Tokens as Commodity

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI uses the token-as-commodity framing as a lens for reading the civilizational stakes of the AI transition. When Jensen Huang declares that tokens are commodities, he is not merely describing what data centers produce; he is prescribing how to think about what intelligence is. And if intelligence is a manufactured good with a cost curve, then the questions of meaning, intentionality, and understanding—the questions that the rest of the cycle grapples with at length—are offloaded to a downstream application layer that the infrastructure manufacturer need not address. The framing is most powerful for what it excludes.

The commodity metaphor has changed how trillions of dollars flow. Pension funds and sovereign wealth funds that could not invest in metaphysics can invest in a factory with a depreciation schedule. The data center that was once a cost center is, under the new framing, revenue-generating capacity. The accounting treatment is different. The willingness of long-duration capital to pour into the buildout is different. In this sense, the token-as-commodity reframing is one of the most consequential metaphors of the decade—not because it accurately describes what tokens are, but because it has shaped what gets built and at what scale. The capabilities the cycle documents—the twenty-fold productivity multiplier, the collapsed imagination-to-artifact ratio—exist at their current scale in part because this metaphor persuaded capital to fund the infrastructure.

Origin

Huang first introduced the data center as AI factory framing in the GTC keynote sequence beginning in 2023, and elevated the specific token formulation to a defining slogan at Computex 2025. The framing did not emerge from nowhere: it is a refinement of the picks-and-shovels logic Huang had articulated for years, applied with unusual precision to the question of what the product actually is. The token is the specific unit that made the factory metaphor tractable: it gave investors a product to depreciate, a throughput to measure, and a cost curve to project.

The historical antecedent Huang draws on, if implicitly, is the logic of commodity markets for natural resources. Oil is priced per barrel, copper per ton, electricity per kilowatt-hour. Each commodity has a defined grade, a known buyer, and a price set by visible markets. Huang’s move was to import this structure into a domain—language and reasoning—where it had never applied. The import is imperfect, as all such imports are: a commodity has a stable unit, but the number of tokens required to express the same idea varies by tokenizer, by language, by model architecture. A more efficient model produces more value per token. A reasoning model produces more value per token at the cost of more tokens per task. The unit is squishy in ways that physical commodities are not. This is not a flaw in the framing from NVIDIA’s perspective; it is a feature.

Key Ideas

The constitutive move. The sentence “these things are called tokens” is not descriptive but constitutive: it is not reporting what tokens already are but establishing what they will be treated as. By constituting tokens as commodity, Huang establishes the data center as factory, the GPU cluster as production line, and the scaling of compute as the industrial variable that determines all others. The constitutive move is more powerful than an argument because it restructures the frame within which subsequent arguments are evaluated.

What the framing leaves out. Mature commodities are fungible within a product grade: a barrel of Brent crude is interchangeable with any other barrel of Brent crude. Tokens are not. A token from a frontier reasoning model is not interchangeable with a token from a commodity-tier inference provider. The factory framing implies a stable production curve; the reality is a constantly reformulating measurement standard in which producers can redefine the unit and capture more value by doing so. The framing also leaves out that intelligence-as-commodity treats human cognitive labor as the substitution target, without explicitly saying so.

The financialization of intelligence. By making tokens measurable, priced, and depreciable, the framing made AI scaling legible to the full apparatus of industrial finance. This has consequences beyond NVIDIA: it has structured how hyperscalers, sovereign wealth funds, and national governments think about AI infrastructure investment. The three-to-five trillion dollar projection that Huang introduced at Davos 2026 as a conditional—“sensible only if the application layer generates matching revenue”—has been treated by most of its audience as a forecast. The move from conditional to forecast is the financialization of intelligence at work.