A growing number of Chinese firms have embraced “All in AI” strategies only to find that the cost of those strategies is quietly migrating from company balance sheets to employees’ wallets. Engineers and other knowledge workers report that tools once supplied by employers — computers, software licences and servers — are increasingly being reclassified as personal productivity aids. The net effect: firms raise efficiency expectations while leaving individuals to shoulder subscriptions and token bills.
For many organisations AI is being framed as a “personal efficiency tool” rather than a foundational production input. That change in classification has real consequences. Timetables, staffing plans and performance targets are now often set on the assumption that staff will use paid models and plugins; yet companies are coy about who should pick up the tab. The logic is stark: efficiency measured as if AI were used, but costs calculated as if it were not.
The phenomenon ranges from informal pressure to systematic deduction. In small start‑ups employees were told to “try” Copilot and other coding assistants and left to sort out payments themselves. In a more explicit case, a large cross‑border e‑commerce company reportedly procured Cursor and Copilot licences centrally and then passed the bill to staff by deducting about RMB 7,200 per person at year‑end. More formally, publicly traded Kunlun Wanwei has ordered mandatory AI coding accounts, set a monthly $100 quota per developer, demanded a 50% jump in developer productivity and promised to fold compliance into mid‑year performance reviews — with bottom‑ranking staff facing elimination.
Those measures show both sides of the shift. Some companies do pay or subsidise tools, signalling that AI is production infrastructure. But capped allowances, hard productivity multipliers and punitive performance regimes reveal another logic: production capacity is now measured in consumed compute and model usage. When quotas are tight, employees who exhaust a company allowance must either throttle their use or pay out of pocket to hit targets.
That tension is intensified by frontier model economics. Top‑tier models now cost many times more than smaller ones, especially in low‑latency or high‑throughput modes. A leading commercial example is Anthropic’s latest flagship, where “fast” or “adaptive” inference can multiply per‑token output costs sixfold compared with baseline rates. For firms chasing marginal gains, the competitive imperative is clear: faster, higher‑quality models shorten iteration cycles and boost revenues — if you can afford them. For individuals on modest salaries, monthly subscriptions to ChatGPT Plus, Grok, domain‑specific legal assistants and other paid services can add up to several hundred renminbi; that is a meaningful slice of take‑home pay for many.
The labour implications are significant. Shifting the burden of AI tools onto employees risks entrenching inequality between well‑funded firms and those with scarce budgets, while redistributing what was previously a corporate investment into personal consumption. It can also distort incentives: higher nominal productivity targets do not necessarily translate into shorter work hours or higher wages but can be used to compress headcount and tighten performance gates. The result is a potential squeeze on morale and an elevated turnover risk among mid‑level technical staff.
This is not just an administrative or HR problem; it is a structural change in the ownership of production inputs. Machines once sat in factory workshops, servers once sat in corporate data centres, and now compute — measured in tokens, latency and API calls — is becoming the unit of production that determines employability. If employers treat compute as a quasi‑private expense, they shift market power and bargaining leverage away from workers.
Companies seeking to scale AI responsibly will face choices. One path is to internalise costs fully and treat compute as a capital expense, absorbing variability through central procurement and governance. Another is to ration usage tightly and incorporate quotas into compensation design, which could create a two‑tier workforce. Governments and regulators may also intervene with transparency rules on reimbursable expenses, taxation guidance or labour protections. For workers, the arithmetic is unambiguous: in the near term some will pay to remain competitive; in the longer term the balance between firm and worker over who owns the means of production will shape career trajectories across the tech sector and beyond.