Han Dieperink: The returns of cheap AI
This column was originally written in Dutch. This is an English translation.
By Han Dieperink, written in a personal capacity
When the Chinese AI model Kimi K3 was released in mid-July, semiconductor stocks reacted as they increasingly have in recent months: they fell. The logic was straightforward: if artificial intelligence becomes ever cheaper, less computing power will be required, and therefore fewer chips will be needed. Yet that conclusion is mistaken.
With 2.8 trillion parameters, Kimi K3 is the largest Chinese AI model to date. Precisely because of its scale, it requires vast amounts of memory: even a compressed version occupies around 1.4 terabytes. Demand for chips and memory is therefore not disappearing. What is happening is that AI is becoming remarkably inexpensive. Kimi K3 delivers state-of-the-art performance at between one-third and one-half of the cost of the most expensive American models. Meanwhile, since the launch of ChatGPT, the cost of replicating this level of intelligence has been falling by around 60% a year.
Why Cheaper Means More
In the nineteenth century, the economist William Stanley Jevons observed that more efficient steam engines did not reduce coal consumption; they increased it. As coal became cheaper per unit of work, people found ever more uses for it, causing total consumption to rise. AI is following the same pattern. As intelligence becomes cheaper, we deploy it everywhere: in customer service, software development, medical diagnostics and thousands of smaller tasks that were previously too expensive to automate. As a result, demand for high-speed memory is growing faster than supply. In the short term, the beneficiaries are therefore the providers of the underlying infrastructure: manufacturers of chips, memory and data centres. Moreover, all that computing power requires electricity. Energy demand rises alongside it, and someone ultimately has to pay that bill.
The importance of productivity
In the long run, however, only one thing truly matters: productivity. New technologies make us more productive, and in theory wages should rise accordingly. Those who produce more should earn more. In practice, that link has weakened. In the United States, wages and productivity moved in tandem until the 1970s. Since then, the gap has widened year after year. Productivity is now more than five times higher than it was in 1947, yet real hourly wages have only tripled. A similar pattern can be observed across Europe and Japan.
Part of this divergence is statistical, as wages and productivity are adjusted using different price indices. Most of it, however, reflects structural changes. The most important is the decline in workers' bargaining power. Trade unions have weakened and temporary employment has become more common. Employees who can easily be replaced have little leverage to demand higher pay. Globalisation has reinforced this trend, as companies can relocate work to lower-cost countries, and the mere possibility of doing so suppresses wage growth. Technology itself has also played a role. Increasingly, profits are generated by software and machines rather than people, with the returns flowing to the owners of capital rather than labour. In addition, a growing share of compensation is absorbed by rising healthcare and pension contributions, leaving net wages lagging behind even as workers become more productive.
The gap between wages and productivity is ultimately the result of policy choices: choices about taxation, bargaining power and how society rewards work. And because it is the product of choices, it can also be changed through new ones. This is where AI enters the picture. Technology can be designed either to augment people or to replace them. The latter path is known as the Turing Trap. It is shaped by incentives. Companies can often deduct investments in machines from their taxable profits, while hiring employees comes with payroll taxes and social security contributions. The incentive, therefore, is to replace labour rather than strengthen it. Now that Moonshot, the company behind Kimi K3, has made its model freely available, any business can run and adapt it. AI is becoming a commodity, as readily available as electricity from a wall socket. It will make us more productive, but much of the economic benefit is likely to accrue to capital.
Uncomfortable yet attractive
For investors, this is both uncomfortable and attractive. Uncomfortable because an economy that fails to share the gains from productivity ultimately risks undermining its own customer base. Who will buy all the goods and services produced by increasingly intelligent machines if purchasing power fails to keep pace? And the wider the gap becomes, the louder the calls for higher taxes and stricter regulation. That political response is itself an investment risk: difficult to price, but impossible to ignore.
At the same time, it is attractive because shareholders stand on the winning side of the equation. Productivity gains that do not translate into higher wages ultimately accrue to the owners of machines, data centres and capital. The crucial question, however, is who those owners are. Not necessarily the developers of the most advanced AI models, because their competitive advantage is eroding by roughly 60% each year. Today's market leader may be overtaken tomorrow by a cheaper competitor. The real AI dividend will be captured more broadly by companies that adopt inexpensive AI most effectively. A bank that automates its back office, a logistics company that optimises its routes, or a manufacturer whose machines become more intelligent can all reap the rewards without building their own AI models. The winners, therefore, will not necessarily be the household names dominating today's headlines, but the businesses that convert cheap intelligence into higher margins. For investors, the lesson is clear: do not focus solely on who has the best technology, but on who generates the greatest economic value from it.
What really matters
Over the long term, productivity growth remains the only metric that truly matters. Cheap AI is almost certain to accelerate that growth. The real question is how the gains will be distributed. The answer will shape not only what our societies look like in ten years' time, but also tomorrow's investment returns. For society, the issue is fairness; for investors, it is performance. The two are far more closely connected than they may appear. Anyone investing today is, in effect, investing in that question.