Alibaba is building the whole thing.
Not just an AI model. Not just a chatbot you can talk to. The chips underneath, the cloud that runs them, the models on top, and the agent software that uses all of it. Top to bottom, one company. In the industry this gets called a “full-stack” strategy, and Alibaba has now put a number on the ambition: $100 billion in combined cloud and AI revenue by 2031.
If you’re not technical, that phrase “full-stack AI” can sound like jargon designed to keep you out of the conversation. It isn’t complicated. Let me walk you through it the way I’d explain it to a friend over coffee.
What a stack actually means
Think about a restaurant. Most restaurants buy their vegetables from a supplier, rent their building from a landlord, and hire a delivery app to bring food to your door. They control the cooking, and not much else. If the supplier raises prices or the delivery app changes its rules, the restaurant just has to deal with it.
Now picture a restaurant that owns the farm, owns the building, and runs its own delivery fleet. Slower to set up. Far more expensive upfront. But nobody upstream can squeeze them.
That’s the difference between most AI companies and what Alibaba is attempting. The layers in an AI stack look roughly like this:
- Chips — the physical hardware that does the math. Alibaba is making its own, including proprietary CPUs and high-performance AI chips. The company says its new AI chip triples the performance of the one before it.
- Cloud infrastructure — the data centers and networking that tie thousands of chips together. Alibaba rebuilt this layer and is scaling infrastructure it already had at large size.
- Models — the AI brains themselves. This is the Qwen family, which Alibaba has been shipping openly for a while.
- Agentic workloads — software that doesn’t just answer you but goes off and does multi-step tasks. Qwen 3.7-Max is built specifically for this.
Why the agent part matters to you
That last layer is the one I care about most on this site, and it’s the tell.
A chatbot answers a question and stops. An agent keeps going. It reads a document, calls a tool, checks the result, tries again if something failed, then reports back. Each of those steps costs computing power, and agents chain together a lot of steps. A single agent task can burn through many times the resources of a single chat message.
So when a company designs a flagship model specifically for agentic workloads and at the same time designs its own chips, those two decisions are talking to each other. You build custom silicon when you know exactly what kind of work you’re going to run on it. Alibaba appears to be betting that the work will be agents, at volume, for years.
That’s a real signal. Chip design takes years and enormous capital. Nobody does it on a hunch about next quarter.
The open-model angle
There’s a twist that makes this different from the usual American tech playbook. Alibaba built its reputation in AI partly by releasing open models — Qwen weights that developers could download and run themselves. That won goodwill and mindshare with builders who couldn’t or wouldn’t pay for closed systems.
The new strategy tries to convert that lead into something more durable. Free models bring developers in the door. The stack underneath is where the revenue lives. Alibaba’s own framing is about turning its open-model position into a full-stack business, and TIME named the company to its 2026 list of most influential companies as that shift took shape.
It’s a smart trade if it works. Developers get genuinely useful free tools. Alibaba gets a generation of engineers who already know its models and find it natural to run them on Alibaba’s cloud.
What I’d watch, without the hype
A few honest caveats, because I’d rather you be well-calibrated than excited.
Announcing a stack and operating one at scale are different achievements. Custom chips have to actually perform in production, not just in a launch slide. Models have to stay competitive as rivals ship — and Alibaba has kept releasing, with Qwen 3.8 Max positioned against Anthropic’s strongest work. A 2031 revenue target is five years of execution away.
What I think is already clear is the direction. The people building AI infrastructure at the largest scale are designing for agents, not chat. They’re spending billions on that assumption. If you’ve been wondering whether AI agents are a passing bit of enthusiasm or something the industry is genuinely restructuring around, watch where the chip money goes.
Right now it’s going toward agents. Alibaba just made that unusually explicit.
🕒 Published: