The new Alibaba AI chip is the Zhenwu V900 — which Alibaba calls the most powerful ever developed in China — unveiled alongside plans to train AI models with 5 to 10 trillion parameters and expand the company’s global data centre capacity past 20 gigawatts by 2032. The announcements, made by CEO Eddie Wu at Alibaba’s annual Apsara conference in Hangzhou on September 22, 2026, mark the company’s biggest push yet to build a fully domestic AI stack — chips, models, and infrastructure — as US export controls keep cutting China off from the most advanced American silicon.
What Alibaba announced at Apsara 2026
Alibaba’s Apsara conference, running September 22–24 in Hangzhou, is the company’s flagship technology event — and this year’s edition was built around one theme: AI at every layer. CEO Eddie Wu’s keynote covered three pillars the company says it is investing across: AI models, chips, and cloud infrastructure.
The headline reveals:
- Zhenwu V900 AI accelerator, from Alibaba’s T-Head semiconductor division, delivering roughly three times the performance of its predecessor, the Zhenwu M890 (released May 2026).
- A plan to develop Qwen models with 5 to 10 trillion parameters, building on the Qwen family whose flagship Qwen 3.8 Max (released July) has 2.4 trillion parameters.
- A target of more than 20 gigawatts of global data centre capacity operated by Alibaba Cloud by 2032.
- Work on Recursive Self-Improvement (RSI) — systems that can improve their own capabilities — as part of the roadmap toward artificial superintelligence.
Investors reacted fast: Alibaba’s Hong Kong-listed shares jumped as much as 5.1% on Tuesday to their highest level in a month.
Alibaba AI chip specs: Zhenwu V900 numbers explained
The new Alibaba AI chip’s published specifications, as reported from the conference, put it firmly in frontier-accelerator territory:
| Specification | Zhenwu V900 | Why it matters |
|---|---|---|
| Memory | 216 GB HBM | Large memory lets the chip hold more of a model’s parameters at once, reducing the need to shuffle data in and out — a key bottleneck in training and inference |
| Inter-chip bandwidth | 1,200 GB/s | Fast links between chips matter when thousands of accelerators train one model as a single system |
| Precision support | FP8 and FP4 native, improved Tensor Core unit | Lower-precision math formats let chips run inference faster and cheaper; FP4 is the frontier for efficient deployment |
| Cluster scale | Up to 500,000 chips per cluster via upgraded supernode servers | A single half-million-chip cluster would support frontier-scale model training |
| Performance vs predecessor | ~3x the Zhenwu M890 | The M890 itself shipped in May and has already moved 560,000+ units to 400+ customers — this is a real product line, not a concept |
| Mass production | Q1 2027 | Manufacturing details were not disclosed |
Two caveats deserve to be stated plainly. First, these are vendor-reported figures — there are no independent benchmarks yet, and “most powerful chip in China” is Alibaba’s own claim. Second, there is a gap between announcement and silicon: mass production is slated for the first quarter of 2027, and the company hasn’t said when it would have enough chips to fill a half-million-card cluster. What is not in doubt is the trajectory: the M890, released just four months ago, has already shipped over half a million units to more than 400 external customers across 20 industries, per reports from the conference. Alibaba is iterating on a product line that is already in commercial deployment.
The 5–10 trillion parameter Qwen plan
Parameters are the variables a model learns during training — a rough gauge of scale and capability. Alibaba’s next-generation target of 5 to 10 trillion parameters would put its ambitions well beyond anything currently deployed: for context, its Qwen 3.8 Max released in July has 2.4 trillion, and Moonshot AI’s model (Alibaba owns a stake in the lab) has 2.8 trillion.
This is a plan, not a launch — the 5–10 trillion parameter model does not exist yet, and third-party reporting suggests Qwen 4 is currently in training, with Qwen 4.5 and Qwen 5 further out. The announcement is best read as a statement of intent about the compute trajectory: training models at that scale requires exactly the kind of infrastructure the V900 and the 20GW data centre target are designed to provide.
Wu framed the ambition in civilizational terms. “The truly groundbreaking products of the Machine Intelligence era have not yet arrived,” he said, comparing today’s AI to the early days of electrification. He predicted machines could eventually generate 1,000 times the “thinking” of all humanity — up from less than 3% today — and described “AI coding” as “simply the light bulb of the machine intelligence era.”
The 20GW data centre push
He set a target for Alibaba Cloud to operate more than 20 gigawatts of global data centre capacity by 2032 — a nation-scale commitment. It’s the same buildout logic driving the AI data centre boom on the other side of the Pacific, where billions are pouring into new AI factories.
He was candid about the constraints, though. “The industry’s mid-to-long-term demand far outpaces our supply capabilities,” Wu said, acknowledging shortages across the global AI data centre supply chain — optical modules, storage, memory, and chips themselves are all limiting how fast capacity can be built.
That supply-demand imbalance is itself a signal: it explains why a cloud provider is spending billions to design its own chips. When you can’t buy enough compute, you build it — and when you can’t buy the chips, you design those too.
Why this matters: China vs Nvidia
The V900 doesn’t exist in a vacuum — it exists because of Washington. Successive rounds of US export controls have limited China’s access to Nvidia’s most advanced chips, leaving the export-compliant H20 — with roughly 15% of the H100’s compute capability — as Nvidia’s main legal offering in China. That pressure has produced exactly what it was designed to prevent: a parallel, domestic AI hardware stack.
