Alibaba's AI Chip Strategy to Bypass Nvidia Restrictions 2026

Alibaba's AI chip strategy is shifting into high gear as the company seeks to reduce its dependence on Nvidia, which has been unable to sell its most advanced chips to Chinese firms due to US export restrictions. The Chinese tech giant recently unveiled its new AI chip, the Zhenwu M890, and outlined an ambitious three-year roadmap that signals a major push toward self-reliance in semiconductor technology.

The move comes amid increasing pressure from Beijing. In early October 2024, the Chinese government recommended that state-owned data centers use at least 50% domestically produced chips in their servers. What began as a recommendation has since evolved into a requirement, accelerating the shift away from foreign technology and boosting domestic suppliers like Huawei, Moore Threads, and Cambricon Technologies. Alibaba, through its T-Head semiconductor division, is positioning itself as a key player in this new landscape.

Alibaba's Zhenwu M890: A Domestic AI Chip

Developed by T-Head, the semiconductor unit Alibaba founded in 2018, the Zhenwu M890 is designed to power AI workloads in data centers. It features 144 GB of HBM3 memory and achieves an inter-chip interconnect speed of up to 800 GB/s. That high-speed interconnect is crucial because it allows multiple chips to work together as a single, more powerful system.

Alibaba's strategy isn't just about building a single chip; it's about creating a scalable infrastructure. By linking many Zhenwu M890 chips together, Alibaba aims to deliver performance that can rival Nvidia's high-end offerings, even if individual chips may not match Nvidia's top models. This approach mirrors the industry trend toward chiplet-based designs and advanced packaging, where multiple smaller chips are combined to achieve greater performance.

Why Interconnect Speed Matters

In AI computing, the ability to move data quickly between chips is often as important as raw processing power. Training large language models and running complex AI inference tasks require massive parallelism, and interconnect bandwidth determines how efficiently multiple chips can collaborate. With 800 GB/s, the Zhenwu M890 is designed to minimize bottlenecks in multi-chip setups, making it suitable for building large-scale AI clusters.

Alibaba Cloud, the company's cloud computing arm, plans to use these chips to offer AI services to its customers. This vertical integration—from chip design to cloud services—gives Alibaba a significant advantage in optimizing performance and cost.

China's Push for Semiconductor Self-Sufficiency

The Chinese government's directive to state-owned data centers is part of a broader effort to achieve technological independence. By requiring at least 50% domestic chips, Beijing aims to nurture its domestic semiconductor industry and reduce vulnerability to foreign sanctions. This policy directly benefits companies like Huawei, which has its own Ascend series of AI chips, as well as Cambricon and Moore Threads.

For Alibaba, the policy creates a guaranteed market for its chips within China. As one of the country's largest cloud providers, Alibaba Cloud can deploy Zhenwu M890 chips in its own data centers and offer them as a service to other enterprises. This aligns with the government's goals and positions Alibaba as a key player in China's AI infrastructure.

• Government mandate: State-owned data centers must use at least 50% domestic chips, driving demand for Chinese-made AI hardware.
• Domestic alternatives: Huawei, Cambricon, and Moore Threads are the leading Chinese AI chip designers, with Alibaba now joining the fray.
• Vertical integration: Alibaba designs chips and operates cloud services, enabling tight optimization and deployment.
• Scalability: High-speed interconnects allow multiple chips to be linked, potentially matching Nvidia's performance for large AI models.

Alibaba's Three-Year AI Chip Roadmap

Alibaba's roadmap outlines a series of chip releases over the next three years, each building on the previous generation. While specific details are scarce, the company has indicated that future chips will offer improved performance, higher memory bandwidth, and better energy efficiency. The goal is to create a comprehensive AI computing platform that can handle everything from training massive models to running inference at the edge.

This roadmap is ambitious, especially given the challenges of developing cutting-edge semiconductors without access to the latest manufacturing equipment. However, Alibaba can leverage its expertise in chip design and its close relationships with domestic foundries like SMIC. The company is also investing in advanced packaging technologies that allow it to combine multiple chips to achieve higher performance.

While the Zhenwu M890's memory capacity exceeds that of the H100, its interconnect speed is slightly lower. However, the comparison is not entirely apples-to-apples, as Nvidia's latest chips, such as the H200 and B100, offer even higher performance. Nevertheless, the Zhenwu M890 represents a significant step forward for China's domestic AI chip industry.

Implications for Nvidia and the Global AI Chip Market

Nvidia has long dominated the AI chip market, with its GPUs being the de facto standard for training and inference. However, export restrictions have forced Chinese companies to seek alternatives, potentially eroding Nvidia's market share in one of the world's largest AI markets. Alibaba's chip efforts, combined with those of Huawei and others, could eventually reduce China's reliance on Nvidia.

That said, Nvidia's software ecosystem, particularly CUDA, remains a formidable moat. Chinese chipmakers will need to develop robust software stacks that can compete with CUDA in terms of ease of use and performance. Alibaba has been investing in its own AI frameworks and tools, but convincing developers to switch will take time.

In the near term, Nvidia will likely continue to dominate outside China, while Chinese companies build up their domestic capabilities. The long-term outcome depends on how quickly Chinese firms can close the technology gap and how effective the US export controls remain. The longer history behind the story is traced in US Semiconductor Companies Thrive in China's Chip Market. More on the surrounding debate is collected in Huawei's New Chip Law Bypasses Moore's Law and US Sanctions.

What to Watch Next

Alibaba's progress with the Zhenwu M890 and its future chips will be closely watched. Key milestones include the chip's commercial deployment, performance benchmarks, and customer adoption. The company's ability to scale production and attract developers to its platform will determine its success. The background to this story is set out in China's RISC-V Strategy to Bypass US Sanctions Gets US Help. A related thread runs through Alibaba Qwen3.8-Max Challenges US AI Chip Restrictions. Earlier coverage in China's AI Infrastructure Strategy Expands to Brazil with Alibaba Cloud shows how the argument developed.

Meanwhile, the Chinese government's policies will continue to shape the landscape. If the 50% domestic chip requirement is strictly enforced, it could create a captive market for Alibaba and its peers, providing the volume needed to justify investments in advanced chip development. A closer look at the same subject appears in Alibaba's New AI Chip Takes Aim at Nvidia's Dominance. Anyone following the wider picture can start with Alibaba Bets Big on AI Agents in Bold New Strategy Move. The same theme surfaces in Alibaba's RISC-V Chips Hit 6 Billion Units in AI Push, which takes a different angle.

For Nvidia, the challenge is to navigate export restrictions while maintaining its technological lead. The company has been developing compliant chips for the Chinese market, but their performance is limited. How Nvidia balances global politics with business interests will be a key story to follow. Context on how this played out before sits in Anthropic Pursues Custom AI Chips to Cut Nvidia Reliance. A companion piece, TOTO's $495M Bet on 1nm Chips: The Hidden AI Giant, approaches the same subject from another angle.