Companies
AI chips / China
Alibaba's vertical moonshot: the chip it calls China's best and models four times bigger
The Zhenwu V900 promises three times the performance of its predecessor and clusters of up to 500,000 chips — while Alibaba's cloud races toward 20 gigawatts of capacity by 2032. Hong Kong-listed shares jumped 5.1%.
TickerGrove Newsroom · Updated
Updated Sep. 22, 2026 at 4:49 PM ET to revise deck wording and Sources presentation.
Sources
Reuters: Alibaba plans AI model with 5 trillion to 10 trillion parameters, unveils new chip. Finance Monthly: Alibaba AI push lifts shares as new chip and model plans emerge.
All dates 2026. Announcements made Tuesday September 22; Zhenwu V900 mass production targeted for Q1 2027; 20-gigawatt Alibaba Cloud capacity target set for 2032; Qwen3.8-Max launched August 2026; HK$80B placement completed August 2026.
Alibaba Group said on Tuesday it was developing an AI model up to four times larger than the company's flagship model and unveiled the Zhenwu V900, a next-generation AI chip it called China's most powerful — sending its Hong Kong-listed shares up 5.1% to their highest in a month.
The double announcement is the whole strategy in one package: at Alibaba Cloud's annual Apsara conference in Hangzhou, Alibaba unveiled the Zhenwu V900, a new AI chip developed by its T-Head semiconductor unit, alongside a model roadmap that puts the company on a path to 5 trillion to 10 trillion parameters — up to four times the size of its current flagship.
On the silicon side, the numbers are aggressive. Chief Executive Eddie Wu said the Zhenwu V900 delivers three times the performance of its predecessor, the M890 — which Alibaba launched in May — and can be linked in clusters of up to 500,000 chips to train and run the largest AI models.
The chip is set for mass production and commercial release in the first quarter of 2027, and Wu said the company expected "significant growth" in annual AI chip shipments — Alibaba's clearest signal yet that it intends to be a merchant supplier of AI silicon, not just a captive consumer of it.
On the model side, Alibaba said it is currently training its next-generation model, Qwen 4, with future Qwen 4.5 and Qwen 5 models expected to scale up to 5 trillion to 10 trillion parameters. The arithmetic checks out against the current flagship: August's Qwen3.8-Max launched with 2.4 trillion parameters, so four times larger lands right around the 10-trillion mark.
Why it matters: Alibaba is betting on the full vertical stack — its own AI chips, its own cloud, and its own frontier models — a playbook that puts it in the unusual position of competing with Nvidia on silicon, the hyperscalers on compute, and the model labs on capability, all at once.
The cloud leg of the bet is getting the heaviest capital. Wu set a target for Alibaba Cloud's global data-centre capacity to surpass 20 gigawatts by 2032, and said the company would begin bringing its AI supernodes online at commercial scale this quarter.
Wu said customer demand for AI was "exceptionally robust" and was accelerating Alibaba Cloud's revenue growth, though supply-chain constraints limit the pace of expansion. "The industry's mid-to-long-term demand far outpaces our supply capabilities," Wu said — the kind of sentence that justifies a buildout. It is also the bull case in one line: if demand keeps outrunning supply, whoever has the chips and the power wins.
The funding is already lined up. In August, Alibaba raised HK$80 billion through a placement of 710 million shares at HK$112.70 apiece, with 60% earmarked for global computing infrastructure and 40% for hyperscale AI data centres — the war chest behind the 20-gigawatt target.
The "most powerful in China" label deserves the framing Alibaba gave it: it is the company's own claim, and no independent benchmark has been published. That caveat matters in a year when every chip launch comes with a superlative attached.
Still, investors liked what they heard. Alibaba's Hong Kong-listed shares surged 5.1% on Tuesday to their highest in a month — a bet that the vertical stack is worth more than the sum of its parts.
The unknowns are material. Benchmarks for the V900 have not been published, pricing and actual shipment volumes are unconfirmed, and Qwen 4's capabilities will be judged when it ships, not when it is announced. The chip itself will not reach mass production until the first quarter of 2027.
What to watch: whether Qwen 4 training stays on schedule, whether the AI supernodes come online this quarter as promised, how the first V900 production milestones track toward 2027, and whether the 20-gigawatt target survives its first contact with real capex cycles.
What parameters, AI chips, and gigawatts mean — and why Alibaba's announcement moved the stock
AI models like the ones Alibaba is building are measured in "parameters" — the adjustable settings inside the model that get tuned during training. More parameters generally means a bigger model that can learn more complex patterns. An AI chip is a processor built for the particular math that training and running these models requires. Alibaba designs its chips through its T-Head semiconductor unit rather than buying them all from outside suppliers. To train a frontier model, companies link thousands of chips together into a cluster so they can split the work. Alibaba says the Zhenwu V900 can be linked in clusters of up to 500,000 chips — its claimed ceiling for training the largest models. A gigawatt is a billion watts of electrical power. Alibaba's target of more than 20 gigawatts of data-centre capacity by 2032 describes an enormous buildout of the physical buildings, power, and cooling that AI computing requires. When Alibaba raised HK$80 billion in August, it sold 710 million new shares to investors at HK$112.70 each — a share placement. Sixty percent of the proceeds are earmarked for global computing infrastructure and 40% for hyperscale AI data centres: the money behind the expansion. A 5.1% jump in the share price means investors collectively decided Tuesday's news made Alibaba more valuable — in this case, a bet that owning the chips, the cloud, and the models is worth more than the parts.
Vertical-stack economics: what 10 trillion parameters actually cost to train
The vertical-stack economics are the point. Captive silicon removes a margin layer, secures supply against export-control shocks, and lets Alibaba tune hardware to its own model architectures — the hyperscaler playbook, except Alibaba is also running one of the world's most competitive open model families. The 4x arithmetic is the tell on ambition. Qwen3.8-Max at 2.4 trillion parameters is already one of the largest openly discussed models; a 10-trillion-parameter Qwen 5 would move the frontier conversation from "how big" to "what does the training run cost" — the compute bill for a model that size dwarfs anything disclosed to date. The 500,000-chip cluster ceiling is best read as intent, not inventory. Nobody has disclosed operating a half-million-accelerator training fabric; the number signals that Alibaba is designing the V900's interconnect for frontier-scale training rather than inference appliances. The financing sequence is disciplined: the HK$80 billion August placement pre-funds the capex before the announcements, with 60% directed to global computing infrastructure — the power-and-buildings layer that constrains every hyperscaler's AI plans. The risk ledger: no independent V900 benchmarks, no pricing, no volume guidance beyond "significant growth," and a mass-production date 15 months out. Qwen 4 will be benchmarked on release, not on announcement — as every model is. The structural read: Wu's demand-outpaces-supply framing is both a customer pitch and a capital-markets pitch. If he is right, the 20-gigawatt target is the moat; if the cycle turns, it is the overhang.
Not yet known
Whether the V900 hits its Q1 2027 mass-production target and at what volumes and pricing; whether Qwen 4 training stays on schedule and how Qwen 4.5/5 benchmark on release; whether the AI supernodes reach commercial scale this quarter; whether the 20-gigawatt capacity target survives real capex cycles; whether independent benchmarks support the 'most powerful in China' claim.
Document trail
Sources & evidence
Sources used for this piece.
Corrections
We do not silently rewrite a published line. Material corrections receive a visible correction note, and we preserve the article’s update history.
Discuss this story. Join the TickerGrove community to talk companies, earnings, and markets, or request future coverage.