Alibaba is not alone in the push. Days earlier, at Huawei Connect in Shanghai on September 17, Huawei laid out an 11-chip AI infrastructure portfolio and pulled its Ascend 960DT accelerator forward to Q1 2027 — the same quarter the V900 hits mass production. Huawei claims Ascend already holds a bigger share of the China market than Nvidia. Zhipu AI’s Z.ai trained its GLM-5 model — competitive with OpenAI’s GPT-5.2 and Anthropic’s Claude Opus 4.5 by some measures — entirely on Huawei Ascend chips, with zero Nvidia hardware. DeepSeek reportedly placed a $2.56 billion order for Huawei Ascend chips.
The honest assessment of where China stands: analysts estimate Huawei’s chips reach roughly 60% of an Nvidia H100’s inference performance, and China remains years behind on training workloads. But the market is tilting toward inference — an estimated 70% of AI compute demand by 2026 — which is exactly where domestic chips are most competitive. Alibaba’s play is broader than Huawei’s: chips plus the Qwen model family plus a global cloud footprint gives it a full stack no other Chinese company controls end-to-end. One more data point from the same month: Beijing blocked Meta’s $2 billion Manus acquisition, a signal that China now treats capable AI assets as national-security property.
The money behind the announcements
The Alibaba AI chip and its surrounding infrastructure push are not cheap — Alibaba is spending at a pace few companies on earth can match:
- RMB 380 billion (about $53 billion) committed over three years to AI and cloud infrastructure — nearly half was already spent by the June quarter.
- ~$10.2 billion raised in August through a Hong Kong share placement — the largest-ever primary follow-on by a Hong Kong-listed company — with 100% of net proceeds earmarked for the “full-stack AI” strategy: chips, infrastructure, models, and deployment.
- $7.1 billion in June-quarter revenue from AI Cloud & Compute Services, up 45% year on year, with AI-related revenue posting its twelfth consecutive quarter of triple-digit growth.
- Adjusted cloud EBITA rose 133% to $830 million — but net profit fell 75% as capex ramped up.
- The expected payback period on AI investments has shortened from 3 years to 2.5 years, per the August earnings call.
The 5.1% share-price jump on Tuesday suggests investors buy the thesis: heavy capital expenditure now, monetized through cloud growth later. The longer-term test is whether Alibaba can keep converting infrastructure — including the new Alibaba AI chip — into revenue faster than the buildout burns cash.
What to watch next
- Independent benchmarks. The V900’s specs are vendor claims until third parties test them. Watch for benchmarks from Chinese AI labs once hardware reaches customers in 2027.
- Qwen 4. The next-generation model in training will show whether Alibaba’s model team can use the scale it is building toward.
- Export-control responses. Every leap in Chinese domestic silicon raises the stakes in Washington’s policy debate over chip restrictions.
- The RSI thread. Recursive Self-Improvement was the keynote’s most tantalizing thread — and its least detailed.
- Supernode deployment. Alibaba says the AI supernodes housing the new Alibaba AI chip go live at commercial scale this quarter — the first real test of whether the domestic stack can deliver frontier training.
FAQ
What is the Alibaba AI chip announced in September 2026?
The Alibaba AI chip is the Zhenwu V900, an AI accelerator developed by Alibaba’s T-Head semiconductor division and unveiled at the Apsara conference on September 22, 2026. CEO Eddie Wu called it the most powerful AI chip in China, claiming roughly 3x the performance of its predecessor, the Zhenwu M890. It is scheduled for mass production in Q1 2027.
What are the Zhenwu V900’s specs?
Published specs include 216 GB of HBM memory, 1,200 GB/s chip-to-chip interconnect bandwidth, native FP8 and FP4 precision support, and an improved Tensor Core unit. Alibaba says the chips can be deployed in clusters of up to 500,000 via upgraded supernode servers. These are vendor-reported figures, not independent benchmarks.
What is Alibaba’s 10 trillion parameter model plan?
CEO Eddie Wu said Alibaba plans to develop Qwen models with 5 to 10 trillion parameters — roughly 2–4x the 2.4 trillion parameters of the current flagship Qwen 3.8 Max. This is a roadmap, not a release; the model does not exist yet. The company is also advancing Recursive Self-Improvement (RSI) research as part of its path toward artificial superintelligence.
Why is Alibaba building its own AI chips?
US export controls have restricted China’s access to Nvidia’s most advanced AI chips, leaving weaker export-compliant variants as the only legal option. Building its own chips through T-Head gives Alibaba a domestic supply of compute for training and running its Qwen models — and positions it against rivals like Huawei’s Ascend line in China’s race for AI hardware independence.
References
- Reuters — Alibaba plans AI model with 5 trillion to 10 trillion parameters, unveils new chip — September 22, 2026
- Morningstar / Dow Jones Newswires — Alibaba Unveils New AI Chip, Outlines Plan for Larger Model — Update — September 22, 2026
- Forkast — Alibaba’s Zhenwu V900 Is China’s Answer to Nvidia’s Absence – and Export Controls Are the Reason It Exists — September 22, 2026
- TechWire Asia — Alibaba AI chip and 20GW plan meet a supply squeeze — September 22, 2026
- ANI via New Kerala — Alibaba unveils Zhenwu V900 AI chip, plans 20GW data centres by 2032 — September 22, 2026
- SkyPress — Alibaba Unveils New AI Chip — September 22, 2026
- Finance Monthly — Alibaba Unveils New AI Chip as Shares Rise 5.1% — September 22, 2026
- CoinCentral — Alibaba (BABA) Stock Jumps 3% After New AI Chip and Data Center Plans — September 22, 2026
- BrandIconImage — Alibaba Unveils Giant AI Model Plans, New Chip as China Races to Challenge Nvidia — September 2026
Leave a Reply