From the stage at the GPU Technology Conference 2026, Jensen Huang delivered his most structured view yet of the AI economy. The Nvidia CEO wasn’t unveiling a mere product; he presented an entirely new framework: the central unit of value is no longer software, nor even the GPU—it’s the token.
Consumption, productivity, performance indicators, and compensation packages; all of it is now organized, he argues, around this digital fragment.
The Data Center Transformed into a “Token Factory”
At GTC 2026 in San Jose, Jensen Huang formalized a concept he has been developing for several quarters: the AI Factory. In this vision, the traditional data center fades away. Production is no longer measured by raw computation, but by the volume of tokens generated by AI agents and models. The token becomes the proxy for an organization’s digital productivity, and perhaps even that of a country.
This approach dramatically restructures the management of infrastructure. The key metric becomes token-per-watt, and every watt of available capacity that does not generate tokens is now counted as lost revenue. As for the scale of ambition, Huang quantified it: investing $100 billion in AI supercenters reaching 1 GW could yield up to $150 billion in annual revenue from token sales.
The Token, the “Bit” of the Generative Era
If the bit was the fundamental unit of classical computing, the token is the unit of generative AI. Already used to fragment text and bill models like ChatGPT or Claude, it takes on a radically new dimension with the emergence of agentive AIs.
Contrary to a standard prompt, which consumes a limited volume of tokens, an autonomous agent operating continuously can spend up to a million times more. This mechanical surge in demand elevates the token to a strategic commodity: the more a company consumes, the more it automates and personalizes its services.
From Tool Rentals to Agent Rentals
The software economy is poised for a structural shift. Today, companies are renters of tools—SaaS, licenses, subscriptions. Tomorrow, they will rent agents that use these tools and consume tokens on their behalf.
For CIOs, this translates into a new line in the budget. Alongside CPU/GPU allocations and SaaS licenses, there will now be dedicated token envelopes for each business function or family of agents to manage. This market, incorporating the value of agents and their flows, could, according to Huang, far surpass the current software market, which runs into trillions of dollars.
Token Budgets as a Recruitment Lever
One of the most unexpected proposals revealed at the GTC 2026 concerns human resources. Huang envisions individual token budgets for engineers, integrated directly into compensation packages.
The argument is economic as much as managerial: an engineer with access to massive computing resources can see productivity multiply tenfold. Allocating $100 to $1,000 in daily inference costs to a standout profile becomes a profitable investment once the agents they pilot create value. The talent market appears already to be taking note of this signal: candidates would begin asking employers about access to available computing resources, making such access a genuine attractor in its own right.
The Token-Cost War
In the face of massive infrastructure investments, Nvidia answers with a simple profitability equation: more compute generates more tokens, which translate into revenue via tiered pricing models—free access for user acquisition, then monetization by volume.
According to Jensen Huang, the ecosystem’s success will hinge on two levers that Nvidia intends to lock in structurally: the cost per token generated on one hand, and the tokens-per-watt ratio on the other.
By positioning its GPUs as the most efficient machines for producing cheaper and more numerous tokens, Nvidia is no longer selling merely hardware. It places itself at the heart of an economy whose unit of measurement it has itself defined, and perhaps that is the most enduring ambition laid out in San Jose this year.
