CEO Eddie Wu detailed a sweeping expansion of China's artificial intelligence capabilities at the Apsara Conference, ranging from new hardware accelerators to massive cloud infrastructure. The roadmap includes scaling the Qwen model family to between 5 trillion and 10 trillion parameters, supported by the Zhenwu V900 AI accelerator. Developed by Alibaba's T-Head unit, the chip delivers 3 times the performance of the previous generation and scales to 500,000 unit clusters, while Alibaba Cloud plans to reach 20 GW of global data center capacity by 2032.
The upcoming Qwen 4 family features Max, Flash, Plus, and 27B versions, utilizing recursive self-improvement to identify weaknesses and synthesize training data. The V900 chip includes 216 GB of memory and a 1,200 GB/s chip-to-chip interconnect, with commercial release scheduled for the first quarter of 2027. Wu stated that machine thinking will eventually exceed human thinking by 1,000 times, noting that AI current capabilities are primarily substituting existing work.
Key sources
- SOURCE@maxforai“包含Qwen4-Max Qwen4-Flash&Qwen4-Plus 还有Qwen4-27B！！！”x.com
- SUPPORT@poezhao0605“Key specs: 216GB memory, 1200GB/s chip-to-chip interconnect, FP8/FP4 low-precision support, 3x performance over the current gen.”x.com
- SUPPORT@poezhao0605“Alibaba Cloud targets 20GW of data center capacity by 2032. That's a 10x increase from 2022.”x.com
- SUPPORT@teortaxestex“Qwen is exploring RSI (recursive self-improvement), enabling the model to identify its own weaknesses based on real task feedback, design experiments, construct data, and then continue training”x.com
- SOURCEmarketbrief.now
- SOURCE@reuters“Alibaba plans to train a new artificial intelligence model with between 5 trillion and 10 trillion parameters”x.com
- SUPPORT@business“targets 20 gigawatts of data center capacity by 2032”x.com
- SUPPORT@choblin29“V900 clusters can scale to 500,000 cards”x.com