"CPU Is Back": Computer's Brain Returns as AI Server Commander

As artificial intelligence development shifts from training to inference, central processing units (CPUs) are drawing intense attention in the semiconductor market. Until last year, graphics processing units (GPUs) for training large language models (LLMs) commanded the spotlight, but in the era of agentic AI — models that autonomously make decisions and execute tasks — CPU performance has become the decisive battleground. After years of being overshadowed by GPUs, rendering the moniker "computer brain" almost hollow, the CPU is reclaiming its status as AI server commander, industry observers say.

When OpenAI's AI chatbot ChatGPT emerged, hyperscalers — operators of massive data centers — scrambled to secure GPUs to meet explosive AI demand. Once thought of solely as gaming graphics cards, GPUs gained attention for their strength in parallel computing. Data centers installed GPUs in bulk to run simultaneous, repetitive machine learning tasks. As OpenAI and Anthropic pushed AI model development and Google, Amazon, Oracle and Meta expanded data centers, GPU demand surged and absorbed all attention.

As its name "central processing unit" suggests, the CPU serves as the brain in computers, laptops, smartphones and data center AI servers alike. CPUs play an essential role in orchestrating numerous GPUs to work together while managing various applications. If GPUs are the players, the CPU is the coach. The CPU calls the plays, deploys starters and substitutes and coordinates the game, while GPUs execute under the CPU's direction.

In the PC era, CPUs attracted consumer attention as the first specification checked when a new computer or laptop launched. Intel dominated computer CPUs and still leads the PC CPU market today. But in the AI market, the dynamic was different. Because training volume and accuracy of AI models mattered most, attention concentrated on GPUs over CPUs. The goal-scoring GPU captured the crowd's spotlight. That is the backdrop to Nvidia's rise to the top of global market capitalization.

Starting this year, however, the mood shifted again as AI agents — which put trained AI to practical use — became the central topic. In the agentic AI era, inference for rapid decision-making and execution has grown critical. To process step-by-step reasoning quickly, CPU performance must improve dramatically. The CPU must retrieve data stored in memory on the fly to relay to the GPU, which then processes the delivered data. Semiconductor research firm SemiAnalysis called it "the return of the CPU."

With projections that the CPU market will outpace GPU growth by 2028, chipmakers are racing to develop CPUs for data center AI servers. Nvidia CEO Jensen Huang unveiled the "Vera CPU" — purpose-built for agentic AI and targeting inference demand — at the GTC conference on March 16. Nvidia emphasized that the chip delivers 1.5 times the performance and twice the energy efficiency of Intel's x86 CPUs, and that a rack of 256 Vera CPUs can be deployed in data center servers.

AMD, which shares the top two positions with Intel in the server CPU market, is also betting on its next-generation CPU. AMD is preparing "Venice," the successor to its EPYC CPU series, targeting data center server replacement demand. AMD emphasized that EPYC CPU-based systems are estimated to deliver up to 2.1 times higher per-core performance and up to 2.26 times better performance per watt compared to Nvidia's Grace-based systems.

Arm, the world's largest semiconductor intellectual property (IP) company, also stepped into the CPU market with its first in-house chip, the "Arm AGI CPU." Bracing for explosive demand, Arm broke its longstanding principle of not competing with its customers, moving beyond its design IP licensing business model to challenge the market as a fabless chipmaker. The company's first proprietary chip in its 35-year history focuses on improving energy efficiency in AI data centers.

A British company founded in November 1990, Arm had maintained a revenue model of providing IP blueprints to semiconductor firms and collecting licensing fees. But as the semiconductor market pivoted toward AI and demand for inference chips surged recently, the 35-year-old principle was abandoned. As competition in inference AI models intensified, data centers are expected to require more than four times the current CPU capacity per gigawatt. HSBC called the in-house chip launch "a game-changing development."

Arm's in-house chip development is expected to send shockwaves through the AI semiconductor market. Intel and AMD CPUs, built on Intel's x86 architecture, have long dominated the data center market, but a formidable competitor has now emerged. Arm could pressure semiconductor manufacturers that are its IP clients to adopt its CPUs. Nvidia's Vera CPU, unveiled last week, was also developed on an Arm-based architecture. In response, AMD emphasized that x86 CPU architecture offers a broad, proven software ecosystem and carries less of the refactoring burden — restructuring to address vulnerabilities — that commonly arises when adopting Arm-based systems.