In the Token Flow

The core idea

Brad Gerstner organizes AI-era investing around a single question: does a company's usage go up or down as computational intelligence gets cheaper and more abundant. Some software companies, Databricks, Snowflake, and Confluent among them, the ones running the database queries, storage, and pipelines that agents exercise more heavily as they run, are clearly in the token flow, since as token consumption rises, the volume of database queries and storage demand rises alongside it.¹ The metaphor is hydraulic: token consumption is a rising current, and the question for any business is whether it sits in that current, where volume pulls it along, or stands across it, where it gets eroded.

Companies in the flow, the enablers, include raw compute and the memory and logic providers that produce tokens, Nvidia foremost among them, plus the data and query layer that gets exercised more as agents run. More tokens mean more queries, more storage, and more pipeline activity, which means more revenue, and these companies earn an above-market multiple as a result. Companies across the flow, the competitors, are application and front-end software whose product surface competes directly with what a model itself can now do; these derate. Gerstner's example is Salesforce, whose front-facing solutions he says compete with the models even as Snowflake enables them, though he leaves the door open for Salesforce to move into the flow if it manages to embed itself as an enabler rather than a competitor.¹

Why it matters

The frame converts a vague worry about software being disrupted by AI into a single testable criterion: does this company's usage go up or down when computational intelligence gets cheaper. If a company benefits from rising intelligence consumption, it earns a premium; if every improvement in computational intelligence makes a company's own product worse off by comparison, it will trade below the market multiple, with more room to fall.¹

This is the bullish, infrastructure-side mirror of the view that durable value sits at the application product layer because raw model access is a one-line swap between providers. The two views are reconcilable: both identify the front-facing application that competes directly with the model as the squeezed middle, and differ mainly on whether the safe ground is the product layer above the model or the infrastructure and data layer beneath it.

Tensions

The line is fuzzy at the model layer itself: a frontier lab is simultaneously the largest token producer in the system and an application, such as a consumer chat product, that competes with everyone else, so whether a frontier lab counts as in the flow or across it depends on whether you are scoring the model or the product built on top of it. The framework also precisely flatters the holdings of the investor articulating it, so while it is directionally close to mainstream thinking, it is not disinterested. And the door is explicitly left open for a company currently classified as a competitor to re-enter the flow if it successfully wraps proprietary data and workflow around a model, which would turn the in-flow-versus-across-it binary into more of a spectrum than a fixed classification.¹