FULL STORY

Alibaba unveils Xuanwu V900 chip and 10T-parameter ambitions

At Apsara Conference 2026, Alibaba unveiled its Xuanwu V900 AI chip, billed as China's most powerful and scalable to 500,000 GPUs, alongside a roadmap toward 10-trillion-parameter models.

2026-09-22 ~ 2026-09-22 · 2 episodes · 9 posts

Episode 1 · Alibaba unveils Zhenwu V900 chip, plans 5-10 trillion parameter model (2026-09-22, 4 posts)

Alibaba announced its Zhenwu V900, billed as China's most powerful AI chip scaling to 500K GPUs per cluster, and reportedly plans a 5-10 trillion parameter Qwen model targeting ASI.

Episode 2 · Alibaba's Cloud Summit: Qwen Eyes 100 Trillion Parameters, 1000x Machine Intelligence Ambition (2026-09-22, 5 posts)

On September 22, 2026, Alibaba Group CEO Eddie Wu unveiled a strategic roadmap for the "era of machine intelligence" at the Apsara Conference 2026 in Hangzhou. The core thesis: machines currently possess less than 3% of human thinking capacity, leaving room for machine intelligence to scale up to 1000x human capability. Reuters and multiple other outlets have confirmed this framing.

Confirmed

  • Models: Alibaba plans to train a new Qwen model with 5 to 10 trillion parameters, far exceeding current mainstream models (m2, m5 citing Reuters).
  • RSI: The Qwen team is exploring recursive self-improvement (RSI), where models identify their own weaknesses from real-task feedback, design experiments, construct data, and continue training (m2, m4).
  • Infrastructure: Alibaba is also building in-house chips and plans 20GW of data centers by 2032 (m3, m4), forming a full-stack commitment across models, chips, and compute.
  • Several posters (m4, m3) summarized this as Alibaba making a "full-stack bet" on the ultra-large model route across three fronts.

Why it matters

  • If 5–10 trillion parameters materialize, it would be several times the size of today's mainstream models, extending the "brute force scaling" route and contrasting with parts of the industry pivoting toward smaller models and inference optimization; its training cost and actual results warrant ongoing attention.
  • RSI suggests the training paradigm could shift from human-curated data to model self-discovery and self-iteration—if viable, it would dramatically reshape the capability growth curve.
  • In-house chips and the 20GW data center plan signal Alibaba's ambition to control the compute layer, with profound implications for China's AI infrastructure landscape.