Every organizational restructuring ultimately exists to serve the execution of a strategic goal. On July 30, 2026, ByteDance issued an internal memo announcing that the entire Feishu product team would be merged into Doubao, forming a new Doubao product team, with Feishu CEO Xin Xie now reporting to Doubao lead Zhao Qi. Feishu's marketing, sales, and customer-service teams were consolidated with Volcano Engine to form a unified ToB organization, the "Creativity Service Platform." A mature office product with annual revenue exceeding 3 billion yuan has been folded into the AI product system and fully integrated. The sales-side consolidation sends an equally clear signal: enterprise customers no longer need to engage Feishu and Volcano Engine separately — a single team now covers the full spectrum of needs spanning office collaboration, cloud computing, and AI applications. Around the same time, Alibaba merged three internally competing office-AI products into "Qianwen Office," deeply embedded within the DingTalk ecosystem; three weeks earlier, OpenAI folded its coding agent Codex into ChatGPT, unifying chat, office, and coding capabilities behind a single entry point; and 360 launched Nano Work, officially entering the office-AI race.
A noteworthy industry commonality is the shift in product naming. Anthropic launched Cowork, OpenAI rolled out ChatGPT Work, Kimi released Kimi Work, ByteDance has TRAE Work, 360 launched Nano Work, Tencent is building WorkBuddy, and Alibaba simply named its new product Qianwen Office. The entire industry's product positioning is collectively moving away from "Chat" and toward "Work." These dense organizational adjustments and product iterations all point to the same strategic judgment: the primary battlefield of AI is shifting from chat-and-entertainment to office productivity. And the core prize in this shift is not the short-term traffic gateway, but the user's long-term switching cost.
Part I. The Endgame of the Chat Model: Cost Structure Dooms the Free-Service Logic
To understand this strategic shift, we must return to the most basic business logic: AI's cost structure is fundamentally different from that of traditional internet products. In the internet era, the marginal cost of a product approached zero — serving one more user added almost no cost — so the playbook of "acquire users for free, then gradually monetize" worked perfectly: first build a massive user base, then commercialize through advertising, subscriptions, and value-added services. User count and time spent were the core assets of that age. But AI's operating logic is entirely different. Every conversation interaction and every deep task consumes real computing resources. The larger the user base and the deeper the usage, the higher the cost the company bears. This is not a choice of business strategy; it is a hard constraint imposed by the cost structure.
The industry events of July 2026 are a direct manifestation of this logic. Moonshot AI released its new-generation model Kimi K3, and within 48 hours of launch, user request volume hit the ceiling of its existing compute cluster, forcing the company to suspend new consumer subscriptions. Industry estimates show that a heavy user completing two deep conversations with million-character context windows per day incurs roughly 2,160 yuan per month in inference-side electricity and hardware depreciation costs, while Kimi's highest-tier membership fee is only 559 yuan per month — meaning a heavy user's monthly compute cost approaches four times the top membership fee. After launch, users of the 199-yuan plan exhausted their quota in under two days, and the 699-yuan plan's allowance was also quickly depleted. Also in July, DeepSeek introduced peak-and-off-peak pricing for its V4 API, doubling prices during peak hours; Doubao had already ended its fully-free model back in June, launching three subscription tiers. Since March 2026, Zhipu, Alibaba Cloud, Tencent Cloud, Baidu AI Cloud, MiniMax, and others have successively raised model prices, behind which lies a persistently widening gap between compute supply and demand. Data from the Ministry of Industry and Information Technology shows that by the end of June 2026, China's total intelligent-computing scale grew 177% year over year; meanwhile, data from the National Data Administration shows that domestic average daily token calls grew more than a thousandfold compared with early 2024, with demand growth far outpacing supply growth.
So-called "down-weighting" of consumer users by large-model companies — restricting features and compressing deep-usage quotas — is not a subjective choice by enterprises, but an inevitable consequence of the cost structure. When "the more you use, the more you lose" becomes an industry norm, no company can sustain a free consumer-traffic model over the long term. Doubao itself is a product of the typical internet traffic playbook, relying on a free model to rapidly build its user base, and ByteDance's proactive pivot of it toward ToB in July precisely signals that the old growth path has reached its end. Changes in revenue structure also confirm this shift: according to public data, Kimi's annualized revenue grew from 100 million to 300 million US dollars within three months, with API revenue accounting for over 70% of the total. The move from "grabbing user scale" to "prioritizing revenue efficiency" is a choice shared across the entire industry. As we noted earlier in "The Token Budget," the logic of consumer internet is "more is better," while the logic of enterprise AI is "precision targeting" — this is not only a company's AI-billing principle, but the survival rule of the entire large-model industry.
Part II. A Global Industry Consensus: ToB Productivity Scenarios Are the Core Path to Commercialization
If the cost structure explains why companies can no longer burn cash on free consumer models, then the global market has already validated the definitive direction of AI commercialization: ToB enterprise services are the true way out. OpenAI and Anthropic happen to be contrasting samples of the two routes. OpenAI continues the growth logic of the previous generation of internet giants — using free products to attract massive consumer user bases, then gradually exploring monetization paths. By early 2026, ChatGPT's weekly active users exceeded 900 million, with paid subscribers around 50 million, a paid conversion rate below 6%, and over 90% of users being free users who consume compute. To sustain this huge free user pool, OpenAI must spend 1.60 to 2.25 US dollars to earn every 1 dollar of revenue, and its gross margin has fallen from 40% to 33%. In early 2026, OpenAI even introduced ad placements in the free version of ChatGPT, attempting to use advertising revenue to subsidize compute costs — precisely the classic monetization path of the previous internet generation.
Anthropic took the exact opposite route — no traffic advertising, no hardware ambitions — with over 80% of its revenue coming from enterprise customers and API calls. Enterprise revenue and consumer subscriptions are completely different business species: after an enterprise contract is signed, the switching cost is extremely high; the deeper the usage, the higher the switching cost, so not only is the renewal rate stable, but the customer's purchase amount also grows year by year. Individual subscribers, by contrast, can cancel at any time, and user retention depends heavily on product novelty. From a transaction perspective, the former is a long-duration asset, while the latter is a short-duration asset. The difference between the two paths ultimately shows up in revenue data. Anthropic's annualized revenue grew from 1 billion US dollars in January 2025 to 30 billion US dollars in April 2026 — a 30-fold increase in 15 months — overtaking OpenAI's roughly 25 billion US dollars in annualized revenue over the same period. By April 2026, Anthropic had surpassed 1,000 enterprise customers with annual fees exceeding 1 million US dollars, doubling within two months; 8 of the Fortune 10 companies use Claude; enterprise spending data shows that Anthropic's share of enterprise AI spending rose from 10% in early 2025 to over 65% by February 2026, with its coding product Claude Code going from zero to 2.5 billion US dollars in annualized revenue in just 9 months.
The most convincing pivot comes precisely from OpenAI itself. In early July 2026, OpenAI folded Codex into the ChatGPT desktop app; after the merger, Codex's user count grew from 5 million to 10 million within days. Although consumer traffic remains its label, the product's center of gravity has clearly tilted toward productivity scenarios, giving the chat entry point real task-delivery capability. The revenue structure of the domestic market also confirms the same pattern: by July average-consumption metrics, ByteDance's large-model business had reached 4 billion US dollars in annualized revenue, exceeding the sum of all other domestic model companies' annualized revenue. The core supporting this scale is Volcano Engine's enterprise services and API business, not consumer chat products. Sustainable profit cannot grow out of a chat box; stable commercialization value is born only in office scenarios. This is not an accidental choice of any single company, but a definitive trend validated jointly by the global market.
Part III. The Battle for the Office Entry Point: The Core Moat Is the User's Switching Cost
Once the direction is clear, the next question follows: ToB application scenarios are so broad, so why do all companies, almost in unison, choose office as the entry point? Two decades of internet development appear on the surface to be competition over business models, but at their core they are contests for the entry point. Whoever controls the entry point controls the distribution of attention and data. The entry point of the PC era was the browser and the search box; the entry point of the mobile era was the super-app. Among all enterprise scenarios, office is the entry point of entry points: knowledge workers spend nearly half their waking hours in office scenarios — high-frequency, essential, with clear task goals — and every interaction deposits the most authentic work data, which is precisely the core fuel that large-model iteration most needs. More crucially, the office scenario is naturally tied to commercial value; every hour of labor saved can be converted into real cost-benefit, so users' willingness to pay requires no extra education.
There is another easily overlooked underlying logic: users already have the habit of opening office software every day, and already have the habit of paying for office software — this is the essential difference between the office scenario and other AI applications. Building a brand-new standalone AI app requires educating users to download, register, and form usage habits, with massive user attrition at every step. But AI capabilities embedded in existing office software face mature users who actively open the software every day and have already paid — no need to change user behavior, only to gradually penetrate along existing workflows, drastically lowering the adoption barrier. Industry data also confirms the scale advantage of the office scenario: as of August 2025, DingTalk served over 26 million enterprise organizations, with 79% of A-share listed companies using DingTalk and 2025 revenue exceeding 4 billion yuan; WeCom (Enterprise WeChat) connects over 14 million real enterprises and organizations; and among Feishu's new customers in the second quarter of 2026, over 90% simultaneously purchased Feishu's AI products. After Qianwen Office and DingTalk were bidirectionally embedded, with authorization they can directly read group chats, schedules, to-dos, documents, and knowledge bases, automatically push results back into workflows, and even directly generate interactive web pages, completing domain, database, and hosting deployment in one stop — greatly lowering the adoption barrier for agents.
Office software itself is undergoing a generational transition. The first generation of office software was locally installed tools, where all work had to be done by the user; the second generation was cloud-collaboration tools, supporting multi-person collaboration but still relying on humans for task execution; the new generation of AI-native office software no longer sells fixed-function tools, but intelligent capabilities that can autonomously generate tools and complete tasks. DingTalk's new CEO Yusen Chen once stated publicly that the minimum gap between an AI-native organization and a non-AI-native organization is more than 10-fold. When AI can directly call data from documents, spreadsheets, meetings, and group chats, users no longer need to open various tool applications separately. When user habits shift from "what tool should I use" to "what task should I have AI complete," the behavioral path becomes irreversible.
Only by seeing through the three layers of logic above can we truly understand the essence of the current office-AI melee. On the surface, it looks like a replay of the internet entry-point wars. Analysys data shows that in June 2026, the combined visits of 17 mainstream desktop AI office agents in China surpassed 60 million, while just three months earlier this figure had only just crossed 20 million — a threefold market growth in a single quarter. The track has gathered at least 15 products, with Tencent's WorkBuddy leading at 20.97 million monthly visits, followed by ByteDance's TRAE and Alibaba's QoderWork. Each player's entry path differs: spilling over from coding capability into general office, penetrating the operation entry point at the OS level, or embedding into existing office ecosystems and workflows. But the core of competition has fundamentally changed. The internet era fought for traffic, which comes and goes with unstable retention; this round fights for everything the user has deposited on the agent — work data, project context, thinking habits, collaboration records — the common feature of these assets being irreversible switching cost. Ecosystem synergy also reinforces the moat: WorkBuddy relies on the Hunyuan large model for underlying capability, integrating WeCom, Tencent Docs, and Tencent Meeting into a tool matrix callable by agents, so users need no repeated authorization when operating, and the smoother the experience, the higher the replacement cost.
Individual users' migration already involves clear path dependence; enterprise users' switching cost is even heavier. Once an organization has deposited its entire workflow, knowledge base, and approval system on a single platform, the cost of switching platforms is no longer a difference in procurement fees, but a complete rebuilding of the organization's work habits. Data also confirms this logic: Anthropic's enterprise customers with annual fees above 1 million US dollars doubled within two months; once an enterprise uses the product deeply, not only will it not easily switch, but it will continue to expand its procurement scale. This is precisely the ultimate bet of the strategic shift from Chat to Work: traffic will drain away, but switching cost will not. Whoever first occupies the user's office scenario can, through the deposition of data and habits, lock the user into the productivity scenario for the long term.
From Chat to Work: AI returning to its productivity essence
Finally, I want to say that this industry shift has an even deeper significance. Not all individual users need AI. AI has never been an entertainment tool, nor a better search box. Even for consumer users, the moment they are truly willing to pay is always to complete some specific task. Doubao grew its user base through free chat, yet did not formally launch a subscription system until June 2026 — which precisely shows that mere chatting cannot breed a stable payment motive; only real task delivery can sustain commercialization.
From this perspective, the shift from Chat to Work is the inevitable process of AI moving from the illusion of entertainment back to its productivity essence. Its core value has never been to give ordinary people one more chat-companion toy, but to improve production efficiency and amplify the ability to create value. And the endgame of this process is never about who replaces whom. Entry points will rotate, data will deposit, but what determines the long-term trajectory of the industry is never which platform wins the entry-point war, but the position in which human agency is placed. The more AI can take on execution-level work, the more firmly humans must stand at the core position of setting goals and making judgments — this is not concession, but division of labor. Humans control direction and guard key nodes in concrete work; AI breaks down steps, calls tools, and delivers results — this is the core logic of super-symbiosis. Entry points and data determine the outcome of short-term competition, but the collaborative symbiosis between humans and AI is the lasting underpinning of this business.
