Evercore presents a new model for AI token consumption, and the numbers hint at a massive explosion in demand for data centers and AI infrastructure.
