Chip development runs years ahead of release, while marketing copy arrives in the final sprint. Two senior AMD executives used that gap to publicly challenge the claim that Nvidia and Arm designed their latest CPUs specifically for agentic AI.
What AMD Actually Said
Both executives rejected the “purpose-built” framing directly.
Madhu Rangarajan, AMD corporate vice president, was direct about the timeline problem. “We didn’t magically know four years ago that agentic was going to be a thing and purpose-built something for agentic,” she said, according to Tom’s Hardware.
Mike Clark, AMD chief Zen architect, echoed that position without hedging. “I’m not going to mislead and act like I knew agentic computing was coming, or we knew it,” Clark said, as reported by Tom’s Hardware.
Neither executive argued that Zen 6 performs poorly on agentic workloads. The dispute, as AMD frames it, is narrower: whether “purpose-built” accurately describes designs that began before agentic AI emerged as a prominent category.
Why Agentic Workloads Complicate the One-Chip Pitch
The variety of agentic tasks is central to AMD’s portfolio argument.
AMD’s deeper argument is structural. Agentic AI systems orchestrate tasks, execute code, call tools, process data, and coordinate multistep workflows across a wide range of demands.
Rangarajan’s position is that this diversity favors a broad CPU portfolio rather than a single optimized design. AMD says its EPYC lineup spans more than 1,000 configuration instances, giving customers options tuned for different throughput, frequency, memory, and efficiency priorities.
Venice by the Numbers
AMD’s sixth-generation EPYC family is the product at the center of its claims.
The EPYC 9006 family, codenamed Venice and built on Zen 6, tops out at 256 cores and 512 threads. AMD claims it delivers approximately 1.9 times the throughput of the previous-generation Turin EPYC 9005 processors, though that figure is an unverified vendor claim made without a disclosed standard benchmark for agentic workloads.
AMD separately claims its 256-core, 400-watt configuration achieves 3.3 times as many agents per watt as Nvidia’s Vera CPU. It also claims 1.8 times as many agents per dollar as Intel’s sixth-generation Xeon, with the same attribution caveat applying to both comparisons.
Tom’s Hardware tested two Venice configurations: the EPYC 9996 targeting peak throughput, and the frequency-optimized EPYC 9686F evaluated in an AI host-node role comparable to Vera.
What Nvidia and Arm Are Claiming
Both competitors use agentic AI as a central marketing message for their data-center CPUs.
Nvidia describes Vera as a CPU purpose-built for agentic workloads, including sandboxing, tool use, code execution, and orchestration. Arm markets its AGI CPU as designed for agentic coordination across agents and accelerators. Custom silicon and AI chips are increasingly central to competitive positioning across the industry.
Arm’s published benchmark comparisons pit its AGI CPU against AMD’s Turin generation, not against the newer Venice processors. Arm’s product strategy also spans multiple core configurations, which itself reflects the kind of portfolio approach AMD is advocating.
AMD’s criticism targets the retrospective framing, not whether Nvidia or Arm hardware can run agentic workloads.
The Roadmap Makes the Argument
AMD’s planned Venice derivatives reinforce its case for a portfolio over a single agentic profile.
AMD is developing additional Venice-based products, according to reporting, including Venice-X with stacked cache aimed at high-performance computing and Verano with LPDDR5X memory targeting AI host-node deployments. Each variant addresses a different set of workload constraints within the same architectural family, lending practical weight to the company’s argument that agentic infrastructure will not reduce to a single CPU profile.
