Tencent Cloud Bets on Agents: WorkBuddy Completes 9 Billion Task Requests in 7 Months Since Launch

The hot spots for agents have cycled through several rounds within a year, and office scenarios have become the new battleground. For cloud vendors, the revenue structure is also changing: in the past they sold training compute and API (interface) calls, but now agents have become a new entry point for token consumption.

At the Tencent Global Digital Ecosystem Summit held from October 29 to 30, Tencent disclosed that WorkBuddy completed 9 billion user task requests in just over seven months since launch, and its enterprise edition customers cover half of the domestic Fortune 500 companies. Such application tasks consume far more tokens than conversational products like Yuanbao. Yang Chen, Vice President of Tencent Cloud, said that over the past three or four months, 2,000 partners working on WorkBuddy have come in, 60% of them new.

At the same time, the data centers that support computing power are still being expanded. To what extent the new customers and token revenue brought by agents can cover this investment is the question that needs to be answered next.

The "hot favorites" among agents have shifted from OpenClaw at the beginning of the year to personal agent Muse and office agent WorkBuddy, among others. At some domestic airports, the AI cloud advertisements that blanketed the scene last year have this year become a battleground for various agents.

The spread of agents brings more computing power demand, and as computing power support, the data center construction wave continues. The reporter learned at the summit that over the past year Tencent has added or expanded data centers in Frankfurt, Germany; Malaysia; Mexico; Osaka, Japan; Riyadh, Saudi Arabia; and Thailand, and in the next year it plans to expand its layout in Asia, the Middle East, Europe and the Americas.

How such AI infrastructure costs will be recovered remains a concern in the industry. Market research firm Gartner recently estimated in a report that by 2030, owners of AI data centers will need to generate $8.5 trillion in revenue to cover capital expenditure, financing costs and achieve a 12% return on investment.

On this point, Dowson Tong, Senior Executive Vice President of Tencent Group and CEO of the Cloud and Smart Industries Group, told the reporter that from a macro perspective, it is difficult to apply a single number to measure (investment and return), as costs vary greatly across application scenarios and ROI (return on investment) performance differs, but cost as the denominator of ROI can be measured. "The cost spent on hardware today is very high. If the same task consumes the same amount of tokens and the cost can be brought down, it will be beneficial to AI development and user adoption," he said, noting that hardware costs include chips, memory and so on.

Tencent Cloud has currently deployed more than 2 million servers globally. Dowson Tong said Tencent will next consider both self-building and leasing data centers. Against the backdrop of scarce global computing power, if the corresponding components can be purchased, Tencent is very willing to build its own.

As for whether AI can work in productivity scenarios, Dowson Tong responded to voices saying "AI doesn't work." He said that if very expensive models and inefficient computing power are used, costs will soar, and the numerator corresponding to ROI, namely the return, will need to be very large to prove that productivity works. But domestic models have been optimized for cost-effectiveness, keeping costs down, thereby changing the ROI result. Dowson Tong believes that AI has already worked in office scenarios, and many enterprises are willing to invest in AI.

From selling servers to selling results

Agents have brought a new way of consuming computing power. Previously, cloud vendors' AI revenue included server resources rented for AI training and tokens generated by large model API calls. Agents, while serving as an entry point for token consumption, have also become a new variable in the cloud business.

From the perspective of tokens, Dowson Tong told media including Yicai that among Tencent's main token consumption scenarios, WorkBuddy is a very major one; it takes relatively long to complete tasks and consumes far more tokens than Yuanbao, which engages in rapid back-and-forth dialogue. In the second half of AI, the industry is paying more attention to product value than model benchmark scores, and token call volume is an indicator for evaluating AI.

WorkBuddy has also brought new business. Yang Chen told the reporter that Tencent Cloud began letting channel partners sell WorkBuddy in May, and over the past three or four months, 2,000 partners working on WorkBuddy have come in, 60% of them new partners. Previously, Tencent's public cloud revenue mainly came from gaming, e-commerce, social networking and other fields, but now the volume of customers using WorkBuddy in manufacturing, retail, healthcare and education is far greater than in other fields.

"In the past we may have been selling server resources, but now what customers need is solutions," Wu Qisheng, Vice President of Tencent Cloud and head of ecosystem management, told the reporter. Previously, when Tencent sold public cloud, it proactively sought out customers and played more of a supplier role; now many potential customers proactively inquire about WorkBuddy. In the early stage of WorkBuddy's adoption, the fastest-moving field was manufacturing, a type of enterprise that previously had little demand for public cloud. Repetitive work such as order management that manufacturing enterprises used to spend dozens of hours on can now be completed in two or three hours.

"Although (selling WorkBuddy to one customer) may bring in only tens of thousands of yuan, subsequent internet and AI transformation may generate more cooperation," Wu Qisheng said, noting that WorkBuddy can serve as a "hook" product. The output value from such traditional industries may not yet be large, but there is potential for extension later.

In addition to manufacturing customers, some smart hardware such as earphones and keyboard microphones has also integrated with WorkBuddy. Besides token consumption, Tencent charges such hardware vendors license fees. Yang Chen told the reporter that this year is the year with the biggest changes in the cloud market in the past three or four years. In 2021 the industry worked on computing power, and over the past three years it worked on models; now agents have brought the industry a third AI growth curve, and this new curve is steep, with related business likely to exceed the growth in computing power and models in the future.

Yang Chen told the reporter that public cloud sales used to follow IT sales logic, "involution" on price and competing over existing stock, whereas WorkBuddy sales follow a value logic, with customers feeling they pay for results. But he also observed that while demand from agent customers is strong, many people still treat it as a tool, and the next step is to push for business process restructuring.

At present, Tencent has no hard numerical targets for WorkBuddy's commercialization. Dowson Tong said Tencent's investment in the product is not unlimited, but at the current stage it remains an important strategic product.