Arm Unveils 136-Core AGI CPU, Its First Production Chip for AI Agent Data Centers

Arm Holdings has unveiled what it calls the AGI CPU — a 136-core data center processor built specifically for AI agent workloads. The move marks a historic shift for the British chip designer, which for decades has licensed its processor designs to other companies rather than selling silicon directly.

Key Specifications

The AGI CPU packs up to 136 Neoverse V3 cores running at 3.2 GHz all-core and 3.7 GHz boost, spread across two dies within a 300-watt TDP. The chip features:

12 DDR5 memory channels delivering 6 GB/s bandwidth per core
Dual 128-bit vector units — deliberately minimal compared to 512-bit SIMD on x86 servers
No simultaneous multithreading (SMT) — a design choice Arm says eliminates wasted die area

Meta is the lead partner for the chip, signaling strong hyperscaler interest in purpose-built agentic AI infrastructure.

A Deliberate Design Philosophy

Arm executive Mohamed Awad made the case that legacy x86 features waste power and silicon area in agentic workloads. "When you increase the frequency, what else do you increase? Power. That's a problem," Awad said.

The company argues that AI agent orchestration — managing thousands of concurrent agent tasks — demands high core counts and memory bandwidth rather than the wide vector engines used for traditional number crunching. "We're focused on exactly and only what the agentic datacenter needs: performance, scale, and efficiency," Awad added.

Intel Pushes Back

Not everyone agrees that AI agents need a new class of CPU. Kevork Kechichian, Intel's Data Center Group leader and a former Arm VP, questioned whether there is real market demand for specialized agentic processors.

On Arm's decision to drop SMT, Kechichian was blunt: "My view is that, if they had the option, they would have put it in." He pointed to Nvidia's competing Vera CPUs, which feature spatial multithreading.

Intel's own Clearwater Forest processor shares surprising similarities with Arm's AGI — 288 stripped-down cores, minimal SIMD, no SMT, and 12 DDR5 channels — but Kechichian positions it primarily for networking workloads like packet processing, not agentic AI.

The Nvidia Factor

Arm's announcement comes alongside Nvidia's Vera CPU, which the GPU giant bills as the world's first processor purpose-built for the age of agentic AI and reinforcement learning. Nvidia claims Vera delivers twice the efficiency and 50 percent faster performance than traditional rack-scale CPUs, using 88 custom Arm-based Olympus cores.

The irony is not lost on the industry: Nvidia's Vera chip uses Arm's own architecture, putting the two companies in direct competition despite their long-running partnership.

From IP Licensor to Chipmaker

This is the most significant strategic shift in Arm's 35-year history. The company has always designed processor architectures and licensed them to manufacturers like Apple, Qualcomm, and Samsung. Now it is competing directly with its own licensees — and with x86 incumbents Intel and AMD.

Arm is projecting $15 billion in annual revenue from the AGI CPU line through 2031, a figure that would nearly double the company's current revenue and reshape its position in the semiconductor industry.

What This Means

The race to build dedicated AI agent infrastructure is intensifying. As agentic AI workloads grow — from coding assistants and autonomous research tools to enterprise automation — the question of what hardware best serves these workloads is becoming a billion-dollar debate.

Whether AI agents truly need purpose-built CPUs or can run efficiently on existing hardware remains an open question. But with Arm, Nvidia, and Intel all positioning for this market, the industry is clearly betting that the agentic era will reshape data center architecture.

Source: Tom's Hardware