At Computex 2026, Arm officially announced the commencement of mass production for its AGI CPU, betting on the explosive demand for CPU computing power in data centers during the Agentic AI era and marking a strategic shift into self-developed chips.
CEO Rene Haas stated that the previously set goal of $15 billion in revenue from self-developed chips is "expected to be achieved ahead of schedule," citing stronger-than-anticipated demand. The target, originally set for the end of the decade, was projected to see self-developed chip revenue surpass the existing IP licensing business.
On Monday, Arm's stock price surged over 15%, closing at a record high of $408.85, bringing its year-to-date gain to 274% and a 224% increase over the past 12 months. NVIDIA launched the RTX Spark notebook chip, which integrates a custom Arm Grace CPU with 20 Arm-based cores and a Blackwell GPU designed for agent computing. Major PC manufacturers including Acer, ASUS, Dell, Gigabyte, HP, Lenovo, MSI, and Microsoft have confirmed plans to release corresponding models.
Rene Haas directly addressed a long-standing question, stating that the moment for the Windows+Arm ecosystem to achieve scaled competitiveness in notebooks and general computing has arrived. NVIDIA's Jensen Huang emphasized that building agent devices requires high-performance CPUs, which is a core reason for choosing Arm architecture, noting the product's 20-core CPU with excellent single-core performance.
Concurrently, Arm's AGI CPU officially entered mass production, with leading AI and cloud enterprises such as Meta, OpenAI, and Oracle Cloud Infrastructure joining the product ecosystem.
AGI CPU Mass Production: Targeting the Data Center CPU Market with Efficiency Advantages
The Arm AGI CPU is positioned to address the incremental CPU demand in data centers driven by the rapid growth of Agentic AI. It is currently offered in air-cooled and liquid-cooled system configurations. The air-cooled rack consumes approximately 36 kilowatts and houses 8,000 cores, while the liquid-cooled rack consumes about 200 kilowatts and contains over 45,000 cores. Rene Haas stated that under the same rack power consumption, the Arm AGI CPU delivers roughly double the performance of competing architecture systems. This means data centers can achieve higher computing power under the same electrical constraints or maintain competitive performance while reducing power consumption.
In his keynote, Haas explained the underlying logic of how Agentic AI increases CPU demand: GPUs and XPUs excel at training and inference, acting as token generators. However, AI Agents run continuously and spawn new Agents, generating a massive backend workload of token allocation, management, coordination, and delivery—system-level tasks that must still be handled by CPUs. He noted that Arm previously estimated data centers might need four times the number of CPU cores within the same power envelope, and current market discussions now include estimates ranging from 4x to 8x or even 10x.
Regarding market size, Arm had estimated the CPU market could reach over $100 to $120 billion in the next five years. Haas indicated that with the continued growth of AI Agent applications, the figures being discussed in the market now exceed those original predictions. Using a data center with approximately 10GW of capacity as an example, he calculated that adopting a CPU solution with higher performance per watt could yield cost savings on the order of up to $10 billion.
RTX Spark Reshapes the AI PC Landscape, Arm Ecosystem Barriers Highlighted
NVIDIA CEO Jensen Huang joined Rene Haas on stage to discuss the positioning of the RTX Spark. Huang stated that NVIDIA's goal is to redefine the personal computer. With the PC having existed for forty years, legacy system software developed by hand will gradually be replaced by various agent applications, a vision driving NVIDIA to re-architect hardware and operating systems. He reiterated that building agent devices requires high-performance CPUs, a core reason for choosing Arm architecture, and praised Arm's customization capabilities for flexibly adapting to overall system design needs.
Product-wise, the RTX Spark comes standard with 128GB of memory. Using NVIDIA's proprietary NVFP4 numerical format compression, the device can locally host a large language model with hundreds of billions of parameters for daily computing, while complex inference tasks can leverage cloud computing power. Huang also revealed that Adobe has announced plans to re-architect Photoshop and Premiere with CUDA acceleration and agent-call adaptation, with industry software from Blender, Autodesk, Dassault, and Siemens also set to complete RTX hardware acceleration adaptation.
Rene Haas pointed out in his speech that Linux, the full range of Apple macOS, Chrome OS, and Windows have all been adapted for the Arm platform. This full-system compatibility, achievable only by Arm, forms an important foundation for the Windows+Arm ecosystem to achieve scaled competitiveness in notebooks and general computing. He credited this milestone to decades of continuous collaboration with Apple, Google, and Microsoft.
Expanding Customer Base, Positive Response to Conflict of Interest Concerns
The Arm AGI CPU has gained support from companies including Meta, Rebellions, SAP, Cerebras, and OpenAI, with Oracle recently announcing it has joined the ecosystem.
Rene Haas directly addressed external concerns about whether Arm's self-developed CPU creates a conflict of interest with existing IC design clients. He stated that the project was driven by customer demand, not Arm's initiative to enter the market competitively. He disclosed that Meta approached Arm with the requirement two years ago, at a time when there were almost no server chip options based on Arm architecture available.
Haas emphasized that current demand for AI computing chips far exceeds supply, with orders for both the Arm AGI CPU and NVIDIA products being extremely strong. He stated that at this stage, there is no evidence of Arm's own products crowding out existing clients, citing positive testimonials and video endorsements from multiple partners at the March 24th launch event as proof of prior thorough communication with key partners.
Market Outlook: Analysts See Strong Demand for Both x86 and Arm Platforms
Mizuho analyst Vijay Rakesh maintained "Outperform" ratings on both Arm and AMD, and a "Neutral" rating on Intel in a research note, stating that "within the overall strong outlook, we continue to see both x86 and Arm platforms benefiting from strong CPU demand growth in the second half of 2026 through 2027."
Independent analyst Richard Windsor wrote in his Radio Free Mobile blog that if AMD were to announce an Arm-based CPU, it would signal abandoning the x86 architecture, at which point "reality would become impossible to ignore."
On the supply chain front, Rene Haas told media during Computex that the supply-demand balance for various memory chips, including high-bandwidth memory, DRAM, and NAND, is tight. Due to insufficient capacity expansion during the previous downturn, memory chips remain the most challenging bottleneck to resolve, and overall supply is expected to remain constrained for some time.
Comments