An AI assistant becomes a different thing when people start using it to finish work rather than talk about it. The signal is not a launch video but a usage number that keeps climbing after the novelty fades. Tencent's third-quarter report offers a look at both halves of that story: what users do, and what it does for the business.
Released alongside the quarter's financial results, the update said AI demand drove cloud-service revenue up 35% year on year. It also said the company's agent product, WorkBuddy, had passed 10 million daily active users and was consuming 100 trillion tokens a day — the highest daily large-language-model token consumption of any Chinese AI product, according to Tencent itself.
Key takeaways
- Cloud-service revenue rose 35% year on year in the third quarter, with the company attributing the gain to AI demand.
- Tencent says its agent (智能体) WorkBuddy passed 10 million daily active users and processes 100 trillion tokens per day. Both figures are self-reported.
- Days before the results, the company shipped Hy4 flash, a 224 billion parameter mixture-of-experts model with 12 billion active parameters, multimodal input and a one-million-token context window.
- Pricing was set at RMB 0.5 per million input tokens (≈ US$0.07 / HK$0.55) and RMB 2 per million output tokens (≈ US$0.28 / HK$2.2), with cache hits at RMB 0.02 (≈ US$0.003 / HK$0.02).
- The strategic point is the loop: cheap models make heavy agent usage affordable, and heavy usage pulls cloud capacity along with it.
The two numbers that matter
A 35% rise in cloud revenue is the kind of figure an investor can check. A daily-active-user count and a token count are the kind a company chooses to release. They are not the same quality of evidence, and it is worth keeping them separate.
The cloud number says that AI demand is showing up in invoices. Infrastructure spending by Chinese technology groups has been widely reported this year, and at some point that spending has to turn into someone else's revenue. Tencent's 35% suggests it is, at least in the segment it chose to highlight.
The usage numbers are more interesting as a description of behaviour. Ten million daily active users is a real audience by any standard. A hundred trillion tokens a day is harder to picture, but the important part is the ratio: it implies that the average active user is asking the system to do a great deal, not just say hello. Agents (智能体) burn tokens on planning, tool calls and verification in the background. High token consumption per user is therefore a rough proxy for whether people trust the product with multi-step work.
The company's claim that this is the highest daily LLM token consumption of any Chinese AI product should be treated as a competitor comparison it has no reason to verify independently. The absolute numbers are plausible; the ranking is marketing.
What a cheap model has to do with it
Released on November 16, just before the results, Hy4 flash has a specification familiar by now: 224 billion total parameters, 12 billion activated per token, multimodal input, a one-million-token context. What matters for this story is the price. At RMB 0.5 per million input tokens (≈ US$0.07 / HK$0.55) and RMB 2 per million output tokens (≈ US$0.28 / HK$2.2), running an agent that makes many tool calls stops being a cost concern for ordinary tasks.
That is the mechanism connecting the two halves of the announcement. An agent that plans a trip, summarises documents and checks its own work might consume tens of thousands of tokens in a single session. At flagship prices, that session is a luxury. At flash prices, it is close to free, which is what allows a product to reach ten million daily users without each one becoming an expensive problem.
Tencent has said it plans to release weights and list the model on a public model hub, though weights were not available at launch. If that happens, the same economics become available to developers outside its own products — a different kind of growth story from the one told in a quarterly report.
Why the loop matters
The pattern here is becoming standard across the industry: train a smaller, cheaper model; make agent products affordable; watch usage climb; sell the capacity that usage requires. Each step reinforces the next, which is why cloud revenue and product metrics now appear in the same sentence.
For readers, the practical effect is that the useful AI products are increasingly the ones attached to a cloud bill someone else has already negotiated. A free or cheap assistant is rarely free in the economic sense; it is subsidised because the usage feeds a larger business. That is not a criticism — it is how a lot of software is paid for — but it is worth knowing when you decide which tool to build your workflow around.
There is also a concentration risk in the pattern. If a product's economics depend on very low token prices, a change in pricing or capacity policy can alter what it can afford to do. The 50% API discount reported alongside the launch was time-limited, which is a small example of how quickly the arithmetic can change.
Honest limitations
This article relies on Tencent's third-quarter disclosures as reported by Global Times and on a sourced model-release tracker for the Hy4 flash specifications and pricing. I have not read the full earnings release, and the 35% cloud figure refers to the segment as the company defined it in that report — not necessarily total cloud revenue under every accounting lens.
The 10 million daily users and 100 trillion daily tokens are company claims with no published methodology. Daily active users can be counted in several ways, and a token count says nothing by itself about how many sessions succeeded or how many users returned. "Highest among Chinese AI products" is Tencent's comparison, not an audited ranking.
The RMB-to-US and RMB-to-HK conversions are approximate and will move with exchange rates; they exclude platform discounts, taxes and regional price differences. The model's benchmark performance and the promised weight release were vendor-reported and, at the time of writing, the weights were not yet available. I have not tested Hy4 flash on my own tasks.
What readers can do now
- If you use an agent product, notice how much it does between your prompts. That hidden work is what the token number measures, and it is the real source of both the value and the cost.
- If you are choosing a tool, ask what happens when the introductory pricing ends. A workflow that only works at a discount is a workflow with a deadline.
- If you follow the industry, watch usage before revenue. A product people use daily usually finds a way to pay for itself; one that only launches rarely does.
The most telling number in this report may not be the revenue growth. It is the gap between ten million users and a hundred trillion tokens — the distance between having an audience and having it do real work. That gap is where the next battle will be fought.
