The wave of AI agents is driving a "super cycle"! Citi: The server CPU market will expand to over $130 billion by 2030

According to Zhitong Finance, Citibank expects the server CPU industry to experience significant expansion in the coming years, projecting the market size to grow from $29.3 billion in 2025 to approximately $132 billion in 2030. The institution indicated that the main driver for this growth will be the emerging category of “Agentic CPU.” This type of chip is primarily used to support increasingly autonomous artificial intelligence workloads.

Citibank anticipates that the general-purpose server CPU market will see a compound annual growth rate (CAGR) of 20%, reaching about $50.9 billion by 2030. At the same time, AI head node processors are expected to have a 21% CAGR over the same period, reaching a market size of approximately $21.1 billion. Agentic CPUs, on the other hand, are expected to become the fastest-growing sector—Citibank forecasts this segment to expand at a CAGR as high as 185%, with the market size reaching about $59.4 billion by 2030.

With the major launch of Claude Cowork by Anthropic and the surge of super AI agent tools like OpenClaw capable of autonomous task execution expected in 2026, the wave of AI agents is rapidly sweeping the globe. The bottleneck in AI computing power architecture is rapidly shifting from GPU-centric matrix multiplication throughput to data center CPUs focusing on control flow, task orchestration, and memory/IO coordination. High-performance CPUs designed for large-scale AI data centers are now in a severe supply shortage.

Over the past two years, the narrative of the AI industry chain has been dominated almost entirely by GPUs, with CPUs making little impression in AI servers. The reason for this is that, during the training era, the primary bottleneck was parallel computing power, with GPUs handling the heaviest matrix computations while CPUs mostly undertook general control and basic scheduling duties.

However, with the explosive growth of AI agents and reinforcement learning (RL) workloads, the strategic position of CPUs in data centers is being structurally re-evaluated. The essence of an agent is not just producing longer answers but breaking down a single request into an entire workflow. Models are no longer just generating an answer—they are executing a sequence of processes. Once AI shifts from “one-time computation” to “workflow execution,” the system’s reliance on CPUs increases significantly. This is because many critical workloads are not suitable for GPU execution, such as task orchestration, thread scheduling, process management, sandboxing, pre/post-processing, cache coordination, and state maintenance—all typical CPU tasks. Especially in multi-agent collaboration scenarios, where multiple agents run in parallel, invoke tools, and share states, there are increased demands on CPU core count, thread count, single-core performance, and memory management capability.

Dylan Patel, chief analyst at renowned semiconductor research firm SemiAnalysis, said directly in a deep-dive interview last month that AI workloads are evolving from simple text generation to complex agents and reinforcement learning, and that CPUs are facing “extremely severe capacity shortages.”

According to market research firm TrendForce, the current CPU-to-GPU ratio in AI data centers is about 1:4 to 1:8, but in the era of agent-based AI, this ratio is expected to evolve to 1:1 to 1:2. In terms of market size, Creative Strategies predicts that the data center CPU market will grow from $25 billion in 2026 to $60 billion by 2030; with the addition of agent-related demand, it could approach $100 billion.

This structural shift has already sparked a chain reaction on both the supply and demand sides. Intel and AMD have already raised prices for some CPU product lines by the end of Q1 2026. Meanwhile, Nvidia and Arm both announced in March 2026 that they are entering the server CPU market—both a GPU giant and an IP licensor making the same move in the same month, which is no coincidence but a concentrated signal from the market.

Intel's Xeon processors have long dominated more than 95% share of the data center CPU market. However, this dominance began to weaken in 2021—yield issues with the Intel 7 process led to nearly a two-year delay in the release of Xeon Sapphire Rapids, opening up market opportunities for AMD’s EPYC Milan.

Intel plans to launch two flagship products in 2026. One is the Xeon 6+ (Clearwater Forest) with the Darkmont architecture, featuring 288 cores/288 threads and a TDP of about 450W; the other is the Xeon 7 (Diamond Rapids) with the Panther Cove-X architecture, up to 256 cores/256 threads and a TDP of up to 650W.

Both products are based on Intel’s most advanced 18A process and, for the first time, introduce Foveros Direct hybrid bonding technology. However, TrendForce points out that due to ongoing yield issues with the 18A process, mass production for these products may be delayed until 2027.

In contrast, competitor AMD is moving more steadily. Its flagship product in 2026, EPYC Venice, will adopt TSMC’s N2 process, the Zen 6 architecture, and be equipped with advanced CoWoS-L and SoIC packaging. Via simultaneous multi-threading (SMT) technology, it will achieve 256 cores/512 threads—the highest thread count in the current market. TrendForce expects AMD to continue eating away at Intel’s market share in 2026.

In addition to these two giants, a group of non-traditional players is rushing into the server CPU race at unprecedented speed, seeking to fundamentally rewrite the competitive landscape. In March, Nvidia announced that the Vera CPU would be sold as a standalone product to meet client demand for more flexible CPU:GPU configurations. Vera uses Nvidia’s self-developed Olympus architecture, is based on TSMC’s N3 process and CoWoS-R packaging, provides 88 cores/176 threads, and is equipped with 1.8 TB/s NVLink-C2C interconnect for memory sharing with Nvidia GPUs. Nvidia also launched the Vera CPU rack, with a single rack integrating 256 CPUs, totaling 22,528 cores/45,056 threads and 400 TB of total memory.

Also in March, Arm announced its first self-developed CPU product, the Arm AGI CPU, ending its 35-year history as a pure licensor. Based on TSMC’s N3 process and Neoverse V3 architecture, it offers 136 cores/136 threads, a TDP of 300W, and supports DDR5-8800 memory and PCIe Gen6. Arm also announced two rack configurations: the air-cooled version integrates 60 AGI CPUs (8,160 cores, about 180 TB memory), while the liquid-cooled version supports 336 CPUs (45,696 cores, 1 PB memory).

Major cloud service providers (CSPs) are also accelerating their layout of self-developed CPUs. Amazon (AMZN.US) AWS will launch Graviton5 (192 cores/192 threads, TSMC N3 process) in December 2025, deploying it alongside their proprietary Trainium 3 AI ASIC to reduce AI computing costs. Microsoft (MSFT.US) will release Cobalt 200 (N3 process, 132 cores/132 threads) in November 2025. Google plans to launch the Axion C4A.metal bare metal version and next-generation Axion N4A in 2026, focusing on the highest cost-performance ratio.

Citibank says that as hyperscale cloud service providers gradually shift from large-scale AI training expenditures to business deployment of AI systems, Intel (INTC.US), AMD (AMD.US), and Arm (ARM.US), among others, are increasingly active in seeking CPU-related opportunities.

In Citibank’s view, by the end of this decade, Intel will still maintain its global CPU market leadership, with an estimated share of 47%; AMD is expected to occupy 34%, and Arm plus other competitors will collectively represent the remaining 19%. Citibank has raised Intel’s target price from $95 to $130 and maintains a “Buy” rating. At the same time, the firm increased AMD’s target price from $358 to $460 but retains a “Neutral” rating.