To be revealedNews of venturing into physical chip manufacturingFollowing this, Mohamed Awad, Executive Vice President of Arm's Cloud AI Business Unit, provided further technical analysis of the processor. This new processor, named "Arm AGI CPU," is a mass-production product built on the Arm Neoverse platform, primarily designed to address the bottleneck issue of CPUs becoming crucial in data center coordination processes during the era of "agent-based AI."
To this end, Arm has completely redesigned the server reference architecture, boasting more than twice the single-rack performance of x86 architecture systems through extreme rack density and multi-threaded performance, and announcing that AI cloud infrastructure has officially entered a new era of "Arm native" development.
The rise of proxy AI: Why has the CPU become the bottleneck of computing power?
Mohamed Awad explained that in the past, the bottleneck in computing lay in "humans"—the speed at which humans input commands determined the execution rhythm of the system. However, in the era of "agent-based AI," this limitation has disappeared. Software agents autonomously coordinate tasks 24/7 and interact with multiple large language models, then make real-time decisions.
However, in this continuously operating and extremely complex environment, the role of the CPU undergoes a qualitative change. It is no longer just a supporting role to the GPU, but must simultaneously manage thousands of distributed tasks, schedule accelerators, manage memory and storage, and even handle the "fan-out" coordination tasks between massive numbers of AI agents. When the load increases exponentially, traditional x86 architecture CPUs often experience problems such as core contention and performance degradation under sustained high loads. This is the core reason why Arm decided to personally develop the AGI CPU.
Built for rack-level efficiency: a single rack can accommodate up to 45000 cores.
To address this pain point, Arm AGI CPUs are custom-designed for supporting high-density rack deployments and massively parallel computing, from their operating clock and memory to their I/O architecture.
Arm has released the hardware reference configuration for the "AGI CPU":
• 1OU dual-node design:This is Arm's standard air-cooled server reference architecture. Each Blade server contains two nodes, configured with two AGI CPUs, dedicated memory, and I/O ports, providing 272 computing cores in a single Blade server.
• 36kW air-cooled rack unit:A standard 36kW air-cooled rack system can accommodate 30 of the aforementioned Blade servers, providing a total of up to 8160 CPU cores.
• 200kW Liquid-Cooled Beast:Arm has partnered with Supermicro to design an extreme configuration that supports 200kW liquid cooling, with a single rack capable of holding 336 Arm AGI CPUs and a total core count exceeding an astonishing 45000.
Mohamed Awad emphasized that this architecture delivers more than twice the rack computing performance of the latest x86 architecture systems. The key lies in the Arm Neoverse V3 core's single-threaded performance and higher memory bandwidth, ensuring that each thread can accomplish more work without the performance crashes that can occur under full load, unlike the x86 architecture.
It's not just about selling chips, it's about defining hardware standards.
It is worth noting that Arm is not just launching chips this time, but also intends to directly define the hardware standards for the next generation of servers.
To accelerate ecosystem adoption, Arm announced the release of the "Arm AGI CPU 1OU Dual-Node Reference Server" compliant with the Open Computing Protocol (OCP) DC-MHS standard size. Arm plans to contribute the server design, supported firmware, system architecture specifications, debugging framework, and diagnostic tools to the OCP open computing community.
Currently, this chip has secured adoption commitments from industry giants including Meta, OpenAI, Cerebras, and Cloudflare, while partners ASRock, Lenovo, and Supermicro have already begun accepting orders for commercial systems.
Analysis of viewpoints
Technical analysis reveals that the Arm AGI CPU is a complete "performance monster," directly targeting the weaknesses of traditional x86 architecture data center racks.
Arm did not choose to compete with x86 in terms of absolute computing power per CPU. Instead, it leveraged the "high energy efficiency" and "high core density" of the Arm architecture to directly raise the battlefield to the "rack-scale".
When cloud service providers evaluate the construction of data centers, they focus on "how much computing power can be crammed into this 36kW rack limit?" In this respect, a single rack can provide 8000 CPU cores, or even up to 4.5 Arm architecture CPU cores without downclocking, which is a significant advantage compared to x86 architecture systems that are limited by heat and power consumption.
At a deeper level, Arm is recruiting AI startups and cloud providers (such as OpenAI and Cerebras) that lack the capacity to develop their own CPUs but urgently need high-performance coordinators. By directly providing mass-produced chips and the OCP open computing hardware design, Arm is essentially paving a highway called the "Neoverse" for the entire AI industry.
This is not only a major transformation of Arm's business model, but also the most severe challenge that the x86 camp will face in the AI server market.
