That is a useful distinction—but not proof that rival engineering claims are false. The debate deserves more scrutiny than simply deciding which company has attached the most fashionable adjective to a CPU.
What AMD is arguing
Rangarajan’s sharpest line was that AMD “didn’t magically know four years ago that agentic was going to be a thing.” Clark separately acknowledged that AMD had not foreseen agentic computing, while maintaining that the company expected CPUs to retain an important role in AI software. AMD’s position is that varied agent tasks favor a broad processor portfolio, rather than one universal hardware profile.
The product examples in Tom’s Hardware illustrate that strategy: a 256-core, 512-thread EPYC 9996 for throughput and a frequency-focused EPYC 9686F for AI-host work. The publication also reports planned 2027 additions: cache-focused Venice-X for HPC and LPDDR5X-equipped Verano for AI hosting. Those are reported positioning and roadmap details, not a universal performance verdict.
Nvidia and Arm really are making the agentic pitch
Nvidia’s own Vera page explicitly describes the processor as purpose-built to accelerate agentic workloads. Its examples include code execution, sandboxing, tool use, data processing and orchestration. Nvidia also positions Vera as both an accelerator host and a standalone CPU—not merely an accessory to a GPU.
There is concrete hardware behind that pitch. Nvidia lists 88 Olympus cores, 176 threads through Spatial Multithreading and up to 1.2 TB/s of LPDDR5X memory bandwidth. These specifications establish what Nvidia emphasizes, not whether Vera wins every agent workload.
Arm likewise describes its AGI CPU as designed for agentic AI, based on Neoverse CSS V3 and intended to coordinate agents and accelerators. But its official specifications add an important qualification to the “single-SKU” characterization: Arm lists 136-core, 128-core and 64-core variants, respectively positioned around maximum core count, total cost of ownership and memory bandwidth per core. A focused product family is not literally a single configuration.
That qualification matters. AMD can make a case for portfolio breadth without treating its competitors as offering no configuration choices.
“Purpose-built” does not settle the design timeline
The competing statements address different questions:
- AMD’s executives: Did the company anticipate today’s agentic trend when its architecture program began?
- Nvidia and Arm’s product descriptions: What workloads do the companies say their processors are designed to serve?
Those statements are not necessarily mutually exclusive. This is an interpretation of the competing claims, not an independently established history of either rival’s design decisions. The available statements do not establish when agent-specific requirements entered each development program.
In other words, a marketing label is not a time machine. But neither is skepticism about that label a substitute for examining the engineering.
What enterprise buyers should take from it
The practical response is to turn the vendors’ own workload descriptions into a test plan. Nvidia emphasizes tool execution and sandboxing; Arm emphasizes coordination and rack-level deployment. Those are different evaluation dimensions, even when both sit under the agentic-AI umbrella.
As an evaluation framework—not a benchmark finding—IT teams should ask:
- What completes the business task? Measure an entire workflow, not only its fastest component.
- What happens under concurrency? Test the intended number of simultaneous agents and isolated environments.
- Where does time go? Separate model-response time from tool execution, data access and orchestration.
- What is the deployment constraint? Compare completed work within the actual power, memory, rack-space and cost limits.
The strongest conclusion is narrower than either side’s sales pitch. AMD’s candor about what it did not predict is welcome; its claim to have the best answer still requires workload-specific evidence. Nvidia and Arm, meanwhile, offer concrete design arguments that deserve testing rather than dismissal. The useful contest is not who first said “agentic.” It is who completes the required work most effectively.