Alibaba Group's semiconductor design unit on Wednesday unveiled the Zhenwu M890, a new AI chip capable of training and running AI models, as China intensifies its drive to reduce dependence on foreign silicon — particularly chips from Nvidia. The announcement marks a significant step in Alibaba's push to develop homegrown AI hardware capable of competing with restricted US-made accelerators.

The Zhenwu M890: What Alibaba claims

According to the company, the Zhenwu M890 delivers three times the performance of its predecessor chip. Alibaba described the new accelerator as 'exceptionally well-suited' for AI workloads, positioning it as a credible alternative for both training and inference tasks. The chip was developed by T-Head, Alibaba's dedicated semiconductor subsidiary established in 2018.

The performance claims, if verified at scale, would represent a meaningful generational leap in domestic AI silicon — a benchmark the broader Chinese tech industry is watching closely.

Why it matters: Beijing's semiconductor strategy

The launch comes as the Chinese government actively pushes domestic technology companies to accelerate adoption of homegrown AI chips, reducing the country's exposure to US export controls that have progressively restricted access to advanced Nvidia hardware since 2022. China's emphasis on semiconductor self-sufficiency has been a cornerstone of national industrial policy, rooted in the Made in China 2025 framework.

For Alibaba, developing competitive in-house silicon serves a dual purpose: insulating its cloud and AI businesses from supply-chain disruptions while positioning Alibaba Cloud as a credible AI infrastructure provider for enterprise customers across Asia and beyond.

The competitive backdrop

Alibaba is not alone in this race. Across China's internet and cloud sector, major players have expanded custom chip programmes to support growing AI workloads under tightening hardware import constraints. Successive generations of domestic accelerators from multiple vendors are narrowing — though not yet closing — the performance gap with leading US-designed chips.

Nvidia has historically dominated AI training and inference hardware globally, including significant sales to Chinese customers prior to export restrictions. The Zhenwu M890's arrival signals that domestic alternatives are maturing faster than many industry observers anticipated.

What's next

The scale of real-world deployment of the Zhenwu M890 — and whether it can perform at claimed benchmarks across diverse AI workloads — will be the critical test. Analysts will watch whether Alibaba Cloud begins migrating significant internal training runs to the new chip, and whether third-party enterprise customers adopt it at scale.

As China's domestic chip ecosystem matures, the global AI hardware market faces a structural realignment that could reshape competitive dynamics for years ahead.