A 5–10 trillion parameter AI model sounds like the kind of thing people say when they want you to stop asking annoying questions about what the product actually does. Bigger is impressive. Bigger is also a convenient way to hide the messy parts: cost, control, and whether the “smartness” shows up in real life or just in a demo.
Still, Alibaba’s announcement matters. Based on what’s been shared publicly, the company says it plans to build an AI model in the 5 trillion to 10 trillion parameter range, far larger than its current flagship model, Qwen 3.8 Max, which it says has 2.4 trillion parameters. At the same event, its CEO Eddie Wu also introduced a new AI chip called Zhenwu V900, described as the most powerful in China. This is being framed as a push on both brains (the model) and muscles (the hardware).
On paper, that’s the right direction. If you’re serious about AI, you can’t only rent other people’s compute forever. You want your own stack. You want leverage. You want to control your costs and your pace.
But I don’t buy the idea that “more parameters” automatically means “better for humans.” It often means “more expensive to run” and “harder to understand,” which then means “harder to put guardrails on.” And when guardrails are weak, the people who pay the price are not the people giving the keynote.
Imagine you run a mid-sized business using Alibaba Cloud tools. You don’t care about a trillion anything. You care that customer support replies stop hallucinating, that invoices don’t get messed up, that your sales team doesn’t send fake information to clients. If Alibaba’s bigger model actually reduces those failures, great. If it mostly raises the price and adds new kinds of weird errors, you’re stuck choosing between “stay behind” and “take a risk.”
Now imagine you’re a developer inside a big company. A stronger model can make you faster. It can also quietly change what your manager expects. Suddenly a one-week task becomes a one-day task because “the AI can help.” The model becomes less of a tool and more of a pressure machine. That’s not Alibaba’s fault alone, but it is a consequence of pushing raw capability without being honest about how it changes work.
There’s also the chip announcement. I’m glad they’re investing in their own infrastructure. When one country or one set of firms controls the key hardware, everyone else lives on their schedule. A strong domestic chip can mean more competition and more options. That’s a real win.
But it cuts both ways. If a company says it has the most powerful chip “in China,” that’s not just a product brag. It’s a signal about independence and capacity. It suggests a future where AI isn’t just a software race; it’s an industrial race. In that world, the winners are the firms that can spend huge amounts for years, and the losers are everyone building on top who has to accept the terms.
And here’s the part that makes me uneasy: parameter counts are becoming a scoreboard for status, not a measure of usefulness. If the internal goal is “be bigger than the last model,” you’ll optimize for size because size is easy to sell. You’ll do the thing that looks like progress even when the real progress is boring: fewer mistakes, better memory, clearer refusal behavior, easier tools for people who aren’t engineers.
If you want an argument against me, it’s simple: scale works. Bigger models often do show better results. They handle more tasks. They generalize better. And companies like Alibaba don’t announce these numbers for fun; they announce them because they think they can get real performance out of it. That’s fair. It may also be necessary if they want to stay competitive.
But even if the model is amazing, there’s a second issue: who gets it, and under what rules. A model that costs a fortune to train and run doesn’t usually become cheap and open overnight. It becomes a premium service. That’s good for the platform owner. For everyone else, it can mean deeper dependence. If your business gets built around this model and the price changes, your margins change. If usage policies change, your product changes. If the model suddenly refuses certain requests, your workflow breaks.
And what about safety? I’m not saying “big model equals dangerous.” I’m saying “big model plus competitive pressure equals rushed decisions.” When the goal is to ship capability, the quiet work of limiting harm tends to lag behind. That’s when you get models that are brilliant and sloppy at the same time.
So yes, Alibaba aiming for a 5–10 trillion parameter model and launching its own chip is a serious move. It could lead to better tools, lower costs over time, and more independence in the AI supply chain. It could also deepen a world where AI power concentrates into a few stacks, and everyone else just plugs in and hopes the terms don’t change.
If you were running Alibaba, would you prioritize building the biggest model you can, or the most reliable model people can actually trust day to day?