Here’s the hard truth about agentic CPUs
They’re not new. They’re not magical. And if you still believe there’s one chip to run every AI agent—well, let’s talk about your router.”
There’s been a lot of noise this year about CPUs built for agents—Arm, Nvidia, AWS, AMD, Intel—they’re all shouting about it. But take a closer look. You’ll see the same pattern we’ve seen for twenty years: general-purpose silicon, repackaged with a shiny sticker that says “AI.” It’s not that agents don’t need CPUs—they do. It’s just that no single silicon recipe fits every agent workload. Some agents run like sprinters, waiting for milliseconds to move data between GPU and CPU. Others run like marathons—pumping through thousands of concurrent tasks with little patience for high latency, only asking for raw throughput.
The truth isn’t boring. It’s liberating: there never has been, and never will be, one CPU to rule them all. The so-called “agentic CPU” isn’t a new architecture. It’s an optimization problem dressed in the latest marketing team uniform. And if you think Arm, Nvidia, or Intel have converged on a single answer, your benchmarks are hiding something.
Let’s walk through the noise and see what’s really cooking under the hood.
The marketing machine moves at light speed
At Computex 2026, the term agentic CPU got its first full-scale dress rehearsal. Arm unveiled Neoverse V3 as the “AGI CPU.” Nvidia CEO Jensen Huang introduced Vera as a “CPU for agents” and doubled down on latency-busting logic to keep its GPUs fed. AWS didn’t miss a beat—Graviton 5 quietly doubled down on the same Neoverse V3 kernel but with more cores and less power to spare.
None of this is inherently wrong. CPUs do run agents. The real question, though, is what those agents are. If they’re all running the same thing—a single AI model talking to one app—the answer is simple. But reality isn’t that clean.
Agents are bridge workloads: they sit between the model and your ERP, CRM, or document summarizer. Some need high-throughput to chew through thousands of user queries a minute; others need low latency because they’re chained into real-time interactive flows. The same chip that handles both? Not unless it has an internal identity crisis.
Look at Nvidia’s Vera: 88 cores, aggressive single-threaded performance, and a memory subsystem that looks like it was designed by a GPU enthusiast with too much bandwidth to spend. That’s great for keeping GPU latency low, because Huang is right—those GPUs cost serious money, and you can’t let them sit idle waiting for a CPU to catch up. But try to run ten thousand concurrent low-latency agents on Vera, and you’ll hit wall-clock saturation fast.
That’s not a flaw. It’s a trade-off, and one the design team made deliberately.
Latency and throughput aren’t just buzzwords—they’re fundamentally different goals
Nvidia’s approach is clear: treat the CPU like an accelerator for the GPU. High single-threaded performance, high memory bandwidth, aggressive prefetching—Vera looks less like a general-purpose CPU and more like an extension of the GPU’s memory subsystem. As Huang put it during his GTC Taiwan keynote, “There will be billions of agents and these agents are going to be using the CPUs with very little patience because the cost of the GPUs they sit next to is too high.” In other words: you can’t afford idle cycles on that $30,000 H200 board, so optimize the interface.
Arm’s take is subtly different. The AGI CPU (another Neoverse V3 variant) strips away things most agents probably don’t need: no simultaneous multithreading, minimal vector extensions, reduced cache latency optimizations in favor of sheer core count. The result? A 136-core monster built for throughput, not latency.
Amazon’s Graviton 5 is the most revealing because it’s essentially a Graviton chip that went to CPU boot camp: 192 cores, stripped-down instruction set, and power-performing like it’s trying to fit in a tablet while delivering rack-scale compute.
But here’s where it gets messy: some workloads thrive on high frequency and low core count. Others need hundreds of cores even if each one runs slower. AMD’s upcoming Venice Epyc—256 Zen4 cores, targeting a 3.3× throughput advantage over Vera per rack—is betting the farm on concurrency, not latency.
If all agentic workloads were identical, there’d be one winning design by now. But they’re not—and that’s the whole point.
Benchmarks tell you more about the test than the chip
Remember those benchmarks Phoronix ran on Vera last month? The ones where Nvidia claimed a 10% edge over AMD’s Epyc 9575F and a 55% lead over Intel’s Xeon 6980P? They’re impressive—if you only run the workloads Nvidia chose.
The truth is, those benchmarks only tell us how Vera behaves in Nvidia’s selected scenarios. They’re useful, but they’re not comprehensive. And once you dig into the details—what apps were tested, what tuning flags were enabled—the story changes.
Some agents run database-heavy workloads where memory latency dominates. Others are compute-bound, running long simulations on unstructured text or JSON payloads. Still others need high concurrency to service thousands of simultaneous users.
That’s why AMD is pushing concurrency metrics. Why Intel showed off a 36,864-core rack at Computex—because for some use cases, raw core count is the win.
And here’s the kicker: no single company can test every possible agent workload. They test what matters to their customers, which means your definition of “agentic CPU” may differ wildly from mine.
One chip won’t run them all
Look back at the history of datacenter chips: AMD’s EPYC, Intel’s Xeon, Arm’s Neoverse, Nvidia’s Grace—none of them were “one chip to rule them all.” They all optimized for different priorities. Some aimed at virtualization density. Others targeted database performance or machine learning inference.
What changes now is the language, not the reality. “AI agents” have become the new buzzword, and every silicon vendor is rebranding their latest chip under that banner. The reality? Agents are just another workload category with wildly divergent needs.
You’ll need a low-latency, high-bandwidth CPU if your agents talk to GPUs constantly. You might prefer high core count and thread density for background summarization or queue processing. Or you could lean into frequency and simplicity if your agents run mostly locally on low-power endpoints.
The choice isn’t about which company wins. It’s about matching the silicon to your use case—just like it’s always been.
This is why AMD’s defensive move last month made sense: they argued that throughput matters more than latency for most large-scale deployments. That’s a perfectly valid stance—just not one Nvidia would ever make.
And that’s okay. In fact, it’s good. A healthy market needs disagreement.
The real danger isn’t that vendors have different chips. It’s that customers start believing there’s one correct answer. That’s how expensive mistakes happen.
So next time you hear “agentic CPU,” ask: What kind of agent? What workload? What’s the bottleneck?
The answer will tell you more than any chip spec sheet ever could.