload. That is a dramatic increase in work per user request. Practically, it means that even if a company is happy with the servers running their internal LLMs in chat use cases, those same servers might not be the right fit for agentic workloads.

To find and size the computing solution that will best match their agentic AI goals, organizations need to understand how many agents a solution can support, what kind of performance users will see, and what that solution will cost. This report examines how one server, the Dell PowerEdge XE7740, performed on those metrics.

Exploring the agentic AI performance of the Dell PowerEdge XE7740

What we tested

We wanted to understand how various configurations of the Dell PowerEdge XE7740 server perform for agentic AI workloads. The server is available in several different CPU and GPU configurations, and organizations considering the server will benefit from this data to size the right configuration for their agentic needs. (Note that customers can order only the 4-GPU and 8-GPU configurations; the 6-GPU configuration is supported, but not orderable.)

We measured how performance scaled with increasing numbers of GPUs. To do this, we configured the server with two Intel Xeon 6747P CPUs and tested with increasing numbers of NVIDIA RTX PRO 6000 Blackwell GPUs: first four, then six, then eight. This real-world data can help you determine which configuration of the PowerEdge XE7740 will be the best fit for your needs.

How we assessed real-world agentic performance

Many existing agentic benchmarks focus only on accuracy, or how well the agent did its job. For this study, we wanted to focus on performance, with the ultimate goal of providing sizing guidance for organizations choosing a server configuration for their agentic workloads. How many agents can the server support while keeping accuracy high and latency below an acceptable threshold? And how does that number change with different GPU configurations?

To answer these questions, we used a custom agentic benchmark that runs realistic AI-agent workflows end-to-end on a computing solution. These workflows are real-world; each of them represents work that an agentic AI system might actually perform in a corporate environment. When we combine our results with up-to-date cost data, this benchmark tells us:

• How many tokens per second (throughput) the server can sustain
• How many AI agents, or workers, a solution can run within a predefined service level agreement (SLA)/latency threshold
• Tokenomics data, including cost to run, amortization of hardware, and cost of energy consumption

We ran our tests on the PowerEdge XE7740 server on the benchmark’s medium difficulty tier, representing a medium to large business doing tasks of moderate complexity. (The benchmark allows us to run at simple, moderate, or complex difficulty tiers, with each tier representing an organization and task set of a different size and complexity.) We tested the benchmark with the well-known large language model (LLM) Llama-3.1-70B-Instruct (quantized to 8-bit precision); vLLM as an open-source engine serving the LLM; and Qwen3.5-2B (unified multimodal vision) as the vision-language model (VLM).

Because agentic AI agents can do a wide range of work, much of it industry- and company-specific, we built the benchmark to simulate a range of realistic industry-specific workflows. For this study, we used the financial services and manufacturing workflows, each of which incorporates eight workflow scenarios.

Agentic AI workloads in manufacturing

Whether they’re building cars or producing pharmaceuticals, many manufacturing facilities produce billions of telemetry rows of data. AI agents in these factories catch quality issues, identify faulty equipment, predict equipment failures, and trace contaminations. Because every pause in production has a dollar cost, keeping systems up and running is critical.

Our benchmark incorporates eight real-world manufacturing workflows:

• Statistical process control, where an agent loads multi-series production telemetry to identify which parameters are off and likely root causes
• Overall equipment effectiveness, where an agent identifies the worst-performing factory lines and the reasons for any performance drops
• Batch genealogy/recall tracing, where an agent traces batches and customers that received material from a simulated contaminated lot and recommends containment actions
• Predictive maintenance, where an agent runs a diagnostic and classifies machines’ risk levels
• Defect trend investigation, where an agent analyzes and investigates trends in production defects
• Vision-based quality inspection, where an agent determines pass/fail quality of a part based on an image
• Vision quality control (QC) + telemetry root cause, where an agent determines pass/fail quality and correlates the defect with telemetry signals to identify a likely cause
• Full QC pipeline, where an agent runs a complete 4-stage QC pipeline

Agentic AI workloads in financial services

Financial firms—such as banks, asset managers, hedge funds, insurance brokers, and others—are using AI agents to catch errors, automate processes, and uncover fraud. AI agents in finance have the potential to automate significant numbers of tasks and assist humans in even more. According to Google, 53% of financial services organizations are using AI agents to improve customer experiences, reduce time-to-market, identify security threats, and much more.⁵

Our benchmark uses a synthetic dataset to simulate a realistic mid-sized financial services firm. It incorporates eight real-world financial services workflows:

