Read the Date First: This Is a Roadmap, Not Something You Can Buy
Mass production and large-scale availability are set for the first quarter of 2027. Which means the chip being called the most powerful in China today is still more than a quarter away from being manufactured at volume. That is not nitpicking, it is the basic method for reading announcements like this. A chip event typically spans three states: **what is already running, what is about to enter production, and what sits further out on a roadmap.** The Zhenwu V900 belongs to the second category; the Yitian 720/730 and Zhenwu J900, slated for Q3 2027, belong to the third. So the category that actually supports a judgment is the first — **what is in large-scale commercial deployment is the previous-generation Zhenwu M890**, whose supernode servers are supporting models above 2 trillion parameters, with the Zhenwu series having served more than 650 enterprise customers. That sentence matters more than any new specification, because it describes something that has happened, and because it shows this domestic line is not starting from zero.
The 216GB Matters at the System Level, Not in Single-Chip Benchmarks
Reading specifications alone pulls you into card-versus-card comparisons. But the structure of this announcement is saying something else: the V900 is explicitly designed to work within a system. ICN Switch links more than 1,000 of them into one unified system, and with the new network architecture a single cluster reaches 500,000 accelerators. That is the point of the approach. Without access to the newest imported accelerators, **the way to close a single-chip gap is interconnect and scale** — stitching more of them into a coherent memory domain large enough to train trillion-parameter models and run agent inference at volume. The Yitian 750 supporting the ICN protocol for direct CPU-to-AI-chip connectivity extends the same idea: not winning on one die, but removing a conversion step along the whole path. Which yields a more useful basis for comparison: **how large a coherent memory domain can be assembled for a given budget and power envelope, and what share of total time goes to cross-chip communication.** There is no third-party data on either yet, and no official end-to-end comparison against any specific imported part on a real training workload. Until independent testing exists, "three times the previous generation" is a vendor-stated generational gain and cannot be converted into a position relative to other vendors' products.
Which Numbers Are Promises and Which Are Facts
The figures here need sorting into three categories. **Already happened:** M890 supernodes in large-scale commercial deployment, support for models above 2 trillion parameters, 650-plus enterprise customers, and the day's share price move. **Dated commitments:** V900 mass production in Q1 2027; J900 and the Yitian 720/730 in Q3 2027. **Long-range commitments:** data center capacity above 20 gigawatts by 2032, and more than $53 billion over three years. Note that the 20-gigawatt target is six years out — a direction, not a metric anyone can check in the near term. The model side needs the same treatment: Qwen 4.5 and Qwen 5.0 are described as between 4 trillion and 10 trillion parameters. A range that wide is itself an indication that nothing is settled — roadmap language rather than a product specification. Writing yesterday about SoftBank borrowing $11 billion in junk bonds to buy OpenAI equity, this site noted a shared pattern that applies here too: **in AI infrastructure announcements, the headline number's horizon is routinely much longer than the news cycle.** This is the fourth instance in September alone — Nscale's contracted backlog, Crusoe's $140 billion of contracted value against 1 gigawatt delivered, SoftBank's staged contributions, and now Alibaba's 20 gigawatts by 2032.
Two Boundaries Worth Stating
First, the competitive picture. Huawei's Ascend processors currently dominate China's domestic AI chip market, and Huawei also announced new AI infrastructure last week. This is not a one-company story; it is one of several domestic supply-side tracks advancing at once. Second, two things from the same company can be assessed separately. On September 21 this site covered Qwen-Image-2.1 pulling its license back from Apache 2.0 to research-only, while today's story has Alibaba discussing 4-to-10-trillion-parameter next-generation Qwen models. Chips, cloud and models were presented at one event to tell a vertically integrated story — but for a user, **whether you can use a given model, and under what license, still has to be checked model by model.** The integration narrative on stage does not change what a specific license file says.
via: TechNode, TrendForce, ABC News, Quartz