CPUs regain prominence in agentic AI

The rise of AI agents restores importance to server processors. AMD bets on its EPYC and a more balanced architecture between CPU and GPU to meet these new uses.

Agentic AI is already changing the distribution of roles in data centers: GPUs handle heavy computation, but CPUs regain a central role to orchestrate tools, APIs, databases, and execution environments. AMD, through its server business, makes this a major commercial and technical argument as agents move from expert tinkering to broader uses. According to Dereck Dicker, AMD's corporate vice president, this evolution pushes infrastructure toward an almost one-to-one ratio between GPU and CPU. And, in this context, the old server processor has not said its last word.

Quand les agents multiplient les tâches, le CPU redevient stratégique

In the early days of generative AI, in 2022, the CPU mainly served as an auxiliary: it prepared data, managed execution, and ran the software around the model. With agents, the mechanics become more complex. A request no longer triggers just an inference on the GPU, but a chain of actions, sometimes across multiple systems, with planning, access to external services, and sequential execution. Dereck Dicker sums up this shift as follows: "It is the shape of the agentic workloads that makes it super unique, where you’re no longer having a single request going into a GPU that runs inference and gives you a response,"

The core issue is here: the GPU computes, but the CPU coordinates. This orchestration becomes even more visible as agents must manage concurrent tasks, calls to tools, and multiple environments. For data centers, this means more CPUs per rack and fewer configurations where multiple GPUs rely on a single host processor. AMD believes this evolution renders obsolete the "multiple-GPUs-per-CPU" architectures that dominated until now.

AMD pousse EPYC comme brique d’infrastructure, pas seulement comme puce

To address this market, AMD introduced Helion, an open rack-scale blueprint designed for agentic AI infrastructure. In CPU-only version, it relies on 6th generation EPYC processors and can scale up to 32,768 CPU cores to run large-scale agents. Another variant pairs these EPYC with Instinct MI350P GPUs to run, in the same rack, agent workflows and model inference.

The Venice family, unveiled at the Advancing AI event in July, also serves this strategy. AMD then raised its projection for the server CPU segment from about $25 billion today to over $200 billion by 2030. The figure is ambitious, but it mainly says one thing: AMD no longer just sells processors; the company sells a way to organize data centers. It’s less glamorous than a chatbot that replies in verse, but much more useful for operators.

Deux familles de processeurs pour deux usages bien distincts

With Venice, AMD segments its offer into two socket-level families. SP7 targets the most demanding workloads, with maximum core density. The flagship model announced reaches 256 cores and 512 threads per socket. SP8, on the other hand, favors balance between compute, memory, and capacity, with a range from 8 to 128 cores. This version targets databases, virtualization, and classic enterprise applications.

This separation meets a concrete need: not all servers should become multi-core monsters. Some must mainly handle load, others offer memory capacity and flexibility. AMD thus tries to cover both agentic AI and traditional IT foundations, to keep its clients within the same hardware family when they modernize their fleet.

La bataille du serveur x86 reste ouverte, avec Intel et NVIDIA en face

In the x86 server market, AMD is still gaining ground on Intel. Mercury Research estimates that in Q2 2026, AMD held 34.5% of x86 server processor shipments, 130 basis points more than the previous quarter, while Intel fell to 65.5% [to be verified]. Looking only at EPYC and Xeon SP data center chips, AMD rises to 46.4% versus 53.6% for Intel [to be verified].

Against NVIDIA, the debate changes nature. NVIDIA pushes its own Arm CPUs for data centers, while its historic domain remains the GPU. Mercury Research estimated Arm server processor shipments at 13.6% in Q2 2026 [to be verified]. AMD defends the advantage of x86 architecture: enterprise software already runs natively on it, simplifying migrations. NVIDIA responds with Vera, its Arm data center CPU, which some cloud provider tests present as very performant on agentic workloads.

Les benchmarks comptent, mais les datacenters regardent surtout le coût réel

Published comparisons between AMD and NVIDIA should be handled with caution. AMD claims its 256-core EPYC 9996 would offer 2.24 times the aggregate throughput of NVIDIA’s 88-core Vera in SPEC CPU 2026, with estimated scores of 2,070 versus 925 [to be verified]. On a closer configuration, AMD also says its 96-core Venice would deliver about 20% more performance per core than Vera [to be verified].

On its side, cloud provider Modal indicated its initial tests showed Vera performing twice as much agentic work as Intel and AMD 5th generation CPUs already deployed there. The company tested 50 Terminal-Bench tasks, covering code compilation, file search, and Git commands, in several containers run in parallel. Results would be better under heavy load thanks to Vera’s memory bandwidth, allowing more work to be loaded on a host without degrading performance as much. AMD counters with its own estimates on Venice. The real winners will likely be those running these chips with the same software, on real tasks. Otherwise, everyone ends up first in their favorite category.

La promesse la plus concrète reste la consolidation des vieux serveurs

Beyond the benchmark race, AMD emphasizes a very pragmatic point: freeing up space, energy, and cooling in already saturated data centers. Dicker explains that many infrastructures lack available space, and consolidation is often the fastest way to free budget and racks for AI. With its 6th generation EPYC, AMD claims that modernizing servers installed about five years ago could achieve a consolidation ratio of 13 to 1.

In other words, thirteen old machines could be replaced by a single newer server for certain workloads. The potential gain is not limited to the number of chassis: fewer servers also means less cabling, less cooling, and often fewer software licenses when charged per core. In an industry where every watt counts, this promise can weigh as much as a peak performance on a public benchmark.

By the numbers

  • 32,768 CPU cores — announced capacity for Helion in CPU-only configuration.
  • 256 cores and 512 threads — flagship SP7 configuration of Venice.
  • 34.5% — AMD’s share of x86 server shipments in Q2 2026 [to be verified].
  • 13 to 1 — consolidation ratio announced by AMD on servers about five years old.
  • 2,070 versus 925 — estimated AMD and NVIDIA scores in SPEC CPU 2026 [to be verified].

To read