Alibaba used its September 22 developer conference to show how quickly China’s AI stack is filling the holes left by restricted access to U.S. chips. The company unveiled the Zhenwu V900 accelerator, which it says delivers roughly three times the performance of its predecessor and is scheduled for mass production in early 2027. Alibaba Group Holding Limited (NYSE:BABA) also discussed future Qwen models scaling to 5 trillion to 10 trillion parameters and its plans to expand global data-center capacity above 20 gigawatts by 2032. For NVIDIA Corporation (NASDAQ:NVDA), the issue is not today’s China revenue. It is whether a protected local ecosystem becomes good enough that customers stop waiting for Nvidia to return.
Alibaba is building the full stack at once
Alibaba Group Holding Limited can use a homegrown accelerator across its cloud, models and internal applications, which gives it a captive workload before it ever sells the chip broadly. That vertical integration is the bull case. If Zhenwu performance improves while Qwen demand rises, Alibaba can capture cloud revenue without paying an outside GPU vendor for every incremental unit of compute. The 20-gigawatt capacity target suggests the company is preparing for a much larger infrastructure footprint.
Pixabay/Public Domain
The obvious limitation is that the V900’s performance claims come from Alibaba and mass production is still expected in 2027. A chip can benchmark well and still struggle with yields, software tooling, memory supply or developer adoption. Alibaba is also financing AI aggressively while its core commerce business must keep producing the cash that pays for the buildout.
Nvidia’s China problem is becoming structural
NVIDIA Corporation still owns the stronger global platform. CUDA, networking and its pace of system-level launches make replacement hard, and export restrictions have already reduced Nvidia’s ability to serve China’s highest-end demand. That means Alibaba’s progress may initially replace sales Nvidia is largely unable to make under current restrictions anyway. The longer-term bear case is different: local customers are being forced to optimize around domestic hardware, turning a temporary supply constraint into permanent ecosystem learning.
Insider Monkey’s database showed 97 hedge funds with reportable Alibaba longs in Q2 2026, down from 102 in Q1. Fisher Asset Management held about 5.10 million shares after trimming its position 1%. Nvidia had 285 hedge-fund holders, up from 275, and Fisher increased its Nvidia stake about 3%. Those filings predate the V900 launch. Alibaba short interest was about 41.98 million shares as of August 14, roughly 2.0% of public float with around 4.6 days to cover.
Alibaba does not need to beat Nvidia everywhere for the stock-level read-through to matter. It needs a chip that is good enough for a growing share of Chinese AI workloads. Every generation that narrows the gap lowers the odds that Nvidia simply walks back into its old China position if policy changes later.
READ NEXT:
The Trillion-Dollar AI Capex Gap: Why Too Much Hardware Could Be Nvidia’s Trap and Microsoft’s Opportunity and Redditors Are Skeptical of SpaceX’s Orbital Data Centers. Are They Right?
