Where does AI stand in GDP statistics? (Eng) 2026

Anton Korinek and Patrick McKelveyMay 2026 The artificial intelligence (AI) economy in the United States isgrowing at extraordinary rates of over 2,000 percent per year yetis leaving only a small mark in the nation’s GDP figures. This is a McKelvey 2026)that nominal AI compute spending grew by morethan 140 percent per year each in 2024 and 2025, raw computecapacity by more than 200 percent per year, and quality-adjustedAI output by more than 2,000 percent per year. The divergencebetween this picture of the AI economy and the one drawn by We argue that US statistical agencies and economicpolicymakers should start now to assemble better data on AIactivity in AI satellite accounts—focused subsets of the national incorporate AI productive-capacity measures into medium-termprojections and scenario analysis. Building this measurementinfrastructure today, while the AI sector is still small in nominal The question of where AI lies in the US GDP statistics hasbecome a recurring mystery in economic commentary. FrontierAI capabilities are advancing at what industry observers considera remarkable pace, with some seeing the possibility of artificialgeneral intelligence within just a few years. Yet when one looks One natural explanation is that AI adoption takes time, andthe kinds of broad productivity gains economists associate withgeneral purpose technologies typically arrive years after the The visibility problem is structural. The conceptual architectureof GDP measurement was developed in the mid-20th centuryto track an economy organized around manufacturing. Thatarchitecture has served well for a long time and continues to do so moderate pace, do not capture what is happening. Moreover,AI-related activity is dispersed across a long list of industries—cloud services, software publishing, data processing, professional AI is the latest in a series of fast-moving technologies that haveraised measurement concerns; semiconductors and the internetgenerated similar debates in their time. But there is a feature of the AI case that distinguishes it from those precedents and that maymake the measurement question much more consequential. In theprior episodes, the rapidly improving technology was acomplementto human labor at the aggregate level: better chips made workersand equipment more productive; free digital services raised These challenges are with us now in modest form, butthe case for confronting them rests on a forward-lookingobservation: statistical infrastructure takes years to build, and once we develop direct estimates of US AI production by combiningseveral data streams: data-center electricity usage and chip-stockcharacteristics, prevailing GPU rental rates, AI inference prices Nominal compute spending.US AI compute spending—measuredon an imputed-rental basis using prevailing graphics processing unit(GPU) rates—rose from $37 billion in 2023 to $90 billion in 2024 Raw compute capacity.As chips became more efficient, each dollarof compute spending bought more physical computing capacity.Measured in H100-equivalent units, US AI computing capacity grew Quality-adjusted AI output.An even larger growth rate becomesvisible once one accounts for algorithmic progress. Inference pricesat fixed benchmark performance fell by roughly 94 percent per yearover our sample, and Ho et al. (2024) estimate that the compute What “quality-adjusted” means here Quality-adjusted inference output measures the AI servicesdelivered to users—chiefly inference tokens—adjusted so that growthreflects both more tokens produced and higher capability per token.Quality-adjusted training output measures investment in new AI Two contributions, two ambitions.It is useful to be explicit abouthow these measures relate to standard national accounting first contribution sitswithinthe existing conceptual framework ofnational accounts. Tracking AI as a coherent sector—by aggregatingnominal compute spending, raw compute capacity, and quality-adjusted output across the dispersed industry codes that AI activityfalls under—requires no change to headline GDP methodology. Itis the same kind of work that produces tradable-sector or energy- The second contribution is more ambitious and morespeculative. To put the production estimates in context relativeto the overall economy, theWorking Paperdevelops a frameworkthat treats the AI sector as a coherent quasi-economic entity inits own right—partitioning value creation according to whetherit is more closely associated with AI computation or with human Implications for total GDP.As an illustrative exercise, one canask what aggregate US GDP growth would have looked like if theAI portion of the economy had been deflated using our quality- come in higher by roughly 2 percentage points in 2024 and roughly4 percentage points in 2025. We present this as an upper boundrather than a point estimate, for two reasons. First, AI inference isoverwhelmingly an intermediate input, not a final good, and theproductio