The Research
A team of researchers analyzed 380 trillion tokens of realized AI consumption across more than 400 large language models, using a proprietary licensed dataset from OpenRouter covering approximately 2% of global monthly AI token consumption. The core finding: a value-weighted long-short strategy based on firm AI consumption intensity earns 64.1 basis points per week — roughly 33% annualized — with the premium concentrated in firms using frontier closed-source models, paying and seasoned users, and long-context prompts.
The detail about long prompts matters. Short queries to commodity AI APIs are background noise in this dataset. The signal concentrates in usage patterns that suggest substantive operational deployment: complex analytical tasks, extended reasoning chains, and sophisticated multi-turn workflows. The factor is not measuring which firms mention AI in earnings calls; it measures which firms are actually running AI at depth in their operations.
What Makes This a Meaningful Factor
Equity factors derived from alternative data tend to decay quickly once discovered, because the signal gets arbed away. Token consumption data from a platform like OpenRouter is different in two ways. First, it's not public information — the dataset required a proprietary license, which limits how quickly the factor can be exploited at scale. Second, the signal reflects operational reality: if a firm is running substantial frontier AI workloads, that spending shows up in token consumption before it shows up in revenue or productivity metrics. There's a natural lag between AI adoption and financial performance that token consumption data could help close.
The 2% coverage caveat is important. OpenRouter's data captures roughly 2% of global monthly AI token consumption. The 64.1 bps figure derives from that sample. As coverage expands through additional licensing agreements or comparable datasets from other aggregators, the statistical confidence around the factor will improve. The current number is a floor estimate, not a ceiling.
The Implications for Capital Markets Builders
If AI consumption intensity is a real equity signal, the firms closest to that data have an information advantage. Token consumption from cloud providers, API usage from AI infrastructure platforms, and enterprise procurement data from AI software vendors all carry similar information. Most of it is currently siloed, unstructured, and unavailable in machine-readable form at the latency investment managers need.
Alternative data providers who can clean, normalize, and license AI consumption metrics at the firm level have a natural product opening here. For quant teams, this is a new category to evaluate — one that hasn't yet been competed down to near-zero because the sourcing infrastructure doesn't broadly exist. The OpenRouter study is the clearest published signal that there's a premium to find before that window closes.
The Factor Decay Risk
Every new equity factor follows the same arc: discovered, published, traded by early movers, and eventually arbed into marginal statistical significance. The AI consumption factor will likely follow this pattern. What extends the window here is the sourcing constraint. Many alternative data factors — satellite imagery, credit card transaction aggregates — are available from multiple vendors simultaneously, which accelerates crowding. AI consumption data at firm-level resolution is currently narrow, which buys more time before the factor becomes broadly traded.
The concentration in frontier closed-source model usage also adds a layer of durability. Consumption of commodity open-source models won't carry the same signal — it's too cheap and too widespread. The premium tracks firms making intentional investments in expensive frontier capabilities, which is a smaller and more observable population.
What to Watch
The follow-on questions that matter: whether the factor holds for smaller-cap firms where signal-to-noise is worse; whether the premium tracks actual AI-driven productivity improvement or primarily reflects investor sentiment around AI commitment; and which alternative data vendors can replicate the OpenRouter coverage at greater breadth. The researchers' finding that long prompts drive more of the premium than short ones suggests the factor measures operational depth — a potential leading indicator for AI-driven productivity gains before they appear in financial statements. That's the thread worth following as the coverage dataset grows.