In Hangzhou, Alibaba's chief executive stood before the company's annual cloud gathering and announced an ambition that reaches well beyond quarterly earnings: an artificial intelligence model of five to ten trillion parameters, paired with a homegrown chip three times more powerful than its predecessor. The announcement arrives at a moment when American export restrictions have severed Chinese technology firms from their traditional suppliers, turning necessity into invention. Alibaba is not merely responding to a market gap — it is attempting to lay the computational bedrock for what its lea
Alibaba unveils 5-10 trillion parameter AI model and next-gen chip amid race for domestic alternatives
The truly groundbreaking products have not yet arrived.
Mark
So Alibaba is building a model five to ten times larger than what they have now. What does that actually let you do that you couldn't do before?
Mimi
The size of a model—measured in parameters—roughly correlates with its ability to handle complex reasoning and longer chains of thought. A bigger model can theoretically tackle more intricate problems, hold more context in mind, and approach what researchers call artificial superintelligence.
Luke
But we should be careful here. "Roughly correlates" is doing a lot of work. Bigger doesn't automatically mean better. Training efficiency, data quality, and architecture matter enormously. The source doesn't tell us how Alibaba's approach compares on those dimensions.
Mark
Fair point. What about the chip—the Zhenwu V900? Why does that matter?
Mimi
You can't train or run a model that large without the hardware to support it. The V900 delivers three times the performance of Alibaba's previous chip and can support up to 500,000 cards in a cluster. That's the kind of scale you need for frontier model work.
Luke
Again, three times the performance—compared to what, exactly? The M890, which only came out in May. We don't have independent benchmarks here. We're taking Alibaba's word for it.
Mark
Is that a problem?
Luke
It's worth noting. Companies have incentives to make their hardware sound impressive. The real test will be whether other companies adopt it, whether it actually works at scale, whether it delivers on the promise. We won't know that until Q1 2027 when it ships.
Mimi
True. But the broader context is real: U.S. export restrictions have made it hard for Chinese companies to get Nvidia chips, so they're forced to build their own. That's not speculation—that's the constraint they're operating under.
Mark
Wu also talked about demand outpacing supply. Is that just Alibaba's problem, or is it industry-wide?
Mimi
Industry-wide. He said the global AI data centre supply chain is the bottleneck. Alibaba wants 20 gigawatts of capacity by 2032, but they can only build as fast as the supply chain allows.
Luke
Which is a useful reminder that this isn't just about innovation. It's about manufacturing, logistics, power infrastructure, cooling systems. The sexy part is the model size and the chip performance. The real constraint is unglamorous infrastructure.
Mark
And Wu's prediction about machines producing 1,000 times humanity's thinking—how confident should we be in that?
Luke
Not very. That's a vision statement, not a forecast. It's aspirational. The source doesn't give us any technical basis for that number. It's Wu's framing of where he thinks the industry is headed.
Mimi
But it does tell you what Alibaba's leadership believes the opportunity is. They're not just building incrementally. They're betting on a fundamental shift in what machines can do.
Mark
So what's the real story here?
Mimi
China's largest tech company is making a serious, well-capitalized bet that it can build the infrastructure for frontier AI independently of American suppliers. They're moving fast—new chips every few months, models scaling rapidly, data centres expanding.
Luke
And we should watch whether it works. Whether the Zhenwu V900 actually performs as promised. Whether the Qwen models at that scale actually deliver on the reasoning and superintelligence claims. Whether the supply chain bottleneck eases or gets worse. Those are the real tests ahead.
El Pulso
- U.S. export controls have cut off Chinese firms from Nvidia's most advanced chips, creating an urgent race to build sovereign AI infrastructure before the gap becomes permanent.
- Alibaba's planned Qwen models — two to four times larger than its current flagship — represent a generational leap that most companies globally lack the hardware to even attempt.
- The new Zhenwu V900 chip, capable of linking 500,000 cards in a single cluster, is Alibaba's answer to that hardware problem, though commercial availability remains months away in early 2027.
- Even with the chips and models in hand, a deeper constraint looms: global demand for AI computing so far exceeds supply that Alibaba's own expansion will be throttled by what can physically be built.
- The company is targeting 20 gigawatts of data centre capacity by 2032, a signal that this is a decade-long infrastructure wager, not a product announcement.
In Hangzhou, Alibaba's chief executive stood before the company's annual cloud gathering and announced an ambition that reaches well beyond quarterly earnings: an artificial intelligence model of five to ten trillion parameters, paired with a homegrown chip three times more powerful than its predecessor. The announcement arrives at a moment when American export restrictions have severed Chinese technology firms from their traditional suppliers, turning necessity into invention. Alibaba is not merely responding to a market gap — it is attempting to lay the computational bedrock for what its leadership calls the Machine Intelligence era, a transformation they compare, without apparent irony, to the Industrial Revolution.
