BIS Working Paper 1381 says AI growth could stall on consumer demand

The AI investment boom will hit its limit in household demand. Chips and energy come second.

That is the argument Phurichai Rungcharoenkitkul makes in BIS Working Paper 1381^[1], published by the Bank for International Settlements on 12 October 2026. In the scenario his model rates as highly plausible, growth speeds up for a short time and then falls below the 2% of the pre-AI era.

The industry is looking somewhere else. CoreWeave sells AI computing power, and its blog has a post on how to break the bottlenecks holding back AI growth^[2]. The market counts megawatts and available GPUs.

The paper counts what goes into the shopping cart.

1820: the year Malthus asked who would buy it all

In 1820 Thomas Malthus published his Principles of Political Economy. His question to David Ricardo was simple. Who buys the goods made by capital that grows faster than consumption?

Ricardo replied that supply creates its own demand, as Jean-Baptiste Say had argued.

In 1902 John Hobson took up the idea again in Imperialism. In his view, surplus capital went looking for returns abroad because wages at home bought too little.

The mechanism is the same in both cases. Investment races ahead, consumption falls behind, and the return on capital drops.

2026 adds a new player: a machine that does the work and also does the research.

Where the model puts today's cycle

The paper uses a task-based model of endogenous growth. In it, innovation is paid for by the rents earned on consumer goods.

The scenarios depend on two main levers:

• Speed: how quickly AI technologies arrive.
• Duration: how long innovators keep their rents.

With transformative AI, growth takes off, reaches 5.5% after two decades and keeps climbing. With the demand bottleneck, it speeds up briefly and then falls below the pre-AI rate of 2%, because innovation slows for good. The scenarios in between offer moderate gains at best, according to Working Paper 1381^[3].

The paper compares itself openly with other research. Some recent estimates it cites see growth above 10% within a few years. The calibrated model projects more moderate growth, even compared with Korinek et al. (2026).

This pattern is structural. The paper measures outcomes over two decades.

Why spending on machines starves research

The mechanism works in three causal steps.

1. Large-scale automation soaks up resources and pushes a growing share of output into investment.
2. Households then buy a smaller share of total output, and the rents on consumer goods shrink.
3. Research and the entry of new firms cost money, and consumer demand pays for them. With smaller rents, there is less incentive to keep innovating with AI.

The paper asks the question directly. Can AI keep growth high while households buy a smaller slice of what the economy produces?

Once fitted to the data, the model picks the demand scenario. It calls it highly plausible on the evidence.

For years I have argued that AI capex is a debt loop, and that the systemic risk sits with whoever lends to data centers. The paper puts that view on new ground. Data center debt gets repaid from revenue that comes from consumption.

Automation squeezes exactly that.

Where to move capital, and when

1. Family offices and sovereign wealth funds, 36 months. Data center returns depend on final household demand. When weighing credit to AI infrastructure, look at the quality of downstream revenue before the pace of construction.
2. Chief Risk Officers, 24 months. The paper stresses high uncertainty about the natural rate of interest. A scenario with high investment and growth below 2% belongs in stress tests alongside the takeoff scenario.
3. CFOs and boards, 18 months. An investor story built on double-digit growth relies on a different scenario from the one the calibrated model finds plausible. Build guidance on the moderate path as well.

My thesis, and the facts that would overturn it

My position is clear. The AI investment cycle will hit its ceiling in household demand before it hits one in chips or energy.

This is a single model, and a working paper reflects its author's views. The outcomes depend on parameters such as how long rents last, which are hard to measure in the data.

Three things would change my mind:

• New consumer products created by AI that expand household spending instead of replacing it.
• Innovator rents that last longer than the calibrated model expects.
• New work from the BIS or other authors that uses more recent data to make the transformative scenario more plausible.

The indicators that will settle it

• Household consumption as a share of GDP in the economies investing most in AI.
• Fed, ECB and BIS estimates of the natural rate, with the width of their uncertainty bands.
• Spreads on private credit tied to data centers, read alongside end-customer revenue.

Malthus lost the 1820 debate. His question is back today in a model from the central banks' bank, and now it is up to those who lend to data centers to answer it.