The Agentic Era Runs on a CPU Foundation — and It's Called AMD EPYC™ 9006 Series Server CPUs

Jul 23, 2026

Curved teal and orange light trails with bokeh particles on a dark background, symbolizing connectivity

AI has learned to act. The data centers that win the next decade will be the ones engineered for every stage of the agentic workflow — and that foundation starts with the processor.

The shift from AI that answers to AI that acts

For the past few years, the strategic conversation about AI has revolved around a single question: how good is the model? That era is closing. The frontier has moved from AI that answers to AI that acts — systems that plan toward a goal, call tools and APIs, write and execute code, retrieve and reason over enterprise data, and carry out multi-step work with minimal human oversight. This is agentic AI, and it rewrites the economics of the data center — and with it, the assumptions behind every infrastructure roadmap. 

Here is what most infrastructure strategies still miss. An agent completing a task is not one model running on one accelerator. It is an orchestrated workflow — a planning step, multiple tool calls, retrieval against your databases, code execution, accelerator inference, and the enterprise services the agent acts upon. And as agentic systems scale from one agent to thousands running concurrently, the work that happens outside the accelerator grows fastest of all. Once you trace where the time and the cost go, the surprising truth emerges: many of the tasks that make up an agentic workflow run on the CPU. 

That reframes the build-out decision facing every CTO and CIO right now. The agentic era is not only a GPU story. It is, first and foundationally, a CPU story. The AMD EPYC™ 9006 Series Server CPUs were engineered for exactly this moment: not as a single CPU chasing a single workload, but as a portfolio built to power every stage of the agentic workflow, and the broad cloud, enterprise, and HPC infrastructure that surrounds it. Put simply, they are the world’s best server CPU Portfolio for the agentic era1.

Why AMD Wins Across CPU Profiles

Leadership Per-Core Performance. Maximum Agentic Capacity. Efficient General-Purpose Compute.

If the agentic workflow has many CPU-dependent jobs, the strategic insight is that those jobs do not all ask the same thing of a processor. They pull in different directions, and optimizing one generic CPU for all of them means compromising on each. The AMD approach is to recognize three distinct CPU jobs in the agentic workflow — three profiles — and to deliver a purpose-built answer for each.

AMD EPYC™ 9006 Series Server CPUs combine leadership per core performance with the flexibility to scale to much higher core and thread counts for greater aggregate agent capacity, plus the general-purpose performance to run the tools and services agents depend on. 

The Agent Sandbox CPU, powered by high-core-count AMD EPYC 9006 Server CPUs, runs the agents themselves — coordinating many concurrent agents, executing the code and tasks they spawn, managing system state and memory, and handling the control-path logic and I/O that all happen outside the GPU. The demand is throughput, concurrency, and responsiveness. AMD delivers leadership per-core performance to keep each agent fast, along with the highest thread density2, scaling up to 256 cores and 512 threads, to maximize agent capacity and enable the most agents per watt, per dollar, and per rack3,4,5.

The AI Host Node CPU, powered by high-frequency AMD EPYC 9006 Server CPUs, hosts and feeds the accelerators. An idle accelerator is the most expensive thing in your data center, and the host CPU is what prevents that idleness. AMD delivers a high-frequency host node at up to 5 GHz with leadership per-core performance, an industry-leading PCIe® Gen 6 I/O subsystem for 2x the bandwidth per lane versus the prior generation, and up to 1.6 TB/s of per-socket memory bandwidth with MRDIMM at 12800 MT/s — enough to keep accelerators fed and to support large-context agentic reasoning. 

The General-Purpose CPU, powered by the broad AMD EPYC 9006 Series Server CPU portfolio, runs the tools and enterprise workloads agents rely on — databases, storage, application services, and AI-support workloads — and delivers strong CPU inference (tokens/s) for real-time small and medium language models (SLMs, MLMs), as well as batch mode. For these general-purpose workloads it offers best performance6, power efficiency, and thread density leadership. 

Multiple jobs, multiple profiles, one purpose-built portfolio— sharing a common software ecosystem and a hardened security foundation. For agentic work that touches sensitive data, AMD EPYC 9006 Series Server CPUs anchor a trusted domain in the CPU and extend it through AMD SEV with Trusted I/O to TDISP-enabled GPUs, networking, and storage, backed by a mature and broad confidential-computing ecosystem. That matters because agents increasingly act on regulated, proprietary, and confidential data, and the trust boundary has to follow the work across every stage. That is the architecture of the agentic data center, and the confidential computing capability behind the AMD EPYC 9006 Series Server CPUs portfolio. 

The AMD EPYC 9006 Series Server CPUs 

The Broadest Server CPU Portfolio for Agentic AI7

Rather than asking one processor to be everything, the AMD EPYC 9006 Server CPUs portfolio is deliberately built around four processor families: AMD EPYC 9006 SP7, 9006 SP8, 9006X SP7, and 9006 LP Server CPUs. Each family targets a specific part of the agentic workflow and includes multiple SKUs for different performance needs and infrastructure constraints. Crucially, each processor family also earns its place across the broader data center, because the same strengths that serve an agentic job serve a class of conventional workloads too. Matching each processor family to the work it serves best reduces the compromises of a one-size-fits-all approach across the agentic workflow: density where you need to run the most agents, frequency and bandwidth where you need to feed accelerators, and efficiency where you need to run the enterprise. 

Our partners are building systems that enterprises will run agentic AI on — and they see the same shift. Here is what our partners see as the agentic era takes shape. 

Want to go deeper? Read more about the portfolio in detail — AMD EPYC 9006 SP7, 9006 SP8, 9006X SP7, and 9006 LP Server CPUs— and how each processor family is built for the specific agentic and broader workloads it serves. Read it here

Footnotes

Agent counts are estimates derived from available CPU thread resources used as a proxy under a consistent theoretical workload. Actual agent capacity and throughput will vary based on workload, model, memory, software, orchestration, and system configuration.

