The competitive landscape of the data centre CPU industry, long led by Intel and AMD, is shifting rapidly.

Big cloud companies such as Amazon Web Services, Microsoft and Google Cloud have moved beyond deploying their own Arm architecture-based data centre GPUs on their cloud platforms. More recently, Nvidia and Arm have also begun deploying data centre CPUs they made themselves. Alibaba unveiled the XuanTie C950, a RISC-V-based data centre CPU built on the open-source chip architecture rather than Arm.

The industry moves are seen as stemming from a growing share of AI inference in data centres, where AI responds to user queries. As demand rises beyond model training to inference, the weight carried by CPUs in addition to GPUs is also increasing, the industry says.

Nvidia showed a next-generation CPU beta at its annual GTC2026 event held recently. It is aimed at AI agents.

According to the company, Vera is designed so that CPUs play a leading role in an AI factory rather than merely providing support. It is based on the Arm architecture and is 50 percent faster and twice as efficient as existing x86-based CPUs.

Nvidia has dominated the AI market with GPUs, but it is now expanding into CPUs as agentic AI spreads and the role of CPUs rises rapidly.

While GPUs are optimised for training and running AI models, CPUs handle large-scale data processing and coordination of AI agents. Nvidia will first deploy Vera CPUs at Meta on a large scale. It is drawing attention because it is a case of providing CPUs on a large scale separately from GPUs.

Arm, which has focused on providing intellectual property to chip developers, has also jumped into the race by releasing its own data centre CPU, the AGI CPU.

Arm's move also squarely targets the spread of AI. According to the company, it supports installing up to 64 CPUs, or about 8,700 cores, in a single air-cooled rack. Arm said performance per watt is double that of an x86 rack. It said that means double the performance in the same space and with the same power. The company said it expects the terms to be attractive to data centre operators facing tight power constraints. The XuanTie C950 unveiled by Alibaba is also focused on supporting AI agents. It is optimised for inference work that runs AI models when installed in data centres.

While GPUs are essential for AI model training, CPUs play an important role when agents perform multi-stage tasks, Alibaba says.

XuanTie C950 was developed based on the RISC-V architecture rather than Arm. RISC-V is a CPU design blueprint that can be used for free and is drawing attention as an alternative that can cut Arm licensing costs. Alibaba is strengthening AI chip development through its T-Head chip unit, and recently also released the Zhenyu 810E AI chip.