Aditya DidwaniyaChinese tech powerhouse Alibaba has rolled out the Zhenwu M890 a homegrown AI chip three times faster than its predecessor alongside the upcoming Qwen 3.7-Max language model, as China races to build its own AI hardware stack free from US restrictions.
Alibaba Bets Big on Homegrown AI Silicon
Alibaba has taken a bold step forward in the global AI hardware race unveiling the Zhenwu M890, a next-generation AI processor built to directly challenge Nvidia’s dominance in the market. The chip, developed by Alibaba’s semiconductor arm T-Head, delivers three times the computing power of its older sibling, the Zhenwu 810E. With 144 GB of GPU memory and an inter-chip bandwidth of 800 GB per second, the M890 is no small upgrade it’s a statement.
The announcement happened at Alibaba’s annual Cloud Summit, drawing significant attention from enterprise customers and technology watchers alike. Alibaba positioned the new chip squarely within the emerging era of AI “agents” software systems that handle complex, multi-step tasks with minimal human intervention. The M890 is purpose-built for these demanding workloads handling long-context retention and real-time model coordination with ease.
Also Read | US Greenlights Nvidia H200 AI Chip Sales to Ten Chinese Firms
A Full Rack, a New Switch, and a Roadmap That Stretches to 2028
Alibaba didn’t stop at just the chip. Alongside the M890, the company introduced the Panjiu AL128 a supernode server system that packs 128 of these accelerators into a single rack, delivering petabyte-per-second bandwidth. T-Head also launched the ICN Switch 1.0, a switching chip with 25.6 Tbps of aggregate bandwidth enabling full-bandwidth interconnection across 64 accelerators at once.
The company went a step further and laid out a multi-year silicon roadmap. A next chip, called the V900, arrives in Q3 2027 promising yet another roughly threefold performance jump over the M890. Beyond that, the J900 follows in Q3 2028. This signals a steady, confident drumbeat of in-house chip development from a company that once depended heavily on outside suppliers.
Over 60% of T-Head’s chips already go to third-party clients especially in the automotive and enterprise sectors showing strong external adoption beyond Alibaba’s own cloud ecosystem.
Qwen 3.7-Max: The AI Model Built to Run for 35 Hours Straight
On the software side, Alibaba announced Qwen 3.7-Max the latest and most capable version of its flagship large language model series. This model is built specifically for advanced coding, complex reasoning, and long-running agentic tasks. Unlike most AI models that degrade quickly under pressure, Qwen 3.7-Max can autonomously execute tasks for up to 35 hours without any performance drop managing over 1,000 tool calls in a single session.
The model handles code generation, office workflow automation, and complex multi-step tasks requiring hundreds sometimes thousands of individual actions. Developers and enterprises worldwide will soon access it through Alibaba’s Model Studio platform. It supports agent frameworks including OpenClaw, Hermes Agent, Qwen Paw, Qoder, and even Claude Code.
China’s Push to Break Free From Nvidia Dependence
The timing of all of this is no accident. US export restrictions have steadily cut off China’s access to Nvidia’s most advanced processors pushing Chinese tech giants to accelerate their homegrown chip programs. Alibaba, widely seen as a “full-stack AI company” spanning hardware, models, tools, and cloud applications, is betting that self-reliance in silicon is the only long-term answer.
Last year, Alibaba pledged over 380 billion yuan roughly $53 billion on cloud and AI infrastructure across three years. That commitment, the largest in the company’s history, reflects a broader industry-wide wager that demand for AI computing will keep surging as businesses shift to agent-based applications. The M890 and Qwen 3.7-Max are the most visible signs yet that Alibaba is fully prepared to back that bet.
The BRICS Times
Akshay Didwaniya
Pratik Agrawal
