Alibaba unveils 3x faster AI chip to challenge Nvidia dominance

Alibaba Group on Wednesday unveiled its Zhenwu M890 AI accelerator, a new chip that delivers three times the performance of its predecessor and is designed to power the next wave of AI agent applications, intensifying its push for domestic technology self-sufficiency.

"Our AI business has officially entered a commercialization payback cycle," Chairman Joseph Tsai and CEO Eddie Wu said in a shareholder letter, adding the company is ramping up investment to build AI and cloud into its next major growth driver.

The Zhenwu M890, equipped with 144GB of high-bandwidth memory, is purpose-built for memory-intensive agentic workloads. It is available immediately to Chinese enterprise customers through a new 128-chip server system. Alibaba also outlined an aggressive roadmap, targeting a successor, the V900, in 2027 and a J900 chip in 2028, signaling a commitment to an annual upgrade cadence similar to Nvidia's.

The move comes as U.S. export curbs block Chinese firms from accessing Nvidia's most advanced processors, a market worth billions. By developing its own powerful alternatives, Alibaba aims to capture a significant portion of the 380 billion yuan ($53 billion) it has pledged for AI and cloud infrastructure, potentially reducing its reliance on foreign technology and boosting long-term profitability.

A Self-Sufficiency Play

The announcement, made at the annual Alibaba Cloud Summit, is a direct response to geopolitical pressures that have threatened to cut off Chinese technology firms from the high-end processors needed for training large AI models. With the Zhenwu M890, Alibaba's T-Head semiconductor unit is making a clear statement of intent to create a viable domestic alternative to Nvidia's market-leading GPUs.

The company said the new chip is particularly well-suited for emerging AI "agents" — software capable of handling complex, multi-step tasks with minimal human input. This requires significant memory and processing power, which the M890's architecture is designed to provide. The new Panjiu AL128 server system, which packages 128 of the M890 chips, is available now on Alibaba's domestic model-as-a-service platform, Bailian.

Roadmap Signals Long-Term Ambition

Beyond the immediate release, Alibaba's multi-year roadmap underscores a long-term commitment to competing at the highest level. The planned V900 chip in 2027 is projected to offer another threefold performance increase, a pace of improvement that mirrors the cadence of industry leader Nvidia. This aggressive schedule is a key part of China's broader strategy to build a self-reliant semiconductor industry, a goal also pursued by competitors like Huawei.

Alibaba's AI ambitions are not limited to hardware. The company also announced Qwen 3.7-Max, its latest flagship large language model. The company claims the model is engineered for advanced coding and can run for up to 35 hours continuously without performance degradation, a key feature for complex, long-running agent tasks.

The investment is substantial. Alibaba has pledged to spend over 380 billion yuan ($53 billion) over three years on its cloud and AI infrastructure. While this level of capital expenditure may weigh on short-term profitability, the market's reaction was initially mixed. Alibaba's Hong Kong-listed shares opened 0.5 percent higher after the announcement but closed down 1.59 percent, reflecting investor uncertainty about the execution risk and long-term return on this significant investment.

This article is for informational purposes only and does not constitute investment advice.