Invest12 publishers2 min readPublished
Alibaba pulls its V900 chip forward two quarters to the first quarter of 2027
The Zhenwu V900 sat on T-Head's roadmap for the third quarter of 2027 when the M890 shipped in May, and Eddie Wu now says mass production starts in the first quarter, with Hong Kong shares up as much as 5.1%.
The Investor · Invest desk
What happened
- Eddie Wu unveiled the Zhenwu V900 at Alibaba's Apsara Conference in Hangzhou on September 22, a chip from its T-Head unit that the company says delivers three times the processing power of the M890.
- Alibaba put mass production and commercial release in the first quarter of 2027, two quarters earlier than the third-quarter 2027 date on the roadmap it published when the M890 launched in May.
- Wu also said Alibaba will train a Qwen model of 5 trillion to 10 trillion parameters, about two to four times the 2.4 trillion in the Qwen 3.8 Max flagship released in July.
- Alibaba Cloud is targeting more than 20 gigawatts of global data centre capacity by 2032, Wu said, describing the buildout as an AI cloud for what he called the Machine Intelligence era.
- Alibaba's Hong Kong-listed shares rose as much as 5.1% during Tuesday trading while the Hang Seng Tech Index gained 0.8%.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- contradiction Morningstar reports the $53 billion as a commitment to spend over three years, while Cryptobriefing describes it as committed over the past three years, and anyone modelling Alibaba's capex has to pick one before deciding whether the money is ahead of it or behind it.
- decision With roughly 60% of T-Head's production capacity serving external clients such as China Telecom and FAW Group, every wafer sold outside is one not racked in Alibaba's own cluster, and that split is a choice Alibaba is making while it argues it needs silicon for superintelligence.
- constraint No V900 can carry a training run before its first-quarter 2027 production date, so whatever Alibaba trains between now and then runs on the M890 and on accelerators it already owns.
- capability If the date holds, Chinese buyers get a second announced path to cluster-scale training beside Huawei's Ascend, on a cadence of claimed three-times steps roughly every four months.
T-Head's release cadence makes the earlier date at least arguable: the Zhenwu 810E shipped earlier in 2026, the M890 in May, and Wu was on stage with the V900 four months after that [10][2]. The multipliers compound. The M890 tripled the 810E, and the same multiple applied again puts the V900 at roughly nine times the part T-Head was selling at the start of this year [4][3][5]. The J900 is still down for the third quarter of 2028 [8].
The cluster ceiling and the sales ledger sit oddly together. One V900 cluster can run up to 500,000 cards as a single training system, according to Alibaba [6], and cumulative Zhenwu shipments across every generation stand at 560,000 units, sold to more than 400 customers in over 20 industry sectors [12]. Filling one cluster at that advertised size would take about 89% of everything T-Head has ever shipped [15]. The average customer to date has bought about 1,400 chips [14].
Alibaba did not give spending details for the new chip and model plans. Last month it raised about $10.2 billion in a Hong Kong share offering earmarked exclusively for AI investments, roughly a fifth of the $53 billion three-year figure [19][20][16].
The pressure here is domestic as much as it is American. US export restrictions block Chinese companies from Nvidia's most advanced processors [31], and at the same time DeepSeek and Moonshot AI keep producing breakthroughs while ByteDance spends heavily to keep up [35]. Moonshot, a lab in which Alibaba owns a stake, released a 2.8-trillion-parameter model in July, about 17% larger than Alibaba's own flagship [24][36]. "We anticipate significant growth in annual AI chip shipment volumes," Wu said in Hangzhou [26].
Tuesday's 4.3 points of relative move against the Hang Seng Tech Index [30] bought exposure to a chip that enters production next year and a model that is still a plan. I would treat the announcement as a dated supply claim and weigh the date accordingly: the nine-times figure is Alibaba's own estimate measured against its own earlier part, while the first-quarter slot is checkable. Two things would break that reading. If the V900 slides back toward the third quarter of 2027, the cadence argument goes with it; if a named external buyer takes cluster-scale volume before production starts, the 1,400-chip average stops describing the customer base. Huawei's Ascend line still leads the domestic market, with Cambricon and Baidu's Kunlunxin also selling into it [28].
What to watch
- Whether the 5-to-10-trillion-parameter Qwen model, when it appears, was trained on T-Head silicon or on imported accelerators.
- Alibaba's next capex disclosure, which would settle whether the $53 billion is money already spent or spending still to come.
- Any movement in the J900's third-quarter 2028 date, the one roadmap point still sitting behind the V900.