China's leading internet companies, including ByteDance, Tencent Holdings, Alibaba Group Holding, and Baidu, have begun imposing strict token quotas on employees as surging generative AI costs force a rethink of the freewheeling usage culture that defined the sector's early AI adoption phase, according to staff who spoke on condition of anonymity.

From badge of honour to rationed resource

When generative AI first swept through China's technology sector, workers faced a clear directive from management: use AI and use it often. Consuming vast numbers of AI tokens — the basic unit of computing power needed to process text or write code — became a badge of honour, with companies often viewing a high token count as a sign that employees were productive and taking initiative. That era is now over. Computing allowances are being sharply rolled back as firms introduce formal token quotas to rein in rising AI-related expenditure, according to multiple employees who described the shift.

ByteDance leads with co-pay model

At ByteDance, one employee said staff using closed-source models must now co-pay out of pocket. The company reportedly reimburses roughly half of external tool costs, capping annual payouts at around US$1,000 for technical staff and interns, and US$300 for non-technical employees, according to the staff member. The co-pay structure effectively transfers a portion of the cost burden directly to workers — a notable reversal from the earlier posture in which unlimited or near-unlimited AI access was positioned as a workplace benefit.

Why it matters

The shift signals that the economics of generative AI at enterprise scale are proving far harder to absorb than initially anticipated. Token costs accumulate rapidly when thousands of engineers and non-technical staff run large language model queries throughout the workday, and the aggregate bill can rival or exceed traditional software licensing expenses. The pullback also complicates the narrative that China's tech giants have been aggressively out-deploying their Western counterparts on internal AI adoption. Rationing suggests that raw usage volume was never a sustainable proxy for productivity or return on investment.

The competitive backdrop

The token-rationing trend arrives as global firms, including those tracked by Goldman Sachs in its enterprise AI spending analyses, grapple with the same fundamental tension: the cost of inference at scale versus the productivity gains that justify it. Platforms such as QoderWork have emerged specifically to help enterprises monitor and cap AI token consumption, suggesting the problem is industry-wide. For Alibaba, Tencent, Baidu, and ByteDance, the stakes are particularly acute because each also operates its own foundation models and cloud infrastructure — meaning internal token consumption directly competes with revenue-generating external workloads for the same compute capacity.

What's next

The most immediate question is whether token quotas will dampen the pace of internal AI tool adoption just as these companies are racing to embed generative capabilities into core products. Employees subject to hard caps may self-censor usage, slowing the feedback loops that help enterprises identify the highest-value AI applications. Watch for whether ByteDance's co-pay model spreads to peers, and whether any company publicly quantifies the cost savings achieved — a disclosure that would offer the clearest window yet into how expensive enterprise-scale AI deployment has become in China.