Daily active users remain high, so why are AI giants getting anxious?
AI giants are shifting their operational focus from daily and monthly active users to token consumption, capital expenditure, and ARR, because user growth continually raises inference and computing costs. Drawing on data from Tencent, Alibaba, ByteDance, and DeepSeek, the article notes that data center delivery timelines and chip supply constraints are limiting expansion, while short-term annualized revenue figures do not necessarily reflect actual cash receipts. Intelligent office applications have become a priority because they are closer to paid use cases, but the traditional model of “acquire users first, monetize later” is coming under growing cost pressure.
Why read this
AI score 68
Examining capital expenditure, token consumption, and annualized revenue, the article explains why high daily active user counts do little to ease the cost and monetization pressures facing AI products.
Topics
Body · AI translation
Daily Active Users Aren’t Low, So Why Are the AI Giants Panicking? Huxiu, November 20, 2026, 14:43:19, from Jilin
In mid-November, free DeepSeek users discovered a new rule: if a conversation had been inactive for more than seven days, they could no longer continue it. The limit for a single input was reduced to about 40,000 Chinese characters, and once the daily allowance was exhausted, a pop-up said that the server was busy and the usage limit had been reached. #DeepSeek Usage Limit Reached# subsequently became a trending topic. Some people returned to meeting minutes they had previously asked it to organize, intending to make a few edits, only to find that the chat had already been closed. Users lined up to demand that the limit be removed, and some even volunteered, “If you launch a paid membership, I’ll buy it.” The official explanation was restrained: “constrained by computing capacity.”
That same week, the ledgers were laid out in exceptional detail: Baidu released its third-quarter results on November 16, Tencent on November 18, ByteDance unveiled Doubao’s workplace performance report on the evening of November 18, and Alibaba delivered the finale on the evening of November 19. In previous years, the competition at this time was over monthly and daily active users. This year, capital expenditure, Tokens, and ARR dominated the headlines.
All of them posted strong results, but none was celebrating. The tech giants were all anxious. Not because growth was too slow, but because growth itself was turning into a bill.
It was not like this ten years ago. An App’s greatest fear was that you would go seven days without opening it. Daily check-ins, red notification dots, push alerts, and re-engagement text messages were all designed to pull you back. Now the product speaks first: if you do not return for seven days, I will not keep your place.
What has changed is not etiquette, but the economics. Almost everything unveiled that week was good news. In the internet era, every set of figures would have warranted popping champagne. Yet all these companies were answering the same question: when will the money come back?
Daily active users are still there—they just no longer count as a victory
In the first half of this year, the industry was still diligently counting heads. According to QuestMobile’s September report, Doubao had 419 million monthly active users and DeepSeek had 123 million, while AI-native Apps collectively had 546 million monthly active users, up 82.1% year on year. The rankings are still published, but this figure is no longer the focus of the publicity.
At Tencent’s earnings call on the evening of November 18, Martin Lau said something rarely heard in the company’s official messaging: Yuanbao had not met initial expectations, the focus was shifting from expanding the user base to improving capabilities, and the company had substantially cut its investment in Yuanbao. When WorkBuddy was introduced at the same event, a different vocabulary was used: more than 10 million daily active users, over 100 trillion Tokens consumed per day on average, and, measured by Token consumption, already the most-used AI product in China.
In the earnings report, the superlative title did not go to the product with the most users, but to the one with the greatest consumption.
In a business that charges by usage, user numbers are merely the denominator; Tokens are closer to revenue. In June, Volcano Engine disclosed that daily average Token calls to the Doubao large model had surpassed 180 trillion, while its business focus was shifting toward programming, agents, and production scenarios. In July, when computing capacity came under pressure, Kimi tightened some consumer-facing services and prioritized its limited capacity for paying users.
For the past two decades, every playbook in China’s internet industry was built on a rule that scarcely needed proving: fixed costs came upfront, while marginal costs approached zero. Once the servers were installed, bandwidth purchased, and App developed, the cost allocated to each person kept shrinking from the first user to the hundred-millionth. “The more users, the cheaper it gets” therefore became common sense. Daily and monthly active users became the hardest currency in valuations, and attracting users for free before gradually monetizing them became the standard move.
In the past, each additional user brought another share of attention that could be sold. Now each additional user brings another inference bill that must be paid. So companies no longer have to explain how many people they have, but how much electricity it takes to support them. The change in language is not modesty; it is because this line item can no longer be contained.
The data centers are not finished, but the bills have already arrived
Expansion in the internet era meant adding bandwidth: when user numbers rose, a company could rent a few more servers and bring them online that same day. Expansion in the AI era means building power plants: pay first, secure chips first, acquire land and connect electricity first, while the machines may not start running for another two or three years. Yet the bills begin accruing the day the order is placed. Money can be approved in a day; a data center cannot be built in a day. Users may arrive tomorrow, but electricity takes years to secure. At full load, a 1GW intelligent computing center consumes about 876 million kWh of electricity annually, equivalent to the total yearly power consumption of a medium-sized city.
Tencent’s third-quarter capital expenditure reached RMB 81.433 billion, up 527% year on year. That single quarter exceeded the RMB 76.8 billion spent throughout 2024, with the money mainly directed toward data centers and advance payments for computing capacity. Free cash flow for the quarter was negative RMB 69.1 billion, swallowed up by those prepayments. Depending on whether new AI products are excluded, the operating margin was 40% or 33%. Behind the figures are server racks, chips, and data centers that have yet to be powered on. Management did not evade the issue on the earnings call: the short-term computing capacity available from high-end GPUs has declined, the next generation of Hunyuan may take longer, and the shortfall can only be made up through domestic supply chains. Even the pace of new releases is being dictated by data-center capacity. This is perhaps the most concrete manifestation of that “anxiety.”
