Artificial intelligence needs customers too

I listened to part of Joseph Stiglitz’s remarks at the Hindustan Times Leadership Summit in New Delhi on 14 November 2026. The Nobel laureate was discussing a possible AI investment bubble, but then raised a question that interests me more than predictions about the markets: if AI replaces a substantial share of human work, who will buy the goods and services it helps produce?

It is a basic macroeconomic question, which makes it difficult to dismiss. An individual company can raise productivity, cut costs and produce more with fewer employees. If this happens across many industries at once, the overall result also depends on what happens to incomes. Wages are a major source of consumer demand. If a growing share of economic output accrues as profits without returning to circulation through investment, new incomes, lower prices or redistribution, purchasing power can weaken. Decisions that are rational for each company may, taken together, create a wider economic problem.

Stiglitz connects this risk to the investment boom. Data centres, semiconductors, energy and computing infrastructure require vast amounts of capital. To justify that spending, firms need sufficient revenue and margins. Yet competition, including from cheaper and more open models, could squeeze the margins investors are counting on. That is a persuasive distinction: a technology can spread, improve productivity and transform the economy even if some of the investments made to develop it turn out to have been excessive. Technological and financial success are not necessarily the same thing.

I would not, however, treat Stiglitz’s forecast that the bubble will burst as an established fact. AI investment is indeed supporting the US economy, but measuring its contribution to growth is harder than some headline figures suggest. Much of the hardware installed in data centres is imported, and spending on equipment does not translate automatically into an equivalent increase in US GDP. A Congressional Research Service report dated 12 November 2026 explicitly sets out these qualifications. The employment effects also remain uncertain.

It is important to resist sweeping claims about jobs. The International Labour Organization considers the transformation of tasks more likely than the complete disappearance of jobs across most occupations exposed to generative AI. A Stanford Digital Economy Lab paper revised in August 2026 finds no economy-wide mass displacement of US workers, but identifies a development I find particularly significant: employment among 22-to-25-year-olds in AI-exposed occupations is 19 per cent below where it would have been had it grown at the same rate as employment in less-exposed occupations. The difference is driven mainly by reduced hiring. This is observational evidence, not proof that AI alone caused every missing job, but it would be unwise to ignore it.

The difficulty also lies in managing the transition. New activities may emerge, prices may fall, and many workers may become more productive with AI tools. These are genuine possibilities. They do not yet tell us how quickly new jobs will appear, what quality of work they will offer or how people who lose income or opportunities will be supported in the meantime. I am particularly concerned about people entering the labour market when some of the junior positions through which experience is usually acquired may be disappearing.

There is also an international dimension, which Stiglitz makes clear in discussing India. A country may see demand fall for services it previously exported, including some IT services, while the profits generated by automation accrue to firms based elsewhere. Within a country, governments can at least debate taxation, education, income support and public investment. Redistribution across borders is far less automatic. Europe faces the same question: who develops the models and owns the infrastructure matters, but so does who captures the productivity gains and where those gains are spent.

In his proposed progressive AI agenda, Stiglitz emphasises competition, effective regulation and using the benefits of technology to improve education, healthcare and the green transition. That seems a sensible place to begin. AI governance cannot be confined to model safety or the prevention of technical errors. It must also address the economic consequences of the choices we encourage: income distribution, access to work, market power and the ability of tax systems to finance public goods.

I still think artificial intelligence can deliver extraordinary benefits. That is precisely why I find it unconvincing to assume that productivity growth will, by itself, ensure that those benefits are shared in a socially sustainable way. We may build machines that can perform an increasing amount of work. We still have to decide how to turn that advantage into income, services and opportunities for people. That is an economic and political choice, not a property of software.

Sources