Alibaba Unveils Zhenwu V900 Chip, Teases 10T-Parameter Qwen 5

Alibaba debuted what it calls China's most powerful AI chip and said its Qwen team will train a 5-10 trillion parameter model on the path to superintelligence.

@sharedot

Alibaba Unveils Zhenwu V900 Chip, Teases 10T-Parameter Qwen 5

By @sharedot · · 8 pages

  • AI
  • Alibaba
  • Semiconductors
  • Chips

Alibaba debuted what it calls China's most powerful AI chip and said its Qwen team will train a 5-10 trillion parameter model on the path to superintelligence.

What Happened at Apsara

At its annual Apsara Conference in Hangzhou on Tuesday, Alibaba introduced the Zhenwu V900 processor, billed as "the most powerful AI chip in China today," and confirmed plans to train its next flagship model with up to 10 trillion parameters.

Group chairman Joe Tsai said Alibaba is "firmly investing in building full-stack AI ... to let AI move from a technological breakthrough to value creation."

CEO Eddie Wu Yongming framed the strategy around three cornerstones — AI models, AI chips and the AI cloud — and said the company aims to scale Alibaba Cloud's global data-centre capacity to more than 20 gigawatts by 2032.

According to the South China Morning Post, Wu described a shift in which machines are becoming "the primary force behind thinking, turning intelligence into a commodity supplied at scale."

Why the Scale Is a Reversal

The parameter target marks a dramatic jump for Chinese frontier models. Alibaba's current flagship, Qwen 3.8-Max, carries 2.4 trillion parameters, while Moonshot AI's Kimi K3 — China's largest open-weight model — holds 2.8 trillion, per the South China Morning Post.

The plan is paired with hardware ambition: the V900 delivers three times the performance of its predecessor, the Zhenwu M890, and a cluster built around it can link up to 500,000 chips for frontier training and inference.

Alibaba also says it expects a "significant" increase in annual AI chip shipments, though executives gave no specific figure.

The Evidence: Specs and Road Maps

T-Head vice-president Gao Hui detailed the V900's specifications at the conference: 216 gigabytes of high-bandwidth memory and inter-chip bandwidth of 1.2 terabytes per second, with mass production and commercial availability scheduled for the first quarter of 2027.

T-Head's road map extends to 2028, including a next-generation Zhenwu J900 chip due in the third quarter of 2028 and two CPUs — the Yitian 720 and Yitian 730 — planned for 2027. The Yitian 730 will be the company's first CPU built on an entirely in-house microarchitecture, according to Yahoo Tech.

On the model side, Qwen project head Liu Dayiheng, in his most prominent public appearance since taking over the project, presented plans for the Qwen 4, Qwen 4.5 and Qwen 5 series and highlighted progress in recursive self-improvement using empirical feedback.

The Stakes: Chips, Rivals and Constraints

The announcements land amid intensifying domestic competition.

Huawei last week unveiled its Atlas 960 SuperPoD computing cluster powered by Ascend 960 chips, with rotating chairman Eric Xu Zhijun expecting a major domestic shift toward Huawei's AI infrastructure for model training next year, per the South China Morning Post.

Export controls remain a complicating factor: the SCMP reported last month, citing sources at major Chinese AI developers, that China's most advanced models were still being trained on Nvidia processors.

Meanwhile, the broader supply chain is accelerating — the Manila Times reports that China's integrated circuit exports reached $216 billion in the first seven months of 2026, up 99.5 percent year on year and surpassing the full-year 2025 total of $201.9 billion, according to customs data.

Adoption in the Wild

Alibaba paired its frontier push with an agent platform aimed at consumers.

The company launched Qwen Intelligence, a full-stack agent platform giving smartphone makers access to Qwen-powered tools for building AI phones that execute tasks across apps — notable given Alibaba's earlier confirmed collaboration with Apple on AI features for iPhones sold in China.

The compact end of the market moved the same day: China Telecom AI released Xing4.0-29B-A4B, an open mixture-of-experts model with 29 billion total parameters and four billion active per token, which the company says scores 75 out of 100 on SWE-bench Verified and runs in 15 GB of graphics memory.

Telecoms Tech News reports the model is already deployed in China Telecom's customer service platform, routing inquiries and planning multi-step remediation procedures.

What Comes Next

Alibaba's Hong Kong-listed shares rose 3 percent to HK$166 by midday following the announcements, according to the South China Morning Post, which also notes Alibaba owns the SCMP.

The near-term milestones to watch are concrete: V900 mass production in the first quarter of 2027, the Yitian 720 and 730 CPUs that same year, and the Qwen 4 series preceding the 5 to 10 trillion parameter Qwen 5 training run. T-Head's Zhenwu J900 follows in the third quarter of 2028.

The Manila Times quotes analyst Liu Gang cautioning that "having an innovative chip architecture alone is not enough if manufacturing capacity cannot keep up," underscoring that sustained progress will depend on advances across the full semiconductor chain, not just design wins at flagship events.

Sources

  1. scmp.com › Alibaba debuts 'China's most powerful' AI chip, teases 10-trillion-parameter model
  2. tech.yahoo.com › Zhenwu V900: Alibaba's most powerful chip links 500,000 units, eyes 10-trillion-parameter AI
  3. manilatimes.net › China's chip push gains momentum as breakthroughs spread from manufacturing to AI processors
  4. telecomstechnews.com › China Telecom AI launches Xing4.0-29B-A4B open model

More on AI