AGI CPU from Arm: server chip for AI agents

Arm Launches Its First Server AGI CPU for AI Agents

Arm Holdings has transitioned from pure architecture licensing to manufacturing complete chips. At the Arm Everywhere conference in San Francisco, CEO Renée Haas introduced the AGI CPU—the company’s first proprietary server processor. Designed specifically for AI agents, the chip delivers powerful CPU logic essential for integrating with external tools, data retrieval, and script execution.

The AGI CPU is built on 136 Neoverse V3 cores, fabricated using TSMC’s 3nm process. With a TDP of 300W, it delivers twice the rack-level performance compared to Intel and AMD x86 solutions. Development involved 1,000 engineers across three labs, including a $71 million facility in Austin. Mass production begins in the second half of 2026.

The market responded with a 20% surge in Arm’s stock price, pushing its market cap past $155 billion. The first customer is Meta, deploying the chip to power Llama infrastructure and autonomous agents in WhatsApp and Instagram.

AI Agent Demands on CPU Architecture

The shift from generative AI models to AI agents is reshaping data center dynamics. While GPUs dominate inference and neural network training, agents require robust CPU capabilities for orchestration—running scripts, managing APIs, and processing real-time data.

Traditional clusters with an 8:1 GPU-to-CPU ratio are outdated. The AGI CPU restores balance by offering high core density on silicon, enabling more agents per rack without increasing power consumption.

Key specs of the AGI CPU:

• Cores: 136 × Neoverse V3
• Process: 3nm (TSMC)
• TDP: 300W
• Performance: 2× higher than x86 at the rack level
• Use Case: AI agent tooling (search, scripting, API integration)

While more power-hungry than mobile Arm chips, this efficiency is justified by demanding server workloads.

Customers and Market Demand

Meta is integrating the AGI CPU to scale the Llama ecosystem. Agents will handle user tasks like product searches, bookings, and code generation.

Waiting list includes OpenAI, Cloudflare, SAP, and SK Telecom—companies focused on maximizing deployment density within constrained data center footprints.

Arm continues its licensing business but now adds direct chip sales. Projections show $25 billion in revenue by 2031, with $15 billion from hardware—six times current levels.

Shifting Chip Ecosystem Dynamics

The launch of the AGI CPU blurs traditional roles: architects (Arm), foundries (TSMC), and vendors (Nvidia, Intel). Competition intensifies as Nvidia now faces direct competition in the server CPU space.

The industry is moving toward integrated AI infrastructure stacks. Developers like Google and Meta are already designing custom accelerators, and Arm is following suit.

Key Takeaways:

• AGI CPU is Arm’s first ready-to-use server chip with 136 Neoverse V3 cores and 300W TDP.
• Delivers 2× rack-level performance for AI agents vs. x86.
• First customer: Meta; others in queue include OpenAI; mass production starts H2 2026.
• Chip revenue: $15B out of $25B projected by 2031.
• Direct competition with Nvidia and x86 in the age of agent-driven AI.

— Editorial Team