Bank of America: Why is the rise of agent AI making CPUs popular again?

```

Over the past three years, AI investment has revolved around a single keyword: GPU. From the explosion of generative AI sparked by ChatGPT to the skyrocketing market value of Nvidia, the market has formed an almost unanimous consensus—the era of future AI computing power belongs to GPUs, and CPUs are gradually being marginalized.

However, a recent semiconductor industry report from Bank of America offers a completely different perspective: With the rise of Agentic AI, not only will CPUs not be replaced, but they may also usher in the largest growth cycle in over a decade.

Bank of America has significantly raised its projection for the global server CPU market in 2030 from $125 billion to $170 billion—an almost fivefold increase from $35.2 billion in 2025—with a compound annual growth rate of 37% over the next five years. In its view, Agentic AI is changing the computational architecture of AI systems, and CPUs are poised to become one of the biggest beneficiaries of this transformation.

More importantly, this is not a story of "CPU replacing GPU," but rather, amid the continuous expansion of total AI infrastructure demand, a huge incremental market previously overlooked by the market is now emerging.

From "Answering Questions" to "Completing Tasks": AI Is Undergoing a Qualitative Change

In recent years, large models have essentially remained "question-answering machines." Users input questions, models generate answers, and the entire process mainly relies on GPUs for massive parallel computation. But the logic of Agentic AI is fundamentally different.

In the future, AI will not just answer questions but will also be able to autonomously complete tasks. It will need to understand objectives, make plans, call upon tools, retrieve information, execute code, manage memory, and coordinate among multiple tasks. For example, an AI assistant might automatically check flight availability, compare hotel prices, arrange itineraries, book restaurants, draft meeting notes, and send emails—a series of operations.

The entire process involves extensive decision-making, scheduling, and system management. These tasks are not suitable for the large-scale parallel computation at which GPUs excel, but rather depend more on low latency, sequential execution, and complex I/O processing capabilities—areas where CPUs are strongest.

Bank of America believes that as AI shifts from "content generation" to "task execution," an increasing proportion of computational workload will flow back to the CPU.

CPUs Are Moving from Supporting Roles to the Scheduling Center of AI Systems

During the era of large models, CPUs mainly played a "logistical support" role, responsible for data preprocessing, task scheduling, storage access, and delivering data to GPUs, while the real value creation happened on GPUs. However, in the era of Agentic AI, CPU importance is rising rapidly.

Bank of America points out that CPUs in future AI data centers will primarily handle three types of work: First, traditional cloud computing and enterprise server services; second, serving as Head Nodes in AI clusters, responsible for managing and coordinating thousands of GPUs; third, wholly new Agentic AI nodes, dedicated to agent workloads. The third category represents an almost entirely new market.

These Agentic CPU nodes manage the inference loop, state memory, tool invocation, and task orchestration for AI agents, becoming the "central brain" of future AI systems. Unlike their previous behind-the-scenes role merely supplying data to GPUs, these CPUs are stepping into the spotlight.

Bank of America projects that by 2030, the CPU market dedicated solely to Agentic AI will reach about $70 billion, roughly matching the AI cluster CPU market scale. In other words, future data centers will not only house GPU racks but also a large number of CPU racks specialized for Agentic AI.

The More Complex AI Systems Get, the Greater the Demand for CPUs

The main debate in the market today is whether increased CPU demand will cannibalize GPU demand. Bank of America’s answer is no. The report argues that AI infrastructure is shifting from being driven by a single component to being driven by system-level architecture. Competition in AI used to be about GPU count; in the future, it will be about the whole system: GPUs for training and inference, HBM for storage, networks for communication, and CPUs for coordinating everything.

With the upgrade of model capabilities, the increase in the number of agents, and the rising complexity of workflows, task scheduling, data exchange, and resource management within the system will grow exponentially. Therefore, CPU demand will not replace GPU demand but will expand in tandem with GPU growth.

Bank of America expects that by 2030, the ratio of CPUs to GPUs in AI data centers may reach 1:1 or higher, compared to just 1:2 or even 1:4 over the past few years. This means the AI industry is shifting from "GPU single-engine driven" to "dual-engine CPU + GPU driven."

Under this trend, CPU manufacturers—especially those deeply tied to leading cloud and AI companies and previously overlooked by the market—may have the opportunity to see a revaluation.

Risk Warning and DisclaimerThe market has risks, and investment should be cautious. This article does not constitute personal investment advice, nor does it take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investing based on this article is at your own risk. ```