Traditional wage dynamics are headed for the scrapheap. While we often view AI agents as a new 'workforce,' that perspective is an intellectual trap. In a recent preprint on arXiv, Siki Zhu of the University of Illinois (UIUC) demonstrates that an agent is not an 'employee,' but a technology for converting capital into units of cognitive labor. Where labor scarcity once dictated market prices, the equilibrium rate is now tethered to the computing market. Welcome to a reality where an analyst’s hourly rate depends on TSMC’s lithography costs and electricity tariffs, rather than the ambitions of top-tier university graduates.
Applying Mankiw’s classical framework, Zhu introduces the concept of the Compute-Anchored Wage (CAW)—a hard ceiling on human compensation. The logic is ruthless: for substitute tasks—ranging from legal due diligence to customer support—your salary is limited by the product of the GPU rental rate and the model’s computational intensity. If a hypothetical GPT-5 can handle a paralegal’s workload for a few dollars an hour, the market valuation of a human staffer will inevitably anchor near that mark. It is an ironic twist of technological progress: we used to fight for talent; now we will fight for infrastructure access.
For businesses, this signals a radical restructuring of cost sheets—a steady migration from payroll expenses to capital expenditures (Kc, or computational capital). The 'infinite elasticity' of AI agents allows cognitive functions to scale with near-zero marginal costs, creating a surplus of supply in sectors where companies once spent years hunting for specialists. In our view, this is a direct path to wage stagnation in sectors where AI serves as a direct substitute. The labor market is losing its autonomy, becoming a mere derivative of the volatile semiconductor market.
However, a 'humanity premium' persists, but only where productivity cannot be scaled by simply adding more GPUs to a cluster. Zhu’s model, based on a CES production function, draws a clear line: if your work can be accelerated by doubling the number of servers, you are in the danger zone. Modern C-suite management must realize that the cost of a professional hour is no longer dictated by labor laws, but by the hard economic logic of hardware and algorithms. You are either managing this capital, or your value is trending toward the cost of the electricity used for inference.