Alibaba unveils XuanTie C950 RISC-V chip designed for AI agents and cloud workloads

Alibaba Introduces XuanTie C950 as Most Powerful RISC-V Server Chip

Alibaba has unveiled the XuanTie C950 AI chip, which the company claims represents the highest performing RISC-V processor ever developed. Announced Tuesday at a conference hosted by DAMO Academy, Alibaba's research arm, the new chip is built using open-source RISC-V architecture and manufactured with a 5nm process1

1

. The 3.2 GHz server chip is designed to power cloud infrastructure, generative AI workloads, high-end robotics, and edge computing devices3

3

. According to Alibaba, the XuanTie C950 performs more than three times faster than its predecessor, the XuanTie C920, marking a substantial leap in the company's in-house chip development efforts3

3

.

Optimized for AI Workloads and Agentic AI Applications

The XuanTie C950 comes equipped with a self-developed Tensor Processing Engine that natively supports large language models with hundreds of billions of parameters, including Qwen3 and DeepSeek V31

1

. This positions the chip as what Alibaba calls "a new type of high-end CPU for the AI Agent era." The processor will be installed in data centers and is specifically designed for AI inference, the stage that allows for the actual running of AI models2

2

. While much attention in semiconductors has focused on GPUs for training AI models, this central processing unit is built to handle the sequential processing required for AI agents to carry out multi-step tasks on behalf of users2

2

. DAMO Academy stated that the XuanTie CPUs "can be customized for specific inference patterns, supporting customers in tailoring the chips for their own use," with over 30% improvement in performance compared to mainstream products thanks to customization flexibility2

2

.

Technical Specifications and Performance Benchmarks

The chip's technical architecture includes the Tensor Processing Engine supporting data types from FP16 down to INT4/FP8, plus micro-scaling formats MXFP8, MXFP4, and RVFP4, achieving 8 TOPS per TPE1

1

. Its memory subsystem comprises a high-performance multi-level cache hierarchy with an ultralow 4-cycle load-to-use L1 data cache latency, a private per-core L2 cache supporting large capacity configurations, and an MMU with multiple RISC-V virtual memory modes1

1

. The chip features a multi-processor mode that leverages the XL-300 interconnect to form clusters of up to 8 cores1

1

. However, Alibaba's single-core general-purpose performance exceeded 70 points in the SPECint 2006 benchmark test, with a SPECInt 2017 benchmark result of 2.6GHz—performance that Google researcher Laurie Kirk noted puts it nearly on par with Apple's M1 chip launched in 20201

1

.

Strategic Shift Toward Open-Source RISC-V Architecture

The XuanTie C950 is based on open-source RISC-V architecture, which serves as a rival to the CPU blueprint created by British firm Arm2

2

. While companies pay Arm royalties to use its CPU design, RISC-V is effectively a blueprint that can be used for free, offering significant cost effectiveness advantages2

2

. Kirk expressed surprise that Alibaba had implemented version 23.1 of the RISC-V RVA, a minor update proposed in August 2025, so quickly1

1

. This move aligns with China's broader push toward technological self-reliance amid ongoing semiconductor restrictions. Whether Chinese chipmakers can build chips of this caliber in large quantities remains uncertain1

1

.

Bridging the Gap in the Chinese AI Market

Alibaba CEO Yongming Wu acknowledged last week that Chinese chips lag behind those from western chipmakers1

1

. He stated the company's response is "to engage in more profound co-design with Alibaba's cloud infrastructure and the Qwen model to provide improved cost effectiveness. This is one key differentiator ... that sets us apart from other chip companies"1

1

. Native support for the company's own Qwen models suggests Wu's vision for a nicely harmonized AI stack is becoming a reality1

1

. The timing coincides with Alibaba's launch of Wukong last week, its enterprise platform optimized for AI agent workflows, and Accio Work on Monday, as companies throughout China adopt agentic AI platforms3

3

. Alibaba is accelerating its in-house chip development through its T-Head semiconductor arm, with the XuanTie series focused on high-performance cloud systems and agentic AI, while the Zhenwu 810E chip series targets AI training and inference3

3

. This business strategy shift comes as Chinese AI models' token prices have dropped dramatically amid fierce domestic competition, pushing Alibaba to find new ways to ensure profitability3

3

.