Alibaba Unveils High-Performance RISC-V Chip as AI Agent Race Shifts to Hardware

Alibaba Group has pushed further into the semiconductor race, unveiling a new high-performance processor it says is tailored for the emerging class of AI agents.

This comes as China’s largest tech firms compete to build the infrastructure underpinning the next phase of artificial intelligence.

The chip, XuanTie C950, was introduced in Shanghai by the company’s research arm, Damo Academy, and is being positioned as the most powerful CPU core in Alibaba’s RISC-V lineup, according to SCMP. The company claims the processor delivers more than three times the performance of the previous-generation C920, a jump that signals a shift from incremental upgrades to more aggressive architectural gains.

The technical specifications point to that ambition. With an 8-instruction decode width and a 16-stage pipeline, the C950 is designed to process a significantly higher volume of instructions in parallel, a requirement for AI workloads that depend on rapid data throughput and low latency. In practical terms, Alibaba is targeting the core bottlenecks that have constrained the deployment of large-scale AI systems, particularly those designed to operate autonomously.

That positioning comes amid the rise of AI agents, systems capable of executing tasks with limited human input, which has quickly become a new competitive front, especially after the emergence of OpenClaw. Chinese firms, including Tencent Holdings and ByteDance, have accelerated efforts to integrate such capabilities into consumer and enterprise platforms, driving demand for chips optimized for inference-heavy, always-on computing environments.

Alibaba’s approach ties hardware directly to that shift. The C950 is engineered to handle cloud-native workloads such as MySQL, Redis, Nginx, and OpenSSL, but its strategic value lies in its ability to support inference for large language models like Qwen and DeepSeek when paired with Alibaba’s proprietary acceleration engines. Those engines, the Vector Acceleration Engine and Matrix Acceleration Engine, are designed to offload and speed up parallel computations, effectively turning the CPU into part of a broader, tightly integrated AI computing stack.

The emphasis on inference is notable. While much of the global AI race has focused on training large models, inference—the process of running those models in real-world applications—is emerging as the more commercially critical phase. It is also more sensitive to cost and energy efficiency, areas where Alibaba is attempting to differentiate. Alongside the C950, the company introduced the C925, a variant aimed at improving performance per watt, a metric that is becoming central as data centers face rising power constraints.

Underpinning the entire strategy is the choice of architecture. The C950 is built on RISC-V, an open-standard instruction set governed by RISC-V International. Unlike proprietary architectures such as Intel’s x86 or ARM’s designs, RISC-V can be freely used and modified, allowing companies to tailor chips to specific workloads without licensing restrictions.

For Alibaba and its domestic peers, that flexibility carries weight beyond engineering. U.S. export controls have limited China’s access to advanced semiconductor technologies, particularly in high-end processors and manufacturing equipment. By leaning into RISC-V, Alibaba is aligning itself with a broader national push to reduce dependence on foreign chip ecosystems and build a self-sustaining supply chain.

The company has been laying the groundwork for several years. It entered the RISC-V space in 2018 with the launch of the XuanTie series and has since expanded adoption through chips like the C910 and C920. The introduction of the C930 last year marked its first move into server-grade CPUs, signaling a shift from embedded and edge applications toward data center infrastructure. The C950 extends that trajectory, suggesting Alibaba is now targeting the upper tiers of performance computing.

Industry observers say the timing reflects a broader transition within the RISC-V ecosystem itself. Lu Dai, chair of RISC-V International, noted that the technology is moving from a development phase into large-scale deployment, with increasing focus on building out software, tooling, and commercial applications around the architecture.

“The architecture was now undergoing a shift from ‘purely technology development’ to deployment, putting a greater focus on the ecosystem, he added.

Alibaba’s own numbers indicate that its chip business is beginning to scale. Its semiconductor unit, T-Head, has shipped more than 470,000 AI chips as of February and is approaching 10 billion yuan (about $1.45 billion) in annual revenue over the past two years. While still modest compared with global chip leaders, the figures point to growing traction in a market where domestic alternatives are gaining urgency.

What is becoming clear is that the contest over AI is no longer confined to software models or consumer applications. Control over the underlying hardware, particularly architectures that can be adapted, scaled, and insulated from geopolitical constraints, is emerging as a defining factor.

With the C950, Alibaba is seen to be making a calculated bet that the future of AI agents will depend so much on chips.