Alibaba Unveils 'China's Most Powerful AI Chip': 216GB On-Package, Clusters of 500,000
One week after Huawei pulled its chip calendar forward three quarters, Alibaba pulled out its own trump card at the Apsara Conference — China's compute arms race just became a two-horse duel.
Three key facts
On September 22, Alibaba unveiled the Zhenwu V900 AI chip at its Apsara Conference in Hangzhou, billed as China's most powerful: roughly 1.8 PFLOPS peak FP16 per accelerator — three times its predecessor, the M890 — with 216GB of on-package memory (more than Nvidia's H200 at 141GB), chip-to-chip interconnect of about 1.2TB/s, and native FP8/FP4 support. Mass production is targeted for Q1 2027.
The cluster figure is the jaw-dropper: a single Zhenwu V900 cluster can link up to 500,000 chips, for 108 petabytes of aggregate memory. Alongside it, Alibaba announced its infrastructure target: over 20GW of global data-center capacity by 2032.
Behind the chip sits the model roadmap: current flagship Qwen3.8-Max carries 2.4 trillion parameters; Qwen 4 is already in training; Qwen 4.5 and Qwen 5 are slated to scale into the 5–10 trillion parameter range. Chip plus model plus data centers — Alibaba laid its "full-stack autonomy" cards on the table.
Wang Dou's Take
Within a single week: Huawei moved Ascend 960DT three quarters earlier, Alibaba flashed a V900 with 216GB of memory. China's compute story changed scripts — it used to be "single-point breakthroughs to dodge the blockade"; now it's "build the full stack and open a proper storefront." What does 20GW of data centers mean? Roughly enough compute capacity to power the entire population of a mid-sized country — before 2032.
Read the parameter numbers carefully. A 10-trillion-parameter Qwen sounds terrifying, but the industry outgrew "parameters = capability" long ago — under MoE architectures, activated parameters are the real cost. Just last week Arcee won unicorn status with Trinity: 400B total parameters, 13B activated. What deserves scrutiny at Alibaba isn't the headline 10 trillion; it's whether a 500,000-chip V900 cluster can actually run at full efficiency. Interconnect bandwidth can be advertised to the sky, but if cluster efficiency drops by half, the ad was free air. Huawei answered that question with SuperPoD; Alibaba now has to hand in its own paper.
The most delicious detail is the timing: the road for Nvidia's China-specific chips keeps narrowing, and V900 mass production lands in Q1 2027 — exactly the window when Nvidia's next-gen platform ramps and Chinese customers are hungriest for alternatives. That's no coincidence; it's positioning. By 2026 the chip war is no longer about who has the bigger spec sheet — it's about who can fill the racks on schedule in the very quarter the other side cuts supply.
Source: The Next Web · Wccftech · Straits Times · Seeking Alpha