---
title: "Tencent Won the AI Race in a Sidebar"
date: 2026-11-19
category: AI in Action
site: NeuroAI
canonical: https://neuroai.site/a/na-app-tencent-workbuddy-1119
language: en
---

# Tencent Won the AI Race in a Sidebar

> In early November, WorkBuddy passed 10 million daily active users. On the same measure it logged more than 100 trillion tokens in a day.

In early November, WorkBuddy passed 10 million daily active users. On the same measure it logged more than 100 trillion tokens in a day.

10 million people, 100 trillion tokens. That is 1 million each.

## Three numbers in one row

The third comes from Martin Lau, president of Tencent Holdings (0700.HK), on the third-quarter earnings call: WorkBuddy's revenue over the past seven days annualizes above $500 million, about 3.334 billion yuan, more than twice Tencent Meeting's.

By daily token consumption per user, Tencent says WorkBuddy is the most used AI app in China. Its own measure, not a third-party ranking.

In the quarter, WorkBuddy's revenue was more than 5 times the second quarter's, MaaS more than doubled sequentially, and cloud rose 35%. The 5 times has no disclosed base.

## A million tokens a day

100 trillion over 10 million is 1 million.

No ratio turns that into words. One person typing cannot produce it in a day. So agents run it in the background. A chat box follows human hours and falls at night. An agent follows tasks and keeps running after the desk empties.

Measuring usage in tokens is measuring work, not headcount. The measure favors Tencent, and it suits AI apps: it swaps how many opened it for how much got finished. The second is harder to inflate.

It has an edge. High tokens can mean more work, or more talk.

Revenue per head: $500 million over 10 million daily users is $50 a year. The division is loose: the denominator is daily users, not payers. As an order of magnitude, ARPU is low. Tools are low early. Hold the people, price later.

## The entrance is in the contact list

On October 14, WeCom, Tencent's workplace app, and WorkBuddy shipped a joint build. Employees do not switch apps; the assistant sits in the contact sidebar. The sidebar matters more than any parameter count.

QuestMobile: in September 2026, WeCom's PC monthly users were 20.30 million, DingTalk 12.57 million, Feishu 1.94 million. That is 1.6 times DingTalk, and 5.79 million more than the two combined. The gap was opened by which window people open first at work.

PC matters. On a phone, an office app means clocking in, approvals, messages. Proposals, spreadsheets and contracts happen on a desktop.

Position sets frequency, frequency sets tokens, tokens set revenue. The model is one link, not the earliest.

## Grown out of code

WorkBuddy is led by Liu Yi, a Tencent Cloud vice president, who brought it out of CodeBuddy.

Version 0.01 came from Wang Shengjie, CodeBuddy's chief product manager, and colleagues, over two all-nighters on the weekend of January 17. Two nights to a 0.01 says the start was low and the direction narrow: serve people who write code. The rename dropped the Code.

Wang says the product is 100% in-house and used no OpenClaw source. That is his statement, with no third-party check. It is here because it arrives abruptly: a product denying use of an open-source project usually means the rumor is already circulating.

Capability widened from code to documents, sheets and research. The entrance did not move. An agent that writes code can write a weekly report. The reverse does not hold.

## 81.4 billion yuan for time

The quarter's most striking line is not AI revenue but capex: 81.433 billion yuan, up 527% year on year, 65.1 billion in cash. Nine months: 166.153 billion, up 179%; R&D 32.362 billion, up 41.8%; revenue 207.336 billion, up 8%; non-IFRS operating profit 68.073 billion, down 6%. Revenue up 8, profit down 6, capex up 527: buying tomorrow with today's profit.

81.433 billion is accrual, 65.1 billion cash; 16.3 billion is not yet paid. Nine months minus the quarter leaves 84.72 billion for the first half, 42.36 billion a quarter. One quarter at 1.9 times the prior average. The spending is a pedal, not a line.

Free cash flow was negative 69.1 billion. Excluding prepayments for compute, Tencent said, it was positive 16.9 billion. The 86 billion between them was not lost; it was paid early for compute. When cards are scarce, paying first buys a place in line. Negative 69.1 billion is a queue ticket.

Annualized revenue of 3.334 billion against capex of 81.433 billion is 4.1%. A year of AI application revenue does not cover a quarter of compute. Capex is also 1.2 times operating profit. Infrastructure is built before it is used, so this need not be a loss. But Tencent is selling seats, not AI margin. Sell the seat first.

Tencent's own date is 2028, when efficiency should improve through infrastructure reuse, cheaper hardware and better compute efficiency. Two of the three are not Tencent's to decide. What it controls is reuse: one set of halls and one scheduler across more businesses.

Pony Ma has called AI a matter of life or death. The line is quoted as resolve. Here it means something else: models can be caught, entrances cannot, because an entrance grows out of organizational relationships. Behind 20.30 million PC WeCom users sit contact lists, approval flows and customer groups another product cannot carry away.

So the moat is not WorkBuddy. It is the column of contacts to its left. That is why Tencent says "most used in China": no third party measures calls from an enterprise IM sidebar.

One question. The quarter does not disclose the model or its scores. For an entrance product, the model is a part, not the pitch. But parts have prices. If rivals' models are stronger a year out and the entrance is still Tencent's, the part can be swapped. If a browser or a system-level assistant routes around it, a stronger model will not help.

When the deciding factor moves from benchmark scores to where work begins, what pricing power is left for a company with a model and no entrance?

3.334 billion yuan annualized is rent on an entrance, not a trophy for a model.

*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-app-tencent-workbuddy-1119
Free to quote with attribution and a link to the original.
