▲ All images in this article are from the internet. Please contact us for removal if any infringement occurs.

Is 'traffic-only' thinking still necessary in the AI era?

This article was first published in Shadow Memo. Written by Mo Yingsheng.

During the 2026 Spring Festival, China’s internet giants staged a dramatic 'money-throwing' spectacle.

Alibaba’s Qianwen invested 3 billion RMB, Tencent’s Yuanbao 1 billion RMB, and Baidu’s Wenxin 500 million RMB in subsidies. Together, these tech giants poured over 4.5 billion RMB into a high-profile battle to attract new users to their AI platforms.

This was no simple 'shake-to-win-red-envelopes' gimmick. Users could only claim red envelopes after completing interactive tasks such as AI dialogues, generating AI content, and experiencing AI effects.

Tech giants like Tencent, Baidu, and Alibaba invested billions in cash, aiming straight for AI entry points. Alibaba’s Qianwen deeply integrated AI capabilities with its 'Gather Five Blessings' campaign and local life services. Tencent’s Yuanbao leveraged WeChat’s over 1.4 billion monthly active users to rapidly accumulate users through social sharing. Baidu’s Wenxin Assistant incorporated red envelopes into nearly 200 AI effects.

Although ByteDance did not directly launch large-scale cash subsidies, it secured exclusive AI cloud partnership status for CCTV’s Spring Festival Gala through its Volcano Engine, leveraging the gala’s massive traffic to push Doubao into lower-tier markets.

This scene harkens back to 2015, when WeChat invested 500 million RMB and used the 'shake-to-win-red-envelopes' feature during the Spring Festival Gala to bind hundreds of millions of users’ payment cards in just a few days, shattering Alipay’s monopoly. Back then, 'throwing money for traffic' was the golden rule of the internet.

Yet, just four months later, the winds shifted abruptly.

On June 24, 2026, ByteDance’s AI assistant Doubao officially launched a professional paid subscription with three tiered pricing plans: Standard at 68 RMB/month, Enhanced at 200 RMB/month, and Premium at 500 RMB/month.

Almost simultaneously, Baidu announced the merger of multiple disperse websites—including Wenxin Yiyan Web, Wenxin, and Wenxin Assistant—into a unified Baidu Wenxin platform. Its pricing strategy remained unchanged: free.

One charges, the other doesn’t. While their actions seem contradictory, they point in the same direction. The AI competition is shifting from competing on the number of entry points and traffic volume to competing on who can actually get things done.

After burning 4.5 billion RMB, the tech giants suddenly stopped chasing traffic. Behind this shift lies a profound change in underlying logic.

The Golden Age and 'Fatal Curse' of Traffic Thinking

To understand today’s transformation, we must first understand why internet giants crazy chased traffic over the past two decades.

In the mobile internet era, a cardinal rule existed: more entry points, more traffic, and stronger bargaining power. For over a decade, nearly all internet companies expanded based on this assumption: flood the market with entry points, capture user time, and build moats through scale.

This logic was perfectly validated during the 2015 red envelope war. WeChat Pay leveraged red envelopes to match Alipay’s nearly decade-long accumulation of bound payment cards in just a few days.

A single product feature rewrote the entire mobile payment landscape. From then on, 'burning money for traffic' became standard practice for internet giants.

By the 2026 Spring Festival, this playbook was transplanted into the AI arena. The giants weren’t betting on a single wave of downloads but on users’ default choices for years to come.

The industry widely expects that general-purpose AI will become the core carrier of the next-generation internet. Each user may only need a few AI assistants to meet all online needs. Whoever seizes the super entry points of the AI era will control digital traffic distribution for the next decade.

But traffic thinking faces a fatal challenge in the AI era.

The cost structure of AI products is entirely different from that of internet products. Internet products have near-zero marginal costs—each additional user requires little more than a few extra lines of server code. But every AI query consumes real computing power. More users sometimes mean heavier bills.

The data is stark. By June 2026, Doubao’s large model averaged 180 trillion daily Token calls, a 1,500-fold increase since its launch and a 10-fold increase in the past year.

Take Doubao 2.1 Pro as an example: output is priced at 30 RMB per million Tokens. At 180 trillion daily Tokens, Doubao’s daily output costs alone reach approximately 5.4 billion RMB.

Earlier reports suggested ByteDance’s AI infrastructure spending in 2026 could hit 200 billion RMB.

This is a brutal arithmetic problem: more users mean greater losses. When traffic becomes a cost rather than an asset, 'traffic-only' thinking turns into a 'traffic trap.'

Even more fatal is that traffic does not equal loyalty. Red envelopes solve the problem of 'getting users to try' but fail to answer the core question: 'Why must users choose you?'

As one industry insider put it, cash subsidies are merely 'door-openers.' Long-term user retention ultimately depends on the core value of AI products.

When AI Assistants Become a 'Memory Burden'

Another byproduct of traffic thinking is the wild expansion of entry points.

The early logic of 'staking claims first, then refining the experience' made sense: whoever retained users first won at the starting line. Thus, every giant crammed AI assistants into all their products.

Microsoft’s predicament best illustrates this awkwardness: open Word, and there’s Copilot on the side. Switch to Teams, and there’s another Copilot. Jump to GitHub, and Copilot’s still there. Edge has Copilot, and even Windows itself has a built-in Copilot.

Users must toggle between five disconnected entry points, none sharing dialogue states or historical context. The AI assistant becomes part of the user’s memory burden.

Entry point redundancy is a 'midlife crisis' for nearly every AI product company.

In the early stages, such redundancy was tolerable. Back then, AI was mostly a feature demo—summarizing meetings, drafting emails, or completing code snippets.

But as the industry evolves from chatboxes to intelligent agents, AI needs more than a chatty tool. It requires a system that can plan, execute, review, and hand off tasks across scenarios. You can’t build that with several assistants wearing the same badge but operating in silos.

When user expectations shift from 'Can it chat?' to 'Can it finish this?', too many entry points become enemies of efficiency.

Thus, the giants almost simultaneously took the same action: consolidating scattered entry points.

Google merged Gemini into Chrome, its traffic hub, as a persistent sidebar, later upgrading it with 'auto-browsing' capabilities.

Microsoft is preparing a 'super app' to unify Copilot Chat, GitHub Copilot, and other products into a single interface. Baidu merged multiple Wenxin Yiyan websites into the Baidu Wenxin platform.

These moves signal the same issue: 'more' entry points are no longer an advantage but a burden.

As AI evolves from a chat tool into a task execution system, users need a unified, coherent intelligent agent that remembers context—not a crowd of disconnected 'AI avatars.'

From 'Entry Point Wars' to 'Tipping Points'

Doubao’s subscription fees and Wenxin’s merger straddle the fault line of this logical shift.

Doubao priced 'task completion' with subscriptions, while Wenxin expanded capabilities for free, turning them into infrastructure. Though their paths diverge, both respond to the same change.

At the core of this change lies a single word: 'tipping point.'

At the 2026 Volcano Engine FORCE Conference, Volcano Engine president Tan Dai introduced this concept: only when model capabilities surpass a 'tipping point' can they truly meet enterprise and individual needs in production scenarios.

Globally, the first video generation model to cross this threshold was Seedance 2.0, while in programming and intelligent agents, it was America’s Claude Opus 4.6.

Tan noted that model usage previously peaked on weekends and holidays, indicating recreational use by individuals. Now, Seedance and Doubao see peak usage on weekdays, signaling integration into productivity workflows.

Doubao 2.1 Pro launched in this context. In programming benchmarks, it matched Claude Opus 4.7; in SciCode scientific computing, it scored 59.8, surpassing Opus 4.7 and GPT-5.5.

Doubao’s office task mode supports local PC operations, browser use, Skills activation, and includes an Office suite for graphic design, video design, and generating shareable app websites.

Users can instruct Doubao to organize computer folder materials, check copy from specific documents on given dates, and post timed Weibo updates via official accounts.

This is what AI should do: not chat with you but get things done.

Doubao Professional’s pricing revolves around this: the 68 RMB/month Standard plan targets new professionals and students; the 200 RMB/month Enhanced plan suits self-media creators and SMEs; the 500 RMB/month Premium plan caters to enterprises, R&D, and professional design—high-compute scenarios. The core difference lies not in 'usability' but in 'capacity' and 'quality.'

