The GPU grabbed the headlines. The CPU may end up doing the heavy lifting. As agentic AI workflows move from demos into enterprise applications, the familiar one-to-one ratio of CPU to GPU is giving way to something strikingly different: six, even eight CPUs for every single GPU.
"Given the amount of work that those agents are doing on enterprise workloads, we're saying upwards of six to eight to one," said Lucas Melo, a cloud solutions architect at Intel, in a video interview recorded at the April 2026 Nutanix .NEXT event in Chicago.
"Six CPUs to one GPU to keep that GPU busy because there's so much time being done on the orchestration side of that workflow. And again, you paid a lot for those GPUs, you want to make sure they're busy," he told The Forecast.
The shift points to a deeper transformation. Agentic AI, the class of artificial intelligence systems that autonomously plan and execute multi-step tasks, is embedding itself into enterprise applications in ways that look less like a chatbot and more like a new microservices architecture. Melo calls it "agentic microservices," a revisiting of the whole microservice framework using agents as the building blocks.
Agentic AI infrastructure is the compute, networking and software stack required to run AI agents that autonomously orchestrate multi-step enterprise workflows, blending GPU-based inferencing with CPU-heavy tasks like data transformation, code generation and file system operations.
The answer lies in what agents actually do, explained Melo. He said Intel's testing reveals that 90% of the agentic workflow is standard enterprise workload: data transformation, code generation, code compilation, code testing and file system operations. Only a sliver requires GPU inferencing on a large language model. The rest runs on the CPU, the same processor that has powered enterprise applications for decades.
"Those workloads have always run best on CPUs," Melo said.
That finding carries an infrastructure implication. If the CPU tier is undersized, GPUs sit idle. If networking is an afterthought, agents stall. The AI system of the next few years will not be a single box, Melo suggested, but a distributed system with GPU racks for dense model compute and CPU racks for orchestration and tool execution.
Melo pushed back on the myth that GenAI equals GPUs. It’s not necessarily true, he said. The better approach starts with understanding the real business challenge, then finding the technology that leads to the intended outcomes.
Melo asked: "What are you really trying to solve for? Is it getting insights from your HR data? Is it minimizing that customer support time?”
He said starting with those questions aligns needs to AI use cases, then SLAs and eventually technologies.
The framework mirrors advice from across the industry. Nutanix and Intel have flipped the script, starting with the business challenge and ending with the best technology to get the job done, rather than buying hardware first.
He addressed the Intel-Nutanix partnership, which has been producing concrete results for some time. Intel has enabled its AI capabilities through Xeon processors and Advanced Matrix Extensions (AMX) via Nutanix Enterprise AI, allowing enterprises to run AI inference directly on CPUs without dedicated GPUs. The two companies are now layering Intel AI for Enterprise RAG, a chatbot retrieval-augmented generation solution.
The RAG solution lets departments feed PDFs, Word documents and PowerPoint decks into a private chatbot. Instead of searching a SharePoint share, employees ask the LLM questions and get answers grounded in their own documentation. Everything stays private and secure, hosted on dedicated Nutanix infrastructure.
"Think about the department RAG solution and Nutanix has then enabled that through their core start program," Melo said. "So it's a simple, easy-to-implement solution where Nutanix and Intel can come in and help customers gain insights from their data using RAG on their existing platforms very easily."
Melo points to the developer local environment as a preview of what is coming to the enterprise. Frameworks running on home PCs are generating energy and experimentation that will eventually migrate into corporate infrastructure. When enterprises start automating applications and workloads with agentic workflows that give agents decision-making authority, the change will accelerate.
"You got to put your DevOps hat on, on how you deploy, redeploy, how you operate," Melo said. "So a lot of that's going to be happening too, but it's going to be exciting times."
