Alibaba plans a 5–10 trillion-parameter AI model, unveils the Zhenwu V900 chip
On Tuesday morning in Hangzhou, Alibaba put a number on China’s frontier-AI ambitions. Chief executive Eddie Wu told the crowd at Alibaba Cloud’s annual Apsara Conference that the company plans to train a new model with five to ten trillion parameters — two to four times the size of its current flagship, Qwen 3.8 Max. The model was only one leg of the announcement. Wu also unveiled the Zhenwu V900, a new AI chip from Alibaba’s T-Head unit that the company says delivers three times the performance of its predecessor, and committed Alibaba Cloud to more than 20 gigawatts of global data-center capacity by 2032.
On Tuesday morning in Hangzhou, Alibaba put a number on China’s frontier-AI ambitions. Chief executive Eddie Wu told the crowd at Alibaba Cloud’s annual Apsara Conference that the company plans to train a new model with five to ten trillion parameters — two to four times the size of its current flagship, Qwen 3.8 Max. It is the largest parameter target any major lab has attached to a named roadmap this year.
The model was only one leg of the announcement. Wu also unveiled the Zhenwu V900, a new AI chip from Alibaba’s T-Head semiconductor unit that the company says delivers three times the performance of its predecessor, and committed Alibaba Cloud to more than 20 gigawatts of global data-center capacity by 2032. The message, delivered two days before Donald Trump and Xi Jinping sit down in Washington with AI on the agenda, was hard to miss: Alibaba intends to build the biggest models, the chips that train them, and the power that runs them — on its own terms.
What was announced#
Wu said Alibaba’s Qwen team is continuing research into model architecture and data optimisation, aiming at “more complex, longer-horizon tasks” and, in his words, advancing toward artificial superintelligence. According to a company statement, the next-generation Qwen 4 is already in training, while the Qwen 4.5 and Qwen 5 series are projected to scale up to the five-to-ten-trillion-parameter range.
He also claimed Alibaba’s proprietary M890 AI supernode already handles inference for models above two trillion parameters — a capability, he said, that only “a handful of companies globally” possess. And he said the Qwen team has made “meaningful progress” on recursive self-improvement: models that identify their own limitations, design experiments, and synthesise data to drive a cycle of self-evolution.
| Measure | Figure |
|---|---|
| Planned model | 5–10 trillion parameters |
| Current flagship (Qwen 3.8 Max) | 2.4 trillion parameters |
| Zhenwu V900 vs M890 | 3× performance (company claim) |
| Maximum cluster size | 500,000 accelerator cards |
| Zhenwu chips shipped to date | 560,000+ to 400+ customers |
| Data-center capacity target | 20 GW by 2032 |
The chip: Zhenwu V900#
The hardware leg of the announcement may matter more than the parameter count. The Zhenwu V900 is the next generation from T-Head, Alibaba’s in-house chip unit, and Wu called it the most powerful AI chip in China — the company’s claim, not an independently verified benchmark. It delivers three times the performance of the M890, which only shipped in May, and T-Head says a single cluster built on the V900 can support up to 500,000 cards for frontier model training and inference.
This is not Alibaba’s first domestic chip rodeo: T-Head has already shipped more than 560,000 Zhenwu-series chips to over 400 customers, showing real commercial traction beyond Alibaba’s own racks. The company says the V900 moves to mass production and commercial release in the first quarter of 2027, and expects “significant growth” in annual AI chip shipments. The backdrop is well understood across the industry: Chinese firms are racing to build domestic alternatives to Nvidia’s processors as U.S. export restrictions keep the most advanced American chips out of the country.
Why the parameter count matters — and why it doesn’t#
Parameters are the variables a model learns during training, and they are a rough gauge of a model’s size. But size has never guaranteed capability: a bigger model can still lose to a better-trained smaller one, and neither OpenAI nor Anthropic publishes exact parameter figures for its newest frontier models, so any head-to-head comparison against them remains guesswork.
What the announcement does establish is intent — and a spending path to back it. A 20-gigawatt data-center target gives the model claim a physical footprint; the chip roadmap gives it a supply chain. Plenty of companies have announced big models. Far fewer arrive at the podium with the hardware and the power to train them.
The timing is the message#
Wu’s announcement landed on Tuesday. On Thursday, Trump and Xi meet in Washington, with AI, trade, and technology rivalry on the agenda; Sam Altman and Jensen Huang are expected at the White House state dinner. China has spent two years insisting it can build frontier AI without unrestricted access to the best American chips — DeepSeek’s low-cost model made that argument loudly in early 2025, and Alibaba’s Apsara announcement makes it again, this time with a specific number, a named CEO, and a hardware roadmap attached.
What to watch#
- Training timelines. Alibaba gave no dates for when the five-to-ten-trillion-parameter models actually arrive — watch for Qwen 4.5 and Qwen 5 milestones.
- The V900’s first quarter of 2027. Mass production and independent benchmarks will test whether “China’s most powerful chip” holds up outside company keynotes.
- The 20-gigawatt bet. Six years out, with Alibaba itself warning that supply-chain shortages could slow scaling.
- Thursday’s talks. Whether the Trump–Xi meeting produces anything on AI — dialogue mechanisms, export-control shifts, or just theatre.
- Whether rivals answer with numbers. Parameter counts are out of fashion in the West; if Alibaba’s framing sticks, someone may feel compelled to reply in kind.