Alibaba AI Chip: Zhenwu V900 Unveiled, 10 Trillion Parameter Plans

The Alibaba AI chip dominating AI headlines does not exist yet. Not in any sense that matters: no customer has one, no lab has benchmarked it, and mass production is not scheduled until the first quarter of 2027. What Alibaba actually unveiled at its Apsara conference in Hangzhou was a promise: the Zhenwu V900, an AI accelerator its CEO called the most powerful ever built in China, plus a roadmap toward AI models with 5 to 10 trillion parameters and data centers drawing more than 20 gigawatts. The gap between the announcement and the silicon is the whole story.

Zhenwu V900: What the New Alibaba AI Chip Actually Is

The unveiling came in CEO Eddie Wu’s keynote at Apsara, Alibaba’s annual technology conference, built this year around a single theme: AI at every layer. Wu framed the strategy around three pillars (AI models, chips, and cloud infrastructure), and the V900 was the chip pillar’s centerpiece. It is designed by T-Head, Alibaba’s in-house semiconductor division, the same unit behind the Zhenwu line that has been shipping to Chinese enterprises for some time.

The company line: the V900 delivers roughly three times the performance of its predecessor, the Zhenwu M890, and Alibaba bills it as the most powerful AI chip developed in China. That superlative is Alibaba’s own. No independent benchmark exists, because no independent lab has the hardware.

Alibaba AI chip specs, as reported from the keynote and subsequent trade coverage:

• 216 GB of memory per chip: enough to hold far more of a model’s parameters on-chip, cutting the data-shuffling bottleneck that slows both training and inference
• 1,200 GB/s chip-to-chip bandwidth: the fast lanes that let thousands of accelerators train one model as a single system
• Native FP8 and FP4 precision support, plus an improved Tensor Core unit: lower-precision math formats that make inference faster and cheaper; FP4 is the current frontier for efficient deployment
• Cluster scale up to 500,000 cards via upgraded supernode servers (a scale target; no operating installation exists yet)
• Volume manufacturing slated for Q1 2027: manufacturing details undisclosed

What separates this from vaporware: the V900 has a predecessor that actually ships. T-Head has moved more than 560,000 Zhenwu-series chips to over 400 customers across 20-plus industries (China Telecom, FAW Group, and Shanghai Pudong Development Bank among the named ones), and Alibaba later raised the customer count past 650. That is a real product line with a real deployment footprint behind it, rather than a concept render. It is also why the Alibaba AI chip announcement landed the way it did: T-Head is selling into a market that already buys from it.

The Specs Are Real. The Benchmarks Aren’t.

Take the Alibaba AI chip story at face value and it reads like a coronation. Read the coverage the way the trade press wrote it and a different picture emerges: TrendForce led with the M890’s shipping numbers and the V900 roadmap, TechWire Asia framed the launch around the supply squeeze, and the Arabian Post led with the share-price pop. An architecture reveal with a production date attached, not shipping silicon. The keynote’s published figures were 216 GB of memory and 1,200 GB/s of interconnect, and “China’s most powerful AI chip” is a company claim no independent test has touched. Nobody outside Alibaba has run code on a V900. That is not an accusation, it is a calendar fact.

The published Alibaba AI chip specs do compare cleanly against the M890 they replace:

All figures vendor-reported. Strip away the adjectives and the Alibaba AI chip story is really about cadence: annual iterations, each roughly tripling the last. Going through the launch coverage side by side, what stood out was that the 3× figure only ever compares the V900 to Alibaba’s own M890, and no outlet could point to a published benchmark behind it. For context on how accelerators actually get compared once they reach buyers, Best Cloud GPU for AI: 8 Providers Compared on Price (2026) breaks down the price-performance lens the benchmarks will eventually use. The relevant question is not whether the V900 beats Nvidia’s flagship on a slide, but what Chinese labs measure once the chips land in their racks next year.

5 to 10 Trillion Parameters: The Qwen Endgame

The other headline number from the presentation was even bigger than the chip. Alibaba plans Qwen models with 5 to 10 trillion parameters. Qwen 4 is in training now; Qwen 4.5 and Qwen 5 are the generations projected to reach that scale. For context, the current flagship Qwen3.8-Max sits at about 2.4 trillion parameters. Alibaba’s preview described it as second only to Anthropic’s Claude Fable 5. A model that size would be roughly two to four times that, and the Alibaba AI chip program exists largely to make training it possible.

This is where a misconception needs correcting, because it circulates in nearly every headline about the plan: more parameters does not automatically mean a better model. Parameters are the variables a model learns during training, a rough gauge of scale. Architecture, training data, optimization, and inference techniques decide what that scale actually buys. A smaller, better-built model routinely beats a bigger sloppy one on real tasks. The 5-to-10-trillion figure is best read as a statement about compute ambition more than a capability guarantee.

The roadmap logic still holds together. Training at that scale is exactly the workload the V900’s half-million-card clusters and the 20-gigawatt data center target are designed for: chip, model, and infrastructure as one plan. And the Qwen ecosystem is genuinely large. Alibaba reported more than 3 billion downloads and 300,000 derivative models. The stated direction of travel is toward artificial superintelligence. The distance between a keynote slide and a model at that scale is measured in years and gigawatts.

