On November 4, two signing books lay on a table at a ceremony. On one side was Hisense, on the other Alibaba. What they signed was a strategic cooperation agreement, and what the outside world can see is mainly a press release.
The release disclosed no amount and no shipment targets. Only four things can be confirmed: co-building the full stack, expanding overseas together, deepening technical cooperation, and upgrading marketing intelligence. It did not say when any of this would land.
Those four phrases are stiff. In plain language: where the AI inside a TV comes from, how to sell abroad, how the technology connects, and how to sell goods more precisely. One more thing was left unsaid: who supplies the brain, and who supplies the screen.
Around that day, the Wuzhen summit was under way, and several big companies were releasing their own news. The day's announcements crowded together. The excitement belonged to others; this agreement got one paragraph in a release, and not much information.
Four People at the Signing Table
Four people witnessed the signing: Alibaba Group CEO Wu Yongming; Alibaba Group Vice President and QwenWork CEO Chen Yusen; Hisense Group CEO Yu Zhitao; and Hisense Group Chief Capital Officer Xia Feng.
The two who signed on behalf of the companies were Hisense Group Chief Information Officer Zhang He and Alibaba Cloud Intelligence Group Vice President Gao Fei. One runs information systems, the other runs cloud. The choice of signatories already shows which layer the cooperation sits on.
This is not the first handshake. Hisense and Alibaba trace their cooperation back to 2010, across e-commerce, optical communications, content platforms, cloud, and AI. This time, scattered cooperation was gathered into one agreement.
A Five-Layer List
According to the release, the two sides will work across five layers: infrastructure, compute, model services, agents, and business and ecosystem. From the bottom of the data center all the way to functions a user can touch, one layer on top of another.
On the product side, the language becomes concrete: jointly build a new generation of smart TVs and smart-home products centered on "agents define functions" and "emotional companionship." The TV is simply the largest screen among them.
The phrase "agents define functions" deserves a second look. A TV used to have a menu of functions with AI stuffed in afterward. Now it is reversed: an agent that understands speech comes first, and functions are called by the agent.
Another line runs inside the company: tools such as Qoder and QwenWork will help Hisense change its own processes. This part is not advertised. It changes the organization, not the product, but it often takes effect earlier than anything on the TV.
The release is also written in a telling order. It starts with the five-layer architecture, then products, then overseas markets. The order is the priority: explain the foundation first, then talk about the living room, and the further it goes, the more it reads like a checklist.
Why TV Makers Don't Build Their Own Models
Everyone in home appliances knows how thin a TV's gross margin is. Panels, chips, logistics, channels, every link costs money. Add a team training a large model, and the math becomes hard to balance. This is basic manufacturing arithmetic.
A large model is not a one-time purchase either. Parameters iterate, data updates, and compute must run year-round. A version trained today may fall behind in six months, while the spending does not stop. There is no buy-it-once option.
The lifespan of a TV follows a different clock. A set bought today is often used for many years without a hardware change. It cannot keep pace with model releases, so the brain has to live in the cloud, where version numbers outnumber hardware models.
So the usual approach for a TV maker is to buy parts: panels from panel makers, chips from chip makers, one layer of system software in-house, and the remaining capabilities from outside. The model is just the newest part, and the most expensive one.
Hisense is not without its own technology. It has display chips, its own TV system, and a foundation in displays and optical communications. But those and training a general large model from scratch are undertakings of different magnitude.
This is not a question of ability but of division of labor. An assembler knows how to fit hundreds of parts into one shell and sell it, not how to make every part itself. That skill took decades to build.
The Model Is the New Panel
Treat the model as a panel and much of this falls into place. The panel sets the ceiling on picture quality; the model sets the ceiling on understanding. Both require long-term investment, and neither is what an assembler does best. Neither can do it for the other.
What Alibaba brings is the Qwen series of models, Alibaba Cloud's compute, and tools such as QwenWork that are already running. What it lacks is the screen in the living room, an entrance that gets opened many times a day.
Nazhen Technology, Hisense's optical-communications unit, is already one of Alibaba's top suppliers in its compute supply chain, providing high-speed interconnect for data centers. That line was connected earlier than the TV line, and is less visible.
One line is the living room, the other the data center. In the same agreement, the two lines run separately but need each other. One wants an entrance, one wants compute; what they trade is different. They may not see equally far, but they do need each other.
The Hardest Screen Is in the Living Room
Television interaction has long been one of the industry's hardest problems. The remote has many buttons, speech recognition struggles in a noisy living room, and every family member has a different accent. A single sentence often has to be repeated several times.
For years, TV voice assistants mostly stopped at understanding commands: change the channel, adjust the volume, search for a title. One step more complex, and they began answering the wrong question, so users gave up and went back to the remote.
"Emotional companionship" was written into the goals, which suggests the next step is not commands but conversation. That demands far more from a model than from hardware, and it is harder to demo on a stage.
But the living room is also the most valuable place. A family is there every evening, the screen is the largest, and the hours are the longest, even if much of the time people are looking at their phones. Whoever gets that screen to speak is closest to the user.
A Deal With No Amount
This agreement has no amount and no sales target. What can be checked is the framework, the people who signed, and the order of the five layers. It reads more like a roadmap than an order, and it has no delivery date.
For ordinary people, it may not change the TV at home right away. In the models you buy, which functions are connected in and which are built in will become easier to tell, and worth asking about.
The fastest way for a TV to get smart is not to grow a brain of its own, but to find a neighbor that already has one. The neighbor has moved in. What happens next is whether it will use the living room, or just stand at the door.
Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.