  1. EPYC-067: Based on AMD internal analysis and modeling of representative infrastructure workloads underlying agentic AI systems as of June, 2026. Performance comparisons reflect estimated rack‑scale throughput within a modeled 100kW deployment and include proxy workloads spanning web services, data infrastructure, and orchestration layers. Results are derived from a combination of AMD testing, third‑party data, and projections, and may vary based on system configuration, software stack, and workload mix. See AMD Agentic AI rack‑scale performance blog and AMD methodology description for details on assumptions, workloads, and measurement approach. 
  2. EPYC-025D: As of July 2026, 6th Gen EPYC 9996 has 256 cores and 512 threads with SMT enabled which is higher than any other publicly disclosed 1P CPU 
  3. 9xx6-014: Comparison based on published Top-of-stack core counts and CPU W, Default CPU Power for 6th Gen EPYC, and TDPs for 5th Gen EPYC, Intel® Xeon®, and Nvidia Vera for estimated Highest Agents / CPU W across AMD EPYC™ 9956 (400W Default CPU Power), AMD EPYC™ 9965 (500W TDP), Nvidia Vera (450W TDP), ARM AGI (300W TDP), and Intel® Xeon® 6980P (500W TDP) powered servers as of 7/22/2026. 2 threads per core (SMT). 1 thread per core for ARM AGI. 
    Starting with the 6th Gen AMD EPYC™ server processor family, AMD uses Default CPU Power to describe processor power consumption, succeeding AMD's historical TDP reference. Default CPU Power reflects total power consumed across the processor's compute and I/O dies for the stated performance target. Default CPU Power and TDP may both serve as processor power references for product comparison, platform planning, and performance-per-watt analysis. Intel Xeon TDP from ark.intel.com. Nvidia Vera TDP from https://developer.nvidia.com/blog/nvidia-vera-cpu-sets-a-new-standard-for-agentic-workloads-in-ai-factories/. ARM AGI Specifications from https://www.arm.com/products/cloud-datacenter/arm-agi-cpu#Specifications  
  4. 9xx6-013: Comparison based on published Top-of-stack core counts and 1Ku pricing for estimated highest Agents / CPU $ and threads / CPU $ AMD EPYC™ 9006 (512 threads), AMD EPYC™ 9005 (384 threads), and Intel® Xeon® 6 (256 threads) SKUs as of 7/22/2026. Threads derived as 2 threads per core (SMT). Intel and Xeon are trademarks of Intel Corporation or its subsidiaries. Source: https://www.amd.com/content/dam/amd/en/documents/solutions/ai/methodology-description.pdf   
  5. 9xx6-012: Based on estimated performance data for Nvidia, Intel®, and AMD EPYC™ Server Processors for Agentic AI, the AMD EPYC 9996 provides the most agents per rack at a 100Kw power envelope per rack. 
    Compared to the cores per rack of Nvidia Vera (88c) powered server racks: 
    - The AMD EPYC 9996 (256C) provides 2.08x the cores (and threads with SMT) per rack 
    - The AMD EPYC 9965 (192C) provides 1.86x the cores (and threads with SMT) per rack 
    - The Intel Xeon 6980P (128C) provides 1.24x the cores (and threads with SMT) per rack  
    Source: https://www.amd.com/content/dam/amd/en/documents/solutions/ai/methodology-description.pdf   
  6. 9xx6-011: SPECrate®2026_int_base comparison based on AMD internal estimates for top of stack 2P 6th Generation EPYC CPU and 5th Gen EPYC measurements and pricing as of 07/22/2026. Preliminary performance estimates based on AMD engineering projections or measurements for respective product launch dates and subject to change. 2P AMD EPYC 9996, 512C total, 600W Default CPU Power, $14904, 16x64GB 12800 MT/s MRDIMM, AOCC 6.0, performance determinism, est. SPECrate®2026_int_base 2250, est. SPECrate®2026_int_base/CPU W 1.875, est. SPECrate®2026_int_base/CPU $ 0.151 2P AMD EPYC 9965, 384C total, 500W TDP, $11988, 12x64GB DDR5-6400, AOCC 5.1, performance determinism, est. SPECrate®2026_int_base 1270, est. SPECrate®2026_int_base/CPU W 1.270, est. SPECrate®2026_int_base/CPU $ 0.106, May 05, 2026, https://spec.org/cpu2026/results/res2026q2/cpu2026-20260210-00037.html 2P Intel Xeon 6980P, 256C total, 500W TDP, $13955, 24x64GB MRDIMM at 8800 MT/s, SPECrate®2026_int_base 1120, SPECrate®2026_int_base / C 8.750, SPECrate®2026_int_base / CPU W 1.120, SPECrate®2026_int_base  / CPU $ 0.080, June, 4, 2026, https://spec.org/cpu2026/results/res2026q2/cpu2026-20260519-00255.html SPEC® and SPECint® are registered trademarks of Standard Performance Evaluation Corporation. Learn more at spec.org. Starting with the 6th Gen AMD EPYC™ server processor family, AMD uses Default CPU Power to describe processor power consumption, succeeding AMD's historical TDP reference. Default CPU Power reflects total power consumed across the processor's compute and I/O dies for the stated performance target. Default CPU Power and TDP may both serve as processor power references for product comparison, platform planning, and performance-per-watt analysis. 
  7. EPYC-068 - The AMD EPYC server CPU portfolio spans the industry’s broadest ranges of data center deployments, from general-purpose enterprise, cloud, telecom, SMB, and HPC systems to emerging AI environments including sandboxed agentic AI deployments and GPU head node servers. AMD EPYC 6th Generation platforms extend this breadth by uniquely combining high core and thread density of up to 512 threads, advanced memory bandwidth of up to 16 channels of 12.8 GT/s MRDIMM support, next-generation PCIe® Gen 6 connectivity, and select SKUs with boost frequencies up to 5 GHz.  
Share:

Article By


Corporate VP, Server Product
Madhu Rangarajan is a seasoned technology executive with over 25 years of experience in the server industry.

Contributors


Related Blogs