• Filing discrepancy analysis, where an agent reads a filing, cross-references numbers, and writes a risk summary flagging any mismatches in the filing
• Transaction reconciliation, where an agent replays every posted transaction across the organization and flags potential issues
• Portfolio valuation/net asset value (NAV) calculation, where an agent reviews positions a portfolio holds and ranks sector exposures by percent of NAV
• Regulatory ratio compute, where an agent loads bank balance-sheet inputs and flags issues in regulatory ratios
• Peer cohort analysis, where an agent analyzes and ranks a stock ticker against its peers
• Fraud investigation, where an agent investigates potential fraud based on bank transactions and escalates issues to a human investigator
• Long-context filing Q&A, where an agent answers specific questions based on a very large dataset
• Board pack narrative, where an agent generates a board of directors-ready presentation about a financial services org

As one example of what the tokenomics might look like on a Dell PowerEdge XE7740 server, we looked at the cost to run our financial services workload at the 8-GPU configuration—the most expensive, and also the highest-performing. For the cost of the server, we used $263,446, a price Dell provided us on 6/23/26. (Your price may vary depending on your configuration, your seller, and other factors.) We calculated energy costs based on the average price of electricity for U.S. commercial customers, which was $0.1385 per kWh as of the most recent available data at the time of this report.⁷ We did not consider costs for cooling, manageability, networking, storage, software licenses, or data center space.

As a comparison point, we also looked at the cost per million tokens, or $/Mtok, to run the same model we used (Llama 3.1) on a representative cloud service. We chose Amazon Bedrock, an AWS service, which at the time of writing this report is $0.90 for both input and output tokens.⁸

For our financial services workload, as Figure 5 shows, the PowerEdge XE7740 delivered dramatically lower $/Mtok, meaning much better value.

Our calculations are just one example. Your agentic AI workloads will be different from ours, you may run at varying levels of utilization, and your costs for items we excluded in this report (management, cooling, etc.) may vary. What is clear, however, is that the lower initial cost of the cloud doesn’t necessarily translate into savings over time. Running agentic AI workloads on-prem with the Dell PowerEdge XE7740 gives your organization significant potential for savings over time.

To get more technical, the CPU is responsible for orchestration and control flow, deciding what the agent does next, when to call which tool, and how to handle the results. When the agent reaches outside the model to query a database, hit an API, run a piece of code, or read a file, those tool calls and the surrounding I/O run on the CPU. As context accumulates across the agent’s session, the host CPU also orchestrates KV cache offloading—directing inactive cache out of GPU memory into system memory to free high-bandwidth memory (HBM) capacity for active inference—and triggering its reload when the session resumes. Plus, when an agent fans out work to multiple sub-agents running in parallel, the CPU tracks the dependencies between them, schedules their inference calls, and assembles their outputs.

In our testing, CPU usage levels varied significantly depending on what part of the workflow was running, from 7% in times when the agents were largely waiting on the GPUs, to over 99% in times when the agents were actively working or coordinating tasks. The CPU tended to be saturated in the start of a run, where the CPU is doing all of the task assignment and agent orchestration work. (This was especially true for the financial services workload, which starts with a highly CPU-intensive workflow.) The variance in CPU utilization reflects the fact that agentic workflows are made up of many different actions, some of which tax the CPU more than others.

The workload also thoroughly stressed the GPUs. In our financial services testing, the GPUs were consistently saturated, meaning that performance was gated on the GPU. In the manufacturing workflows, the story was similar though more complex due to the nature of the workload. Here, the GPUs burst to near-saturation (approximately 90% utilization) for the first minute of the run, then hovered around 50% utilization. This is because in the manufacturing testing, we split the GPUs between the LLM and the VLM (vision language model), and not all workloads use the VLM. In some cases, therefore, the GPU might reach 50% utilization and be fully subscribed to the work it was allowed to do: either just the LLM, or just the VLM.

Both GPU and CPU performance are vital factors for overall agentic AI throughput. Every time the AI agent calls the model, whether to plan a step, reason about a tool result, or generate a final response, the GPU is running the model that produces the agent’s decisions. In a single agent task, that may happen many times. It can also mean, however, that the GPU spends a meaningful share of the workload waiting on the CPU to feed it the next inference call.

If GPUs can’t run the models quickly enough, users won’t get the outcomes they need. And if the CPU can’t keep up with orchestration, tool execution, and cache management, expensive, difficult-to-access GPUs sit idle.

Conclusion

You may already be considering how your organization can benefit from agentic AI. As AI capabilities continue to increase, the sky’s the limit. Agentic AI has the potential to enhance some types of work and open up brand-new avenues for productivity and innovation. But to take best advantage of these new technologies, you need a computing solution that can support your ambitions both today and tomorrow.

We found that the Dell PowerEdge XE7740 server, featuring Intel Xeon 6 processors and NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, can meet businesses’ real-world agentic needs. On our financial services and manufacturing workloads, it sustained from 8 up to 74 simultaneous AI agents and required substantially lower costs per million tokens than the costs for the same LLM on Amazon Bedrock. Your specific agentic AI plans will determine whether the four-GPU, six-GPU, or eight-GPU configuration is the best fit for you. But one thing is for sure: The Dell PowerEdge XE7740 is ready for agentic AI.