Alibaba Launches XuanTie C950 CPU for Agentic AI

//php echo do_shortcode(‘[responsivevoice_button voice=”US English Male” buttontext=”Listen to Post”]’) ?>

Last week, Alibaba’s DAMO Academy unveiled the XuanTie C950, a 64-bit multi-core CPU designed for agentic AI workloads. It aims to challenge proprietary Western architectures with open-source competition. Alibaba leans heavily into the open-standard RISC-V instruction set architecture.

The company said it pursues architectural sovereignty and shields its operations from volatile U.S. export controls. These controls have severely restricted the flow of advanced Western AI accelerators into China.

Breaking the Western duopoly with Taiwanese manufacturing

To assess the importance of the XuanTie C950 release, it is essential to trace Alibaba’s cadence of microarchitectural refinement over the past several years. Teresa Cervero, RISC-V ambassador and senior research engineer at the Barcelona Supercomputing Center (BSC), told EE Times, “The family of XuanTie processors has been evolving at an amazing pace over the last five years, helping bridge the gap with the market leaders.”

The release of the C920 in 2024 optimized the execution pipeline for cloud workloads, while the debut of the C930 in 2025 was heralded as the industry’s first server-grade RISC-V CPU.

The C950, introduced in 2026, goes beyond previous incremental updates. The processor is verified on an advanced 5-nm process node. Sources indicate that TSMC would manufacture it. It operates at a competitive maximum clock frequency of 3.2 GHz.

The C950’s performance benchmarks consistently deliver a SPECint2006 score above 70 points. This single-thread integer performance shows that a design based on an open-source ISA can compete at the highest levels traditionally monopolized by Intel’s x86 and Arm’s proprietary designs.

“​This is an important milestone for the RISC-V community, especially in the HPC domain,” Cervero argued. “The RISC-V HPC Special Group of Interest has been stating for several years that RISC-V is HPC from both the software and hardware stacks, but the lack of high-performance hardware has undermined this fact… until now.”

Independent technical analyses suggest that the C950’s performance profile and execution efficiency place it nearly on par with Apple’s transformative M1 processor. This indicates Alibaba’s ability to quickly close a multi-decade gap in architectural engineering. This momentum is reflected in the broader market: By January 2026, RISC-V had reached roughly 25% of global processor share, suggesting that the entrenched duopoly between x86 and Arm is beginning to loosen.

The open-standard RISC-V ISA allows chip designers to “customize instruction sets and accelerate specific AI workloads with no or low licensing fees,” Alibaba said.

Alibaba versus HiSilicon and SMIC

Assessing the strategic impact of the XuanTie C950 requires comparing it with Huawei Technologies and its chip design division, HiSilicon. Both aim for computational independence, but differ in their approaches. Huawei uses a proprietary Arm architecture in its Kunpeng 920 CPU, built on 7-nm technology. Alibaba relies on the open-standard RISC-V architecture in its C950. Huawei develops the next-generation Kunpeng 930; the C950 currently outperforms the Kunpeng 920 in single-thread integer throughput, processing over twice as many instructions per clock cycle.

The battlefield is equally fierce in the realm of discrete AI accelerators, a sector long dominated by Nvidia. In response to severe U.S. export controls, Alibaba has shifted from consuming Western technology to producing domestic alternatives. Last year, Alibaba launched the Zhenwu 810E (T-Head parallel processing unit) to compete with Nvidia’s H20 and A800 chips. 

This processor reduces reliance on Taiwanese foundries. Industry reports indicate that a domestic Chinese partner, widely believed to be SMIC, is manufacturing the chip using 7-nm technology. Meanwhile, Huawei introduced the Ascend 950PR for inference tasks and aligned its software to CUDA to attract developers.

Alibaba is following a dual strategy spanning short-term procurement and long-term development. Alibaba indeed plans to place massive orders for Huawei’s new Ascend 950PR to meet its immediate cloud infrastructure needs. At the same time, it aggressively funds its own T-Head subsidiary to build the vertically integrated Zhenwu accelerators and the open-source C950 CPUs.

Engineering for agentic AI

The C950 features native, closely coupled AI acceleration. Traditional computing architectures rely on peripheral buses to offload intensive matrix computations to discrete, power-hungry Nvidia GPUs. Alibaba circumvents this by embedding a proprietary tensor processing engine directly into the CPU core complex through the RISC-V Attached Matrix Extension.

This architectural shift mirrors broader industry movements. “As AI is gaining prominence in HPC, XuanTie addresses the convergence/coexistence of traditional HPC and AI by integrating an AI engine into its new CPU,” Cervero explained. “A trend that is aligned with one of the RISC-V standardization activities for matrix extensions.”

Furthermore, by decoupling the matrix from traditional vector computation, the processor enables greater data reuse, the company claimed. This reduces memory bandwidth demand and improves energy efficiency. The tightly integrated hardware is designed to natively support and execute inference for massive foundation models with hundreds of billions of parameters, such as Alibaba’s Qwen3 and DeepSeek V3.

DeepSeek V3 can now run natively on a RISC-V CPU. The model’s Mixture-of-Experts architecture requires hardware capable of sparse computing and dynamic computational graph optimization. The C950 provides the heterogeneous architectural profile needed to leverage these innovations. It significantly reduces the memory access latency that typically plagues GPU-accelerated inference, the company claimed.

This hardware leap coincides with the viral explosion of agentic AI in early 2026, exemplified by open-source frameworks such as OpenClaw that can autonomously operate computers and manage enterprise workflows. In direct response to this market demand, Alibaba launched “Wukong,” an AI-native enterprise platform integrated into its DingTalk collaboration suite, and “Accio Work,” an international agent designed to automate complex, cross-border business operations. 

Deploying large-scale autonomous agent systems requires low-latency, cost-efficient inference hardware. Alibaba claimed its C950 provides the architectural profile required to operate these platforms profitably, insulating the company from the severe margin compression that comes with relying on expensive Nvidia or Huawei arrays for high-frequency, logic-intensive tasks.

Chinese state support and T-Head potential IPO

Alibaba’s corporate efforts align with state-backed initiatives for technological independence. The DAMO Academy has forged strategic agreements with the Chinese Academy of Sciences to jointly develop the next-generation Xiangshan core. They pool vast resources to tackle complex engineering challenges, such as advanced on-chip interconnects. By collaborating with state academic institutions, Alibaba accelerates and commercializes foundational RISC-V research through its supply chain.

The commercialization of this ecosystem is spearheaded by Alibaba’s semiconductor subsidiary, T-Head, which has achieved cumulative global deliveries of over 470,000 AI chips. Annualized revenue has rapidly approached 10 billion yuan (~$1.45 billion) over the past two years. This growth trajectory has prompted persistent, credible market speculation regarding a potential independent initial public offering for T-Head. 

Furthermore, Alibaba CEO Eddie Wu has explicitly outlined a long-term strategic vision to transition Alibaba into a full-stack AI technology provider, encompassing everything from foundational silicon to upper-layer cloud infrastructure. A successful spinoff and public offering of T-Head would not only unlock massive shareholder value but also inject billions of dollars of capital directly into expanding the Chinese RISC-V ecosystem.

---

See also:

RISC-V Pivots from Academia to Industrial Heavyweight

China Unyielding Ascent in RISC-V

Alibaba Unveils Own AI Chip, Mounting Direct Challenge to Nvidia

RISC-V Solidifies Presence in China as Global Momentum Builds