Gross domestic product measures the economic transaction in an economy, according to quantities bought and sold and market prices. But does this method work for AI? Anton Korinek and Patrick McKelvey work through the question in “Where is AI in GDP Statistics (Peterson Institute for International Economics, May 2026, both a readable Policy Brief and an underlying research paper are available).
Here’s the basic issue. The authors calculate that the total amount currently spent on AI computing power is (roughly, depending on underlying assumptions) about $250 billion per year. This measures the AI bought and sold in the market, and thus is clearly part of GDP.
However, the AI chips are becoming more efficient:”As chips became more efficient, each dollar of compute spending bought more physical computing capacity. Measured in H100 equivalent units, US AI computing capacity grew at more than 200 percent per year, outpacing nominal spending.” Also, the AI algorithms are becoming more efficient, so that the quantity of computing power needed to achieve a fixed level of AI has been falling by about two-thirds per year. Put these together, and the quality-adjusted increase in AI is more than 2000% (that is, a twenty-fold increase) per year.
So here’s a situation where the amount spent on AI and captured by GDP is rising quickly, but the actual capabilities of that AI spending are rising much, much faster. How should measures of the size of the economy deal with this situation?
The question isn’t totally new. Government statisticians have for some years now used “hedonic” adjustment. The idea is to define a good, like a computer or a television or a car, as a set of characteristics. The qualities of these characteristics are improving over time. Thus, when we say whether the “price of a computer” or the “price of a car” has changed, we want to do an apples-to-apples comparison holding the characteristics of that product constant. Thus, the Consumer Price Index measure for computers and related goods looks like this:
Just to be clear, this graph doesn’t mean that the amount you personally paid out of pocket for a computer fell by three-quarters from 2005 to 2018. It means that the price of buying a computer with the same characteristics as a standard 2005 computer fell by three-quarters by 2018–but of course, few of us in 2018 would have wanted to buy that version of the computer from 2005, even if it was available. The authors explain:
That AI’s growing footprint is so faintly visible in national GDP statistics has, in part, a straightforward accounting explanation. Nominal AI revenues grow only moderately because per-unit prices for any given level of AI capability fall almost as fast as quality-adjusted output rises. In the semiconductor industry, this pattern played out for decades: each generation of chips was dramatically cheaper per unit of performance than the last, so the semiconductor share of GDP remained modest even as quality-adjusted output expanded enormously.
This kind of hedonic adjustment could readily be used for AI as well, but for purposes of measuring GDP, it has limitations as well. GDP is a measure of market transactions, which happen when the technology behind supply-side production of goods and services meets the preferences and tastes behind demand-side purchases of those goods and services. Hedonic adjustments focus on adjusting for changes in the quality of what is produced, but don’t directly take into account what value buyers may place on these quality changes.
For example, perhaps the speed and power of AI tools have considerable value to users up to some point, but then additional speed and power have diminishing marginal value beyond that point. Just because quality-adjusted increase in output of AI has risen by 2000% per year in the last couple of years doesn’t mean that the value of AI tools to users has risen by 2000% per year.
Beyond the details of questions like these, Korinek and McKelvey are making a broader point: sensible thinking about the economic effects of the emerging AI tools needs to be rooted in data, but economic data about a very fast-changing industry can be hard to collect and can involve delicate issues of interpretation.