On Tuesday, Alibaba chief executive Eddie Wu took the stage at the company's annual cloud conference in Hangzhou and described a model of artificial intelligence so large it would make the company's current systems look modest — somewhere between five and ten trillion parameters, two to four times the size of its existing flagship. He framed the ambition as part of a longer arc toward artificial superintelligence, noting that Alibaba's Qwen team has been working on recursive self-improvement, a process in which models examine their own limitations and redesign themselves accordingly.
The announcement did not arrive in a vacuum. U.S. export restrictions have made Nvidia chips increasingly difficult for Chinese companies to obtain, accelerating the push to build domestic alternatives. Alibaba's answer is the Zhenwu V900, a next-generation chip developed by its in-house T-Head semiconductor unit. It delivers three times the performance of its predecessor, supports clusters of up to 500,000 cards, and is scheduled for commercial release in the first quarter of 2027 — a striking pace given that the previous generation only shipped in May.
Wu was candid about the constraints. Demand for AI computing, he said, far outpaces what the industry can supply, and Alibaba's own expansion will be limited by what can physically be constructed and deployed. Still, the company has set a target of 20 gigawatts of global data centre capacity by 2032, a commitment that signals this is a long-horizon infrastructure bet rather than a near-term product play.
Wu closed by placing the moment in sweeping historical terms, comparing the Machine Intelligence era to the Industrial Revolution and predicting that machines would eventually generate more than a thousand times the cognitive output of all humanity combined. The truly transformative products, he acknowledged, have not yet arrived. For now, Alibaba is building the foundation on which they might.
Alibaba is building bigger. On Tuesday, Chief Executive Eddie Wu stood before the company's annual cloud conference in Hangzhou and announced plans for an artificial intelligence model so large it would dwarf anything the company has built before—somewhere between 5 trillion and 10 trillion parameters, roughly two to four times the size of its current flagship system. The move signals how seriously China's largest technology companies are moving to construct their own AI infrastructure, independent of American suppliers and the export restrictions that have made Nvidia chips increasingly difficult to obtain.
The scale matters because parameters are the variables a model learns during training, and they function as a rough measure of a system's capacity and capability. Alibaba's existing Qwen 3.8 Max model operates at 2.4 trillion parameters. The new generation—Qwen 4, already in training, with successors Qwen 4.5 and Qwen 5 planned—represents a significant leap forward. Wu framed this expansion as part of a longer journey toward what researchers call artificial superintelligence, or ASI. The Qwen team, he said, has been making progress on what's known as recursive self-improvement, a process in which models identify their own limitations, design experiments to test themselves, and use the results to evolve on their own.
But building a model of that size requires infrastructure most companies simply do not possess. Alibaba already operates what it calls the M890 AI supernode, a piece of hardware Wu said can handle inference—the computational work of running a trained model—for systems above 2 trillion parameters. He noted that only a handful of companies globally have this capability. Now the company is introducing its successor: the Zhenwu V900, a next-generation chip designed and built by Alibaba's T-Head semiconductor unit. The new chip delivers three times the performance of the M890 and can support up to 500,000 cards in a single cluster, the kind of scale required for training and running frontier models. Mass production and commercial availability are scheduled for the first quarter of 2027. The M890 itself only arrived in May.
This rapid iteration reflects the intensity of competition among Chinese technology firms to build domestic alternatives to American processors. U.S. export restrictions have made it increasingly difficult for Chinese companies to acquire the chips they need, forcing them to accelerate their own development timelines. Wu said Alibaba expects significant growth in annual AI chip shipments, a signal of confidence in both the technology and the market demand.
The infrastructure push extends beyond chips. Wu set a target for Alibaba Cloud's global data centre capacity to exceed 20 gigawatts by 2032, a massive expansion from current levels. He acknowledged, however, that the company faces a fundamental constraint: demand for AI computing far outpaces what the industry can currently supply. "The industry's mid-to-long-term demand far outpaces our supply capabilities," Wu said. Alibaba Cloud is beginning to bring its AI supernodes online at commercial scale this quarter, but the shortage across the global AI data centre supply chain means expansion will remain limited by what can physically be built and deployed.
Wu positioned this moment as the beginning of what he called the "Machine Intelligence" era, comparable in historical significance to the Industrial Revolution. He predicted that machines would eventually produce more than 1,000 times the cognitive output of all humanity combined, up from less than 3 percent today. He drew a distinction between early applications—he cited AI coding as analogous to the light bulb in the electrical age—and the truly transformative products still to come. "The truly groundbreaking products of the Machine Intelligence era have not yet arrived," he said. For now, Alibaba is laying the foundation: the chips, the models, the data centres, and the computational capacity that will be needed when those products do arrive.
Citas Notables
The industry's mid-to-long-term demand far outpaces our supply capabilities.— Eddie Wu, Alibaba CEO
The truly groundbreaking products of the Machine Intelligence era have not yet arrived.— Eddie Wu, Alibaba CEO
Quieres la nota completa?
Lee el original en Reuters ↗
Technology