The "CPU Renaissance" is underway! Arm (ARM.US) is more confident in achieving its $2 billion outlook, as AI agents spark a huge wave of CPU demand.

According to WisdomTree Finance APP, Arm Holdings (ARM.US), owner of the ARM instruction set architecture, CEO Rene Haas said in an interview with the media on Wednesday that as AI agents increasingly penetrate various industries, driving a growing demand for CPUs, he is increasingly confident that the chip design company's exclusively launched AI datacenter CPU—AGI CPU—will be available in sufficient supply to meet customer demands reaching as high as $2 billion or even higher. The prevalence of ARM architecture CPUs in AI computing infrastructure clusters, coupled with Arm's move to directly sell its own complete chip designs, are forming two mutually reinforcing growth paths.

The $1 billion outlook announced in March 2026 already had corresponding supply arrangements; by May, customer demand had surged to over $2 billion, but with new supplies yet to be secured, the revenue outlook remained at $1 billion; in July, confidence in expanding supply and surpassing the $1 billion revenue grew; in the September interview with Jim Cramer, Arm CEO Haas further expressed that his confidence in achieving $2 billion or even stronger numbers was significantly greater than in July. The $2 billion demand outlook for Arm's AGI CPU business is planned for fulfillment during fiscal years 2027 to 2028 (spanning calendar years 2026 to 2027).

For the past two years, AI narratives have been dominated almost exclusively by GPUs, with CPUs having seemed like background “supporting actors” in the AI arms race. However, as open-source, agent-oriented AI workflows (i.e., AI agents) like OpenClaw lead the charge in inference workloads, data orchestration, task scheduling, memory access, network communication, and multi-tool integration, the market has come to a thorough realization: without a powerful CPU as the system core, GPU clusters cannot operate efficiently.

CPUs have re-emerged from “undervalued infrastructure” to the central stage of the chip world, featuring a distinct “Renaissance” flavor. Previously, AMD’s earnings showed that their Q2 datacenter revenue reached $6.7 billion, a YoY growth of 107%, jointly driven by EPYC and Instinct, while Intel's Xeon extended from enterprise computing foundations to agent execution and heterogeneous inference infrastructure; the company's Q2 datacenter and AI business revenue was about $6.3 billion, up 59% year over year, and they’ve showcased rack-level inference and agent systems based on Xeon with partners.

Amid the AI agent frenzy and surging CPU demand, Arm's share price has soared 120% year-to-date. ARM's reduced instruction set computing architecture gives server CPUs based on its designs huge advantages in high performance and low power consumption for AI inference and training tasks, compared to Intel's x86 architecture. This trait makes ARM particularly suitable for AI datacenter servers and enables efficient cooperation with AI GPUs to meet virtually unlimited AI inference and training compute needs.

Arm can be considered one of the biggest winners of the global AI boom: NVIDIA’s self-developed Grace CPU and the recently hot-selling Vera CPU are both ARM-based; Amazon's self-developed datacenter Graviton processors also use ARM architecture. Similarly, Google’s self-developed datacenter CPU Axion, based on ARM, and Microsoft's Azure Cobalt 100/Cobalt 200, custom ARM architecture datacenter CPUs based on high-performance Arm Neoverse, are all contributors, making ARM architecture a foundational part of AI cloud era's computing infrastructure, evolving from the “king of smartphones.”

Arm CEO: More Confidence in Meeting $2 Billion CPU Chip Demand

Arm CEO Rene Haas said in an interview with renowned business channel CNBC’s “Mad Money” host Jim Cramer on Wednesday that he’s increasingly confident the company can meet Wall Street’s higher revenue expectations for its new datacenter chip called AGI CPU.

“We feel pretty good about it. We really feel very good about it,” Haas said during the “Mad Money” show with Jim Cramer in San Francisco.

The comments are significant because, for Arm’s first self-designed datacenter central processor, demand has never been in question. Instead, investors are focusing on whether the company can secure enough supply and convert that demand into revenue amid the manufacturing capacity crunch caused by the AI enthusiasm and the intense competition among chip manufacturers.

This execution ability is particularly crucial, as this CPU marks a major expansion of Arm’s business model. The company has traditionally made money by licensing its chip designs to customers, while this CPU reflects its shift into selling its own complete chips.

Arm first revealed in its May earnings call that demand for the AGI CPU had reached $2 billion, double the $1 billion outlook suggested with the March launch of the first custom CPU. However, as Arm was still working to secure more supply to meet new demand, it maintained the official $1 billion revenue forecast, leading to a 10% drop in its share price.

In the July earnings call, management expressed increased confidence in securing the needed supply. The stock rallied over 7% the next trading day.

On Wednesday, Haas further told Cramer that his confidence had dramatically increased.

“So I remember it was around the May earnings call when we said we already saw $2 billion of demand; and then in the most recent earnings call, we said, from May to July, our confidence in achieving $2 billion strengthened,” Haas explained in the interview. “Now, Jim, it’s September, and I can tell you that, compared to our July earnings call, I’m even more confident about achieving it today.”

