Why Exported Labor Automation Breaks Standard Models of Trade
Marginally Cheap AI Contradicts Heckscher-Ohlin’s Fixed Capital/Labor Assumption
Basically, there will be a near-zero marginal cost for each additional automated worker created. Under H-O we treat every extra unit of output as requiring additional capital and labor to produce. The labor and capital required to produce an AI agent, by contrast, should be almost entirely up front. It comes from a single huge initial sunk cost of model training, after which the costs of subsequent production are low. What follows is basically just the cost of inference and so it is partially excludable, non-rival, and intangible. Once the model weights are trained, cloning another instance uses almost no extra capital or labor.
A near-zero marginal cost may only be true in the short-term, though, as markets may constantly desire newer models, meaning that skilled labor will always be demanded to produce subsequent AI models in the long term. Of course, in the yet-longer run, subsequent models might be trained by AI’s instead of humans.
In any case, unlike Hecksher-Ohlin assumes, export of this trade encourages a potentially destructive form of trade specialization. While investment and production curves shift toward exporting more AI agents, it shifts away from other sectors of the economy. The resulting structural transformation likely results in less of all kinds of labor, rather than to more skilled labor.
Competition is Imperfect
H-O also assumes constant returns and perfect competition. The economics of digital platform services goes against that grain. Large fixed costs for R&D creates economies of scale and network effects, where a few industry champion platforms set prices above cost in markets where competition is monopolistic. This conforms with Krugman & Helpman New-Trade Theory, where scale drives trade flows, not relative factor abundance.
Factors are Highly Mobile
Under H-O, capital stays “at home” while goods move abroad, but digital trade shows that capital, data, and compute infrastructure can relocate at will to reduce electricity costs or regulatory hurdles, for example. Cross-border data flows show firms of all sizes renting foreign cloud capacity instead of shipping hardware abroad, erasing the home-country tie that H-O presumes. As training and operating AI models requires increasingly vast power supplies, data centers will be developed in energy-rich countries like the United Arab Emirates and Norway.
Massive investment in artificial intelligence promises to augment productivity and eventually automate job tasks and whole jobs. Historically, automation in factories has displaced humans on assembly lines.
Today, large language models are capable of doing some white collar knowledge work. But eventually, models and robots seem likely to be increasingly capable of performing cognitive and physical tasks and jobs. Labor automation is often framed in terms of national politics, such as in a recent piece from the office of Senator Bernie Sanders that calls for swift action to protect American workers. It is worth remembering, however, that automation will not be a one-country issue.
Artificial intelligence is on the verge of becoming an exportable factor of production. Instead of trading goods made by workers, countries will increasingly trade the workers’ functions themselves, packaged as agentic AI services delivered over the cloud. That shift matters because our dominant trade frameworks assume that labor is domestic, rival, and costly to scale. A labor-automating AI service is none of those things.
If firms in the United States or China can sell millions of near-zero-marginal-cost “digital workers” to foreign firms, the result will be faster productivity growth but also downward pressure on wages, new terms-of-trade frictions, and a set of distributional problems that would look more like immigration politics than like ordinary digital trade. States will need new instruments, some perhaps akin to work-visa regimes, to tax, monitor, and pace the inflow of automated labor.