20 Gigawatts of… What, Exactly?

Alibaba Cloud wants to operate more than 20 gigawatts of global data-center capacity by 2032. Read that number carefully, because most coverage does not: it is a measure of electrical capacity, not computing power. It says how much power the buildings can draw rather than how many models they can train. Reaching it means years of construction (power delivery, cooling, networking, servers) on a scale that has as much to do with civil engineering as with AI. It is the other half of the Alibaba AI chip bet: silicon without power is a paperweight.

Why it matters anyway:

• Demand is outrunning supply by Alibaba’s own admission: Wu said the industry’s mid-to-long-term demand “far outpaces” what the company can currently supply, with shortages across optical modules, storage, memory, and the chips themselves.
• Citigroup estimates infrastructure at that scale could eventually support roughly $160 billion in external cloud revenue (the bank’s math, not Alibaba’s guidance).
• Wu’s own target: more than $100 billion a year in combined cloud and AI revenue within five years.
• The context is the same buildout logic driving the AI data center boom on the other side of the Pacific. If the gigawatt numbers feel abstract, what AI data centers actually are explains where all that power goes.

The inversion most coverage misses: Alibaba is not building chips to become a chip company. It is building chips because its cloud cannot buy enough compute. The V900 is infrastructure insurance, and 20 gigawatts is the premium.

Washington Built This Chip

The V900 does not exist in a vacuum. It exists because successive rounds of US export controls cut China’s access to Nvidia’s most advanced silicon, leaving weaker export-compliant variants as the legal option. Forkast puts the sharpest version of the argument: the V900 as ‘China’s answer to Nvidia’s absence,’ with export controls as the reason it exists. There is something to that, the pressure did push money and talent into a domestic hardware stack. But the stronger claim, that the controls achieved the precise opposite of their stated goal, runs into this article’s own evidence: Liang Wenfeng puts Huawei’s supernode two years behind Nvidia’s GB300, and Bloomberg’s reporting notes DeepSeek still trains on Nvidia hardware because Huawei’s stack is not yet trusted for the heaviest training runs. The controls manufactured urgency; they have not manufactured parity.

The evidence is piling up faster than the policy can adapt. At Huawei Connect in Shanghai, Huawei pulled its Ascend 960DT accelerator forward to the first quarter of 2027, the same quarter the V900 is meant to hit mass production, and laid out an 11-chip AI infrastructure portfolio spanning compute, interconnect, storage, and management. Zhipu AI’s Z.ai trained its GLM-5 model, a 744-billion-parameter system that rivals GPT-5.2 and Claude Opus 4.5 on some evaluations, entirely on Huawei Ascend chips. Zero Nvidia hardware. DeepSeek placed a roughly $2.56 billion order for 160,000 Huawei Ascend chips for a gigawatt-scale data center in Inner Mongolia, meant for inference workloads; Bloomberg’s reporting notes the lab still trains on Nvidia hardware, because Huawei’s stack is not yet trusted for the heaviest training runs.

The line I kept returning to while reading the reaction wasn’t a headline. It was Liang Wenfeng’s ‘two years behind’ assessment of Huawei’s supernode.

The sharpest public yardstick came from DeepSeek founder Liang Wenfeng: a Huawei supernode, he said, can do everything Nvidia’s GB300 can do, at four Huawei GPUs to one Nvidia card, and two years behind. That is the honest state of the race in one sentence: competitive on inference, trailing on training, closing fast.

And Beijing has started treating AI assets the way it treats oil fields. When Meta agreed to buy the AI-agent startup Manus for about $2 billion, regulators blocked the deal on national-security grounds, ordered it unwound, and barred its co-founders from leaving the country, per the FT’s reporting. The Alibaba AI chip program operates inside that same logic. Compute is now strategic infrastructure, and nobody in Beijing or Washington is pretending otherwise.

The $53 Billion Receipt

The Alibaba AI chip push is not cheap. Alibaba committed more than RMB 380 billion (roughly $53 billion) over three years to AI and cloud infrastructure, and had spent about half of it by the June quarter. A Hong Kong share placement raised another ~$10.2 billion. The quarterly scoreboard, from the company’s June-quarter results:

• Revenue: RMB 269 billion, up 9% year over year: growth is intact at the top line.
• Net profit: RMB 10.5 billion, down 75%: the capex bill arriving all at once.
• AI Cloud & Compute Services: RMB 48.4 billion, up 45%: the AI demand story, in one number.
• Capital expenditure: RMB 67.7 billion, up 75%: the buildout, in another.

The chip division behind the Alibaba AI chip sits at the center of the bet: its annual iteration cadence (M890, V900, then the J900) is supposed to turn that spending into a domestic compute supply the cloud business can actually rely on.

What to Watch Before Believing the Slide Deck

The Alibaba AI chip story will be decided in labs and earnings reports more than at launch events. How to tell when the plan is converting into reality: the same checklist this article’s reporting was built on:

The Bottom Line on the Alibaba AI Chip

Strip away the keynote glow and the Alibaba AI chip story is a bet with a two-year fuse: a threefold generational leap on paper, a half-million-card cluster target, and a 5-to-10-trillion-parameter model roadmap, all converging on hardware no independent lab has touched. If the V900 ships on schedule and the benchmarks hold, it is the moment China’s domestic stack stops being a backup plan. If either slips, it is a $53 billion reminder that announcements are cheap and silicon is not.

FAQ