Arm’s share price has held onto the gains since the July results, but after a parabolic rise in the first half of the year, it remains about 45% below the June peak of $452. Cramer’s charitable trust previously held Arm shares but sold the position after the stock’s sizable run to lock in profits. The stock remains on the Club’s Bullpen watch list for the investment club.

AI Agents Drive CPU Demand, Arm Accelerates Data Center Chip Expansion

In September, Arm disclosed that Google is already running agent sandboxes on the Axion-based Google Kubernetes Engine, Microsoft Azure is using Cobalt 200 to accelerate tool execution within sandboxes, and NVIDIA’s Vera is also serving agent workloads. Meta is the primary partner and co-developer for Arm’s own AGI CPU product line, with ByteDance Volcano Engine bringing agent sandboxes based on this chip to market. The former path expands the application base for Arm architecture, while the latter path directly involves Arm in commercializing complete CPU products.

In the context of rapidly growing AI compute industries, the most significant positive takeaway from Haas’ interview is that confidence in supply assurance has continually strengthened from May, through July, to September, making the $2 billion of visible customer demand have a much clearer revenue realization outlook.

According to disclosure on Jim Cramer’s show “Mad Money,” Arm’s revenue outlook in March was $1 billion, customer demand reached $2 billion in May, but the official revenue forecast stayed at $1 billion as they waited for more supply to become available; the September interview revealed increasing supply confidence. Thus, the bullish narrative around ARM is shifting from “Can the new chip attract customers?” to “How can continuous CPU demand be converted into real delivery and revenue?” while Arm’s business model expands from design licensing and royalties to include sales of their own complete chips.

Global AI compute expansion is growing from billions of dollars in cloud procurement to infrastructure lock-in plans spanning more than a decade, launching CPUs into a new cycle of demand expansion. In April, Meta and CoreWeave signed an expansion agreement worth about $21 billion, extending compute supply to December 2032, primarily supporting inference workloads. Anthropic then also reached a multi-year agreement with CoreWeave to obtain more compute for the development and deployment of Claude, with related capacity coming online this year. On September 16, Blockfusion further announced its subsidiary signed a formal lease with CoreWeave for the Niagara Falls AI park in New York, with an initial 15-year term and two five-year renewal options, coupled with an expansion agreement. The park is being upgraded into high-density liquid-cooled AI infrastructure, leveraging an existing power grid, hydropower supply, and redundant fiber as long-term demand ramps up.

These orders and facility setups show that new cloud platforms are not just filling shortfalls in compute, but are also becoming a key part of long-term deployment plans for model companies and tech giants. Meanwhile, improvements in Astra for programming, browsing, computer operations, and complex task execution have expanded the actual domains in which AI can participate. Astra’s exponential boost in complex task capabilities is expected to further expand datacenter CPU demand. With compute demands booming, OpenAI announced the suspension of new registrations and upgrades for the $200/month Pro 20X plan starting September 10. From a demand perspective, multi-year datacenter orders lay the groundwork for expansion, and new-generation models like Astra broaden the scope and depth of tasks, driving ongoing inference—and growth opportunities—can now spread from GPU clusters to CPUs, datacenter memory/NAND storage, high-performance Ethernet equipment, and other components in the full AI datacenter infrastructure chain.

The core driving force behind the “CPU Renaissance” is AI’s evolution from generating answers to executing tasks: accelerators handle model computation, while CPUs are increasingly responsible for real-world work surrounding the models. In a typical heterogeneous agent system, accelerators such as GPUs handle large-model inference, while CPUs manage workflow orchestration, tool calling, code execution, database access, data preprocessing, sandboxing, and permission management; for every action the model proposes, subsequent execution often requires actual software programs and business systems, which then return results to the model for processing.

Arm’s description of agent infrastructure, as well as Intel’s Xeon, SambaNova, and NVIDIA GPU collaborative architectures, all reflect this division of labor. From an engineering logic perspective, branches, serial dependencies, memory access, and I/O operations in tasks mean systems need not only more parallel computing but also low-latency, stable general execution power. As agent numbers, concurrent tasks, and tool-call iterations increase, CPU core-hours, memory capacity, and network processing demands will also grow. Based on this, Arm predicts that after agents are deployed at scale, datacenter CPU processing capacity per gigawatt may be at least four times current levels—this is the company’s latest capacity demand forecast.

The AGI CPU offers up to 136 Neoverse V3 cores, 12 channels of DDR5 memory, and a preset 300-watt TDP, with a core focus on balancing core density, memory provision, and power constraints to support more concurrent executions. Thus, CPU demand now comes not only from hosts paired with GPUs but also from independently expanding agent execution resource pools. The more real and complex tasks AI can perform, the larger the addressable market becomes for general-purpose computing.