Tencent is not alone in facing these construction timelines. Alibaba delivered its results immediately afterward: AI cloud and computing-capacity revenue reached RMB 60.973 billion, up 53% year on year; quarterly capital expenditure was RMB 50.477 billion, up 60%; and free cash flow recorded a net outflow of RMB 34.472 billion. On the conference call, Eddie Wu said bluntly that computing capacity had been particularly scarce in the Chinese market over the past two years. In February 2025, the company announced that it would invest RMB 380 billion in cloud and AI infrastructure over three years, while shortening data-center delivery times to 100 days. ByteDance raised its 2026 AI infrastructure budget to RMB 200 billion, including about RMB 85 billion for chips and RMB 90 billion for construction of intelligent computing centers in places such as Inner Mongolia and Xinjiang. This money becomes fixed assets and enters the income statements of the coming years through annual depreciation; the expense does not simply end once the money has been spent.
The clearest example is DeepSeek. From the outset, it did not rely on charging consumers: its app and website remain free for individuals, and the API is its only direct paid offering. According to an October report by Science and Technology Daily, it is still planning a gigawatt-scale data center in Ulanqab that will deploy at least 160,000 domestically produced AI chips. The entire cluster is not expected to begin coming online gradually until late 2027 or early 2028. Even a company with the least need to court users has been forced to build its own data center.
One coincidence that does not seem entirely coincidental: Tencent expects its operating pressures to ease in 2028, on the grounds that equipment procurement prices will have fallen significantly by then and the same infrastructure will be able to process more tokens. DeepSeek is also targeting around 2028 to power up its data center. Alibaba has laid out the same timeline for domestic computing capacity: Eddie Wu said that the supply chain for T-Head’s domestic chips would not begin scaling up substantially until after 2028, when the vast majority of model training would rely entirely on domestically produced chips. The three companies did not coordinate; they simply ran into the same physical construction timeline. CCTV Finance reported in early September that orders at an intelligent computing park in Ulanqab were booked out for three years. Demand for domestically produced AI chips is expected to reach about 4 million units in 2026, while actual deliveries will total roughly 3 million. Money can speed up decisions, but it cannot buy inventory that does not exist.
At an internal meeting held on November 11 to mark Tencent’s anniversary, management required every core business team to start afresh for the AI era and explicitly called for breaking down organizational silos. Pony Ma put it in unusually stark terms, urging teams to change tactics and “get aboard and join the fight.” Those four words were repeatedly quoted in notes made public on November 16. An organization does not use language this forceful because it is performing poorly. A product can be updated three times a week; a data center cannot. A service can gain 100 million users overnight; electricity supply cannot. The playbook the internet industry knows best—acquire users first, monetize later, and move faster than competitors—is failing for the first time when confronted with construction timelines measured in years.
Revenue from intelligent office products in financial reports is all being presented on an “annualized” basis
There is another question even more worth asking: Why has AI-powered office software become so popular, and why are all the tech giants crowding into the office market? Because it is currently the consumer use case that most resembles a real business.
Look at the tech giants’ latest financial reports and you will notice an interesting phenomenon: revenue from their intelligent office products is invariably described in annualized terms.
On the evening of November 18, Tencent disclosed that WorkBuddy’s revenue over the previous seven days exceeded US$500 million on an annualized basis, that third-quarter revenue was more than five times that of the second quarter, and that 60% of China’s 100 largest companies were already customers. That same evening, ByteDance announced that Doubao Work had surpassed 8 million daily active users three months after launch and that annual recurring revenue (ARR) from enterprise customers had exceeded RMB 2 billion. What is ARR? One of its core features is that it is calculated on an annualized basis. Alibaba’s AI-related products generated RMB 17.918 billion in quarterly revenue, marking triple-digit growth for 13 consecutive quarters, and more than US$10 billion on an annualized basis; Eddie Wu set a fiscal-year-end ARR target of RMB 30 billion for MaaS. Daily active users and standalone revenue for individual products still were not reported separately.
China Business News noted in an article that one month is the shortest reasonable annualization period, while a quarter is more stable. Annualized from seven days, that US$500 million corresponds to roughly RMB 64.55 million in revenue over the seven-day period. ARR is not money already in the bank either; it is a projection of what a full year would produce at the current pace. Using it to rank companies can easily be misleading.
Why does everyone favor annualized figures? Because annualization can magnify the pace seen in a brief window into a year-long story. But that does not work.
Annualization assumes that over the next 12 months prices will remain unchanged, models will not be replaced by newer generations, customers will not leave, computing capacity will not become scarce, and competitors will not upend the market. Yet in the AI industry, all five of those things could happen within a single year. Multiplying seven days of revenue by 52 does not produce actual revenue; it produces an assumption about “what would happen if nothing changed.” The longer the period, the more things can change, and the more an annualized figure resembles an empty promise.
The problems annualization cannot solve ultimately come back to costs.
Imagine this: 419 million people use Doubao every month, yet almost none of them pay. According to LatePost, it generates less than RMB 1 million in revenue per day. But every question those 400 million people ask costs money to process in the background. Doubao is a closed-source model, so outsiders cannot calculate its costs precisely; all they know is that the more users it has, the larger the bill becomes.
Whether that money can be recouped depends not on daily active users, but on whether each invocation performs valuable work for someone. It is not that the tech giants have failed to understand this. They understand it, but they cannot stop: the ship has already set sail, the fuel gauge is moving, and the shore is still nowhere in sight.
The internet industry’s old playbook of “acquire users first, monetize later” has stopped working here.
“Special notice: The above content (including any videos, images, or audio) was uploaded and published by a user of Dafeng Hao, a self-media platform owned by Phoenix New Media. This platform merely provides information storage services.
Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services.”
AI Trends