Arm and Alibaba Launch Specialized CPUs for AI Agent Workloads

Arm Breaks 35-Year Tradition with First Proprietary Chip

Arm Holdings has made a historic pivot by releasing its first-ever proprietary processor — the Arm AGI CPU — designed for data centers and optimized for AI inference. Meta has signed on as the first customer for the new chips.

This launch represents a fundamental shift in Arm's business model. For 35 years, the company operated without its own manufacturing capabilities, earning revenue exclusively through licensing and royalties. Now Arm is expanding its Neoverse platform beyond IP and Arm Compute Subsystems, giving customers a broader range of deployment options: building custom chips, integrating platform solutions, or using processors designed by Arm itself.

Arm CEO Rene Haas projects that by 2031, this new line of business will generate approximately $15 billion in annual revenue. The company's total revenue could reach $25 billion, with earnings per share hitting $9.

Why This Matters

The rise of agentic AI is reshaping computational infrastructure requirements. Until now, the industry has been predominantly focused on GPUs — graphics accelerators dominated by Nvidia — which are essential for training large-scale models through massive parallel computation.

CPUs, however, operate on a fundamentally different principle: sequential execution. This makes them the optimal choice for AI agents that carry out specific chains of actions. According to Haas, demand for such processors is expected to quadruple.

Arm explained that in modern data centers, CPUs manage thousands of tasks: coordinating accelerators, handling memory and storage, distributing workloads, and moving data. With the emergence of agentic AI, the CPU also takes on the role of coordinating multiple AI assistants simultaneously.

The Arm AGI CPU delivers high per-task performance while running thousands of cores in parallel, staying within power consumption and cooling constraints.

Alibaba's Answer: The XuanTie C950

Chinese tech giant Alibaba has unveiled its own competing solution — the XuanTie C950 processor, also targeting agentic AI workloads. The chip can handle the multi-step tasks performed by AI assistants and is designed for data center inference deployment.

The XuanTie C950's distinguishing feature is its use of the open RISC-V standard as the underlying architecture — a direct alternative to Arm's proprietary designs. The processor can be customized for specific inference scenarios, enabling clients to tailor it to their particular use cases.

The AI Chip Race Intensifies

Arm is best known as the architect behind the processor designs powering most modern smartphones. In 2018, the company entered the server chip market to compete with x86 processors from Intel and AMD by launching the Neoverse platform. The release of its own CPU now elevates competition in AI server processors to a new level.

The AI chip race continues to attract major players. In December 2025, Amazon Web Services introduced Trainium3, the latest iteration of its in-house AI chip. The emergence of specialized CPUs from both Arm and Alibaba signals the formation of a new market segment focused specifically on inference and AI agent operations, rather than solely on model training.