Arm Holdings has unveiled its own data center chip optimized for AI inference.
The first client for these new CPUs is Meta. According to Arm's CEO Rene Haas, this development is expected to generate around $15 billion in annual revenue by 2031. The company's overall revenue could reach $25 billion, with earnings per share projected at $9.
CPUs are regaining high demand due to the rise of agent-based AI, which is altering computational requirements.
"In modern data centers, CPUs manage thousands of tasks: coordinating accelerators, managing memory and storage, distributing workloads, and moving data. With the advent of agent-based AI, they also coordinate the work of numerous assistants," the company stated.
The release of the Arm AGI CPU marks a significant shift in the company's business model. Previously, it lacked its own manufacturing capabilities and had earned revenue solely through licensing and royalties for 35 years.
"The Arm AGI CPU delivers high performance per task while operating thousands of cores in parallel, all within energy consumption and cooling constraints," the company emphasized in its blog.
Arm is well-known for its architecture, which underpins most modern smartphones. In 2018, it began competing with Intel and AMD's x86 server chips by launching the Neoverse platform.
"For the first time in over 35 years, Arm is releasing its own chips—expanding Neoverse beyond IP and Arm Compute Subsystems to give clients more options for deploying computing: from creating custom chips to integrating platform solutions or using processors developed by the company itself," the announcement stated.
Alibaba's Counterpart
Chinese tech giant Alibaba has introduced its own solution—the XuanTie C950 CPU for agent capabilities.
This chip can handle multi-step tasks performed by AI assistants and will be used in data centers for inference.
Until recently, the industry primarily focused on graphics accelerators (GPUs), dominated by Nvidia. These chips are essential for training large models, as they can perform numerous calculations in parallel.
In contrast, CPU architecture is designed for sequential operation, making them an optimal choice for the evolving ecosystem of AI agents focused on executing specific action chains.
The XuanTie C950 "can be customized for specific inference scenarios, allowing clients to tailor processors to their tasks." The processor is based on the open RISC-V standard, which serves as an alternative to Arm's architecture.
It is worth noting that in December 2025, Amazon Web Services introduced a new version of its AI chip, the Trainium3.