1,400 RMB per month.

That's the starting AI token quota Tencent has set for ordinary developers. According to InfoQ, Tencent is rolling out a developer token quota system internally: ordinary developers start at 1,400 RMB/month in token quotas; senior developers, algorithm roles, and specific project members can apply for higher quotas, with some exceeding 10,000 RMB/month. This isn't charging — it's turning AI usage into a quantifiable, traceable, governable cost item.

From “Use Freely” to “Use Wisely”

Just a year ago, internet companies' attitude toward AI was still “AI for all,” “AI First” — handing out ChatGPT Plus to everyone, buying enterprise licenses, encouraging people to try AI tools; how much it spent didn't matter, what mattered was not falling behind. Back then AI budgets were “innovation budgets,” strategic investments; nobody calculated ROI because everyone felt “if we don't use AI now it'll be too late.”

In H2 2026, the wind has clearly shifted. Tencent's token quota system is a landmark: it means enterprises are starting to manage AI as a normal IT cost — just like giving employees computers, phone allowances, or cloud server resources: there are quotas, standards, overages require approval, and unused balances don't roll over. AI is no longer a “faith-based investment” — it's a production tool whose ROI must be calculated.

Behind this is explosive growth in AI usage. In 2025 people primarily used AI for chatting, looking up information, and writing a few paragraphs of copy — token consumption was limited; in 2026 Agents have proliferated, and a single AI coding tool's daily token consumption can be tens to hundreds of times that of casual chat — a heavy Claude Code user burning several thousand RMB/month in tokens is normal. Tencent's QClaw product manager noted in a response that the token mechanism would become more granular in July, essentially acknowledging: without governance, AI costs will spiral out of control.

Doubao's pricing confirms cost pressure from another angle. Doubao 2.1 Pro is priced at 30 RMB per million output tokens, and Doubao's daily token calls have reached 180 trillion; even by extremely conservative utilization estimates, daily token costs are in the hundreds of millions of RMB. Free consumer users are customer acquisition costs, but B2B enterprise clients and internal usage must have clear accounting.

What Will an ROI-Driven Approach Change?

Enterprise AI entering the cost-accounting phase will trigger several chain reactions.

First, AI tool procurement criteria will change. Previously AI tools were purchased based on “how smart the model is” or “how many features it has”; going forward it will be about “can it help me save/make money.” However cool a tool may be, if it can't clearly quantify how much efficiency it improved, costs it reduced, or incremental revenue it generated, it won't pass the CFO. For AI startups this is both a challenge and an opportunity — the era of telling stories and showing demos is over; you need real ROI data.

Second, the cost-performance advantage of self-built and open-source models will become prominent. When token consumption reaches a certain scale, third-party API costs exceed the cost of self-deploying open-source models or even self-built models. Meituan restricting Doubao to push LongCat, Alibaba pushing Tongyi Qwen company-wide, Tencent pushing Hunyuan — all are fundamentally rational choices after cost accounting. That's also why privacy + value AI platforms like Venice AI can reach a $1 billion valuation and $70 million ARR in two years — enterprise customers want AI capabilities, but don't want vendor lock-in or wasted spend.

Third, employees' AI usage habits will change. With quota limits, people won't casually have AI generate piles of useless content; instead they'll use AI more deliberately — just as when mobile data was unlimited people watched video freely, but with limited data they cached on WiFi first. This isn't a bad thing; it forces people to ask “does this task really need AI? Is the ROI of using AI high?” — which actually improves real AI utilization efficiency.

From the campaign-style promotion of “everyone must use AI” to the rational usage of “let's calculate whether AI is worth it,” this marks the AI industry maturing. The sign that a technology has truly landed isn't when everyone is discussing it — it's when nobody discusses it anymore, because like water or electricity it has become infrastructure that requires monthly payment and metered usage.