Related:
Video transcript:
Lucas Melo: Really what's going to be huge in the turning point for enterprises, is that Agentic AI workflow getting embedded into enterprise workloads and actually being part of enterprise applications. I like to call it the Agentic microservices. So kind of revisiting the whole microservice architecture by using agents. And sure, there are some new components there with AI tool calling and there's a step where the agent is doing AI inferencing on LLM, but we already see doing our test that 90% of the agentic workflow is standard enterprise workload. Data transformation, like I said before, code generation, code compilation, code testing, file system operations. We're even starting to see where the ratio of CPU to GPU needs to change. Today it's mostly one-to-one, you got a whole CPU and a GPU. But given the amount of work that those agents are doing on enterprise workloads, we're saying upwards of six to eight to one. So eight, six CPUs to one GPU to keep that GPU busy because there's so much time being done on the orchestration side of that workflow. And again, you paid a lot for those GPUs, you want to make sure they're busy.
Ken Kaplan: Talk about the future you see. Is it going to happen quickly in the whole world of agents, more inferencing happening in more places?
Lucas Melo: Yeah. So if you're following what's happening in the kind of developer local environment, like home PC environment with open claw, a lot of energy behind that, a lot of cool stuff happening. I think that type of framework would eventually start to get implemented on the enterprise. Like I said before, I think that's going to be when we're going to see a lot of change on enterprise when we're starting to automate enterprise applications and workloads with agentic workflows, providing agents agency to make decisions is going to get extremely interesting. Of course, you got to then on the enterprise space, monitor all that. You got to put your DevOps hat on, on how you deploy, redeploy, how you operate. So a lot of that's going to be happening too, but it's going to be exciting times.
Ken Kaplan: You seeing customers building on existing technologies, adding new things? What is it that they want to do with what they have and where they want to go?
Lucas Melo: Yeah, there's definitely a challenge with the hype that AI or GenAI equals GPUs and a heavy focus on thinking about technology first. So what we're trying to get customers to do is to, one, focus on the business challenge. So what are you really trying to solve for? Is it getting insights from your HR data? Is it minimizing that customer support time? So starting with that in mind, that then aligns to AI use cases, eventually SLAs, and then eventually technologies. So we try to move the conversation from what models should we use, what size of models, what tokens per second should we be trying to achieve to what is the business use case you're trying to solve for? And then we can look at what you have available today that can support that. So the myth that GenAI equals GPUs, that's not true. So depending on the SLAs and the business use case that you're trying to solve for, let's align your infrastructure, the technologies to go support that.
Ken Kaplan: Can you talk about how Intel is working with Nutanix?
Lucas Melo: Yeah, and I guess I'll focus on the latest integration partnership that we've been doing with Nutanix. I mean, it's been there for a while, but we have recently, of course, enabled our AI capabilities through Xeon and AMX through Nutanix Enterprise AI and now been working with Nutanix on then layering on our Intel AI for enterprise RAG, which is a chatbot RAG solution so that it's an easy start for customers. Think about the department RAG solution and Nutanix has then enabled that through their core start program. So it's a simple, easy to implement solution where Nutanix and Intel can come in and help customers gain insights from their data using RAG on their existing platforms very easily. Again, the idea is to showcase what's possible.
Ken Kaplan: And what kind of data are you getting, what is being exposed and what can you do with the data?
Lucas Melo: It can be any departmental data, so PDFs, Word docs, PowerPoints, so just infrastructure guides. So just think of, you want to gain insights from your departmental data, can easily deploy a solution, kind of adding your documentation to the RAG and instead of searching for that data in a SharePoint share, you can now ask the LLM questions and fully private secure because it's all being hosted by Nutanix on your infrastructure, dedicated infrastructure, but it could also be Nutanix cloud infrastructure, but it's still your infrastructure.
Jason Lopez is executive producer of Tech Barometer, the podcast outlet for The Forecast. He’s the founder of Connected Social Media. Previously, he was executive producer at PodTech and a reporter at NPR.
Ken Kaplan contributed to this story. He is Editor in Chief for The Forecast by Nutanix. Find him on X @kenekaplan and LinkedIn.
© 2026 Nutanix, Inc. All rights reserved. For additional information and important legal disclaimers, please go here.