Meanwhile, Baidu chose another path. Wenxin Yiyan merged multiple entry points and added features like AI college application reports, AI PPT generation, and deep academic research aids. Pricing remained free.

Baidu’s strategy is to deepen AI capabilities, making them indispensable infrastructure rather than competing for user attention through multiple entry points.

Both routes diverge but align logically: shifting from 'acquiring users' to 'getting things done,' from 'traffic scale' to 'capability depth.'

Why Did the Giants Stop Burning Money?

The answer is simple: traffic has peaked.

Mobile internet user growth has stalled. QuestMobile data shows that in Q1 2026, Doubao had ~140 million daily active users and ~345 million monthly active users, exceeding the combined MAU of the second- and third-ranked apps.

In China’s AI app market, Doubao’s penetration has hit its ceiling. Most potential users have already arrived; no amount of burning money can attract more.

When the increment market becomes a stock market , competition shifts from 'acquisition' to 'retention.' And what drives retention? Not red envelopes. Users stay only if something is genuinely useful.

First Finance’s 'One Rating' points out that China’s large model industry has moved beyond pure C-end traffic wars and price battles. Commercialization now hinges on B-end enterprise productivity cycles.

C-end scenarios face dual bottlenecks: low user willingness to pay and inverted reasoning costs, making long-term profitability unsustainable. In contrast, B-end scenarios like AI coding and long-chain agents offer high value density.

Industry data backs this trend. By 2026, 45% of AI Agent companies adopted some form of usage-based pricing.

The entire AI Agent market is projected to grow from $7.38 billion in 2025 to $47.1 billion by 2030.

The giants didn’t stop burning money by choice—the logic no longer works in the AI era. Every yuan burned becomes a computing bill.

Instead of acquiring 'zombie users' who leave after claiming red envelopes, it’s wiser to invest in model upgrades so users willing to pay for value stay.

Two Paths, One Answer

Doubao’s subscription fees and Wenxin’s free model may seem opposite, but they converge on the same point.

Doubao’s logic: the free version meets daily needs for most users. For complex office and productivity scenarios, higher quotas and stronger models require payment. Free secures the base; paid unlocks deep value.

Wenxin’s logic: instead of confusing users with multiple entry points, integrate capabilities into one platform and deepen them. Free expands the user base; depth builds ecological barriers.

Both paths answer the same question: when traffic dividends vanish, what is the true moat?

The answer is capability.

Not 'how many use you,' but 'how much you can get done.' Not 'how much time users spend with you,' but 'whether things still work when they leave.'

When AI transforms from a 'toy' to a 'tool,' from 'chatting' to 'doing,' traffic is no longer the core metric. Completion rates, accuracy, and efficiency gains become the hard currency.

Reviewing AI’s trajectory over the past few years, a clear path emerges:

Phase 1: 'Concept Era'—large models launched, parameters compared, leaderboards chased. More parameters meant superiority; higher rankings meant dominance.

Phase 2: 'Traffic Era'—money burned for user acquisition, entry point wars, red envelope blitzes. More users meant victory; more downloads meant success.

Now, the industry enters Phase 3: the 'Value Era.'

Doubao starts charging, Wenxin undergoes merger, access points are consolidated, and capabilities are deepened. All these signals indicate one fact: traffic is just a bubble, and quality is the ultimate criterion.

If the users acquired by spending 4.5 billion can't be retained, it will all be for naught. However, an AI that can help programmers write code well, assist designers in generating materials, aid white-collar workers in organizing documents, and assist students in completing research will naturally retain users, who will also be willing to pay for it.

This is not a debate between the 'free' and 'paid' routes, but a value choice between 'shallow' and 'deep'.

When access points are no longer scarce, traffic is no longer cheap, and computing power becomes a hard cost, the 'traffic-only' theory will eventually be replaced by the 'value-only' theory. Big companies are proving one thing with their actions:

Instead of having 100 million people use it occasionally, it's better to have 1 million people rely on it every day.

The bubble will eventually recede. Only products that truly create value can survive through cycles.