- AMD CEO Lisa Su stated on the Q1 2026 earnings call that agentic AI workloads are shifting the traditional CPU-to-GPU ratio from 1:4 or 1:8 toward nearly 1:1, significantly increasing CPU demand.
- Agentic AI’s orchestration, planning and tool execution layers are CPU-bound, making high-core-density processors like AMD EPYC a critical part of the stack, not just the GPUs.
- AMD’s new Instinct MI350P PCIe accelerator targets enterprise AI inference within existing infrastructure, reflecting a broader strategy around open, balanced hybrid AI deployments.
Agentic AI is staging a quiet comeback for the server CPU. AMD CEO Lisa Su told investors on the Q1 2026 earnings call that the long-established GPU-heavy compute ratio is shifting, with agentic workloads pushing deployments toward as many CPUs as GPUs. For enterprise infrastructure teams, that’s a structural change with real budget consequences.
Agentic AI Rewrites the Enterprise Compute Equation
For years, the dominant data center model paired one CPU with four or eight GPUs, treating the processor as little more than a traffic controller for GPU workloads. Agentic AI breaks that assumption. Unlike a chatbot responding to a single prompt, an agentic system plans multi-step tasks, calls external tools, queries databases and routes sub-tasks to different models in a continuous loop. That control plane is CPU-bound, and an undersized CPU tier means expensive accelerators sit idle waiting for it to catch up.
AMD also introduced the Instinct MI350P PCIe accelerator this week: a dual-slot, air-cooled card designed to bring AI inference into existing server environments without requiring new racks, upgraded power or liquid cooling. It targets agentic AI and Retrieval-Augmented Generation (RAG) pipelines, the kind of workloads now appearing across regulated enterprise deployments.
The CPU’s Resurgent Role in Agentic Workflows
The workload profile of an agentic system looks less like a single neural network pass and more like a distributed operating system. The GPU handles heavy numerical computation during inference, but everything around it runs on the CPU. That includes:
- Task Planning and Orchestration: Breaking complex goals into executable steps and managing the flow between agents.
- Memory Management: Handling vector database queries, RAG pipelines and dynamic multi-turn context.
- Tool Execution: Invoking APIs, web searches and code execution environments.
- Model Routing: Directing sub-tasks to specialised large language models or other AI models.
- Data Pre- and Post-Processing: Preparing inputs for GPU inference and handling outputs.
- Monitoring and Security: Enforcing policies, logging activity and securing execution.
Each of these steps runs serially before or after a GPU call, which means concurrency scales with CPU thread count, not just GPU FLOPS. Undersizing the CPU layer creates a bottleneck that no amount of additional GPU capacity can fix. For teams exploring agentic deployment frameworks, this compute balance is a first-principles constraint worth addressing early.
AMD’s EPYC and Instinct Power the Agentic Era
AMD’s EPYC 9005 Series processors are well-matched to this workload profile. High core density, strong single-thread performance and large memory bandwidth make them capable of handling the concurrency that agent threads and microservices demand. The x86 architecture means existing enterprise software stacks integrate without modification.
AMD projects that the total addressable market for server CPUs will grow significantly through the end of the decade, driven largely by agentic workload demand, reflecting how substantially agentic AI is reshaping server CPU economics.
The Instinct MI350P sits at the entry point of AMD’s accelerator stack. It supports MXFP4, MXFP6, FP8 and INT8 precision formats, carries 144GB of HBM3E memory and delivers estimated peak performance near 4,600 TFLOPS at lower precision formats, according to AMD. The dual-slot, air-cooled form factor is the key differentiator: enterprises can add AI inference capacity to standard rack-mounted servers without touching their power or cooling infrastructure. That compares favourably with higher-end accelerators like the Instinct MI300X, which delivers greater raw performance but demands more infrastructure headroom.
AMD’s ROCm open software stack underpins both product lines, according to AMD, providing developer tooling for GPU-accelerated workloads and compatibility with leading AI frameworks. The open approach is a deliberate counter to proprietary ecosystems, and gives enterprises more flexibility in their deployment choices.
Strategic Implications for Enterprise IT Leaders
The practical takeaway for infrastructure teams is straightforward: agentic AI requires a balanced architecture, not just more GPU capacity. Planning for higher CPU-to-GPU ratios, sufficient memory bandwidth and well-structured networking is now a baseline requirement, not an optimisation.
The partnership between Rackspace Technology and AMD, announced this week, illustrates where this is heading. The two companies signed a Memorandum of Understanding to build a governed enterprise AI cloud integrating Instinct GPUs and EPYC CPUs into a managed stack for regulated workloads. The goal is a secure, compliant and cost-predictable path to production agentic AI precisely what regulated industries have been waiting for.
As agentic deployments scale, the infrastructure teams that plan for CPU-GPU balance now will avoid the retrofit costs that come from GPU-first architectures hitting orchestration ceilings later. The architecture decisions made today will determine how smoothly those systems scale. For more coverage of AI chips and infrastructure, visit our AI Hardware section.
