AI Premium: stock markets are pricing in AI gains before companies earn them
Companies linked to AI are outperforming the stock market by 0.64% every week — not because they've boosted earnings yet, but because investors expect them to. That finding comes from a Yale AI Premium study covering 380 trillion AI tokens processed through the OpenRouter platform between January 2024 and April 2026. For anyone with a pension, index fund, or tech-heavy portfolio, this premium is already baked into valuations you're holding.
The data behind the number
The researchers tracked real AI consumption — tokens processed by large language models — rather than relying on company self-reporting or press releases. That distinction matters. The "AI Premium" reflects what investors believe will happen to margins and productivity, not what's appeared on any quarterly earnings report. Firms with large physical assets — durable goods manufacturers, capital-intensive industrials — capture the biggest share of the premium, because automation promises the sharpest efficiency gains there.
The premium holds clearly in the US and Europe. China and emerging markets show a weaker correlation between AI usage and stock valuations, per the NBER working paper that underpins the study.
The Yale study tracked 380 trillion AI tokens across the OpenRouter platform to measure real-world AI consumption by businesses.
The agentic shift changes the math
How AI is being used has shifted dramatically. In early 2024, most tokens came from people chatting with bots. By April 2026, agentic AI — systems that execute multi-step tasks autonomously, without constant human oversight — accounted for 52.2% of all tokens processed. That's a structural change, not a trend. Enterprises are deploying AI to run workflows, not just answer questions.
That shift has a cost dimension worth watching. Agentic systems consume tokens at scale, and early enterprise deployments have exposed runaway spend patterns. If margin compression from AI infrastructure costs accelerates faster than productivity gains materialise, the premium could deflate — markets are pricing in the upside but may be underweighting the operating costs.
What the premium ignores
The study is bullish on expectations but silent on friction. In the US, SEC scrutiny of "AI washing" — companies exaggerating AI involvement in disclosures — adds a liability layer that isn't priced cleanly into valuations. The Yale data also doesn't capture the divergence between US manufacturing job forecasts (around 2 million roles at risk by 2026) and ECB data showing AI-intensive European firms are net hirers, not net firers.
The market's 0.64% weekly bet on AI is real and measurable. Whether it survives contact with actual earnings cycles — and regulatory headwinds on both sides of the Atlantic — is the question no token count can answer yet.