NBER paper pegs AI-beta stock premium at 64.1 bps a week

A new NBER working paper does something the AI-exposure literature had mostly done by proxy: it tries to measure which firms are actually AI-exposed by watching what their stocks do, and then asks whether that exposure earns a return. The size of the answer is what makes it interesting.

Nicola Borri, Yukun Liu, and Aleh Tsyvinski build an "AI Factor" from roughly 380 trillion tokens of realized consumption across more than four hundred large language models, drawing on a licensed OpenRouter dataset they estimate covers about 2 percent of current global monthly AI token consumption. From that factor they estimate firm-level AI Betas, meaning how much a company's returns comove with AI usage, and sort stocks on that beta. According to the working paper on NBER, a value-weighted long-short strategy earns 64.1 basis points per week. That is a very loud number for a documented equity factor, and the authors treat it as such.

The more careful finding is that the premium is heterogeneous. It concentrates on what the paper calls the frontier-oriented margin of AI use, meaning closed-source models, paying and seasoned users, and long prompts, while casual or open-weight consumption barely registers. And it does not stay inside tech. The premium reportedly extends into consumer-facing and capital-heavy parts of the economy, but is absent in emerging markets, including China. On the labor side, an occupation one standard deviation higher in interaction-and-communication content has 0.36-standard-deviation higher market-implied AI exposure, while analytical, scientific, and operations-control skills show negative exposure.

The honest caveat is that this is one working paper on one proprietary dataset covering a small slice of global AI usage, and enterprise consumption behind private endpoints is largely invisible to it. Weekly return premiums this large also tend to shrink out of sample once a strategy is publicly documented, and the paper is a snapshot rather than a long-horizon test. What the reporting doesn't give you is the causal channel, whether high-AI-beta names are earning more because of realized revenue, cost savings, or investor expectations, or how stable the ranking is across regimes. Take the specifics as reported, not settled.

The forward-looking piece the authors flag, and the one worth watching if you allocate capital or hire, is that the premium tracks frontier, paid, closed-source use, not "does the company mention AI." A pre-print version is also up on arXiv. If the result holds up, an AI Beta becomes a real portfolio input rather than a marketing story, and the paper's line about "the rise of the agentic economy" starts to look less like a phrase and more like something a factor sheet has to price.