Table of contents
Introduction to Arm
Business segments
Business model
Increasing value per chip through Armv9 and CSS
Emerging production silicon opportunity: Arm AGI CPU
Increasing CPU content in AI workloads
Competitive strengths
Management team
Financials
Valuation
Key risks
Conclusion
Increasing CPU content in AI workloads
Thus far, the AI story in the early days has been all about GPUs and Nvidia.
Today, the balance of power is shifting towards CPUs.
The focus for AI has shifting from training models to deploying these models at scale, but perhaps more importantly, moving from simple prompt-and-response applications into autonomous software agents capable of planning, reasoning, using tools, interacting with enterprise software, and completing complex multi-step tasks.
Agentic AI fundamentally changes the compute profile of AI infrastructure.
The key insight is that training is accelerator-centric, whereas agentic AI is system-centric.
As AI systems become increasingly autonomous, the amount of CPU compute required alongside GPUs grows substantially.
Rather than replacing GPUs, CPUs become increasingly valuable because they orchestrate, coordinate, and enable the increasingly complex workflows surrounding AI inference.
Traditional AI inference is relatively straightforward.
A user submits a prompt, the model performs inference on a GPU, and the generated response is returned. The GPU performs nearly all of the computationally intensive work, while the CPU largely serves as a coordinator responsible for feeding data to the accelerator and handling basic system functions.
Agentic AI fundamentally changes this architecture.
Instead of simply generating an answer, an AI agent must first understand the objective, formulate a plan, retrieve relevant information from databases or vector stores, determine which external tools should be invoked, execute those tools, validate intermediate results, potentially revise its own reasoning, and finally generate a response.
A single user request can trigger multiple inference passes, several database lookups, API calls, code execution, browser interactions, and security checks before the task is complete.
The workload therefore shifts from executing one large parallel computation to coordinating dozens of interconnected computational tasks.
GPUs remain responsible for executing the neural network itself, but the surrounding orchestration increasingly falls on the CPU.
The reason CPUs become increasingly important is that they are fundamentally better suited than GPUs for general-purpose computing.
GPUs excel at performing identical mathematical operations simultaneously across thousands of cores. CPUs excel at handling sequential logic, decision-making, scheduling, branching instructions, memory management, and operating system functions.
Agentic AI happens to require far more of the latter.
Every time an AI agent decides whether to search the web, retrieve internal documents, execute Python code, call an enterprise application, or verify an intermediate result, the workload becomes less about tensor computation and more about coordinating heterogeneous computing resources. These orchestration tasks are inherently CPU-intensive.
One useful way to think about modern AI infrastructure is to separate it into a compute plane and a control plane.
The compute plane consists primarily of GPUs and AI accelerators. These devices execute transformer models, perform matrix multiplication, calculate attention mechanisms, and generate tokens.
The control plane is largely the responsibility of CPUs.
CPUs schedule work across GPU clusters, move data between memory and storage, coordinate networking traffic, enforce security policies, manage virtual machines and containers, perform retrieval operations, execute external applications, and maintain the state of AI agents.
In essence, the CPU functions as the operating system of the AI data center.
This distinction becomes increasingly important because future AI systems will likely consist of millions of simultaneously executing autonomous agents rather than millions of isolated chatbot sessions.
Coordinating these agents requires sophisticated scheduling, memory management, caching, security enforcement, and communication between services, areas where CPUs excel.
Industry participants have increasingly argued that inference clusters serving autonomous agents require substantially greater CPU capacity than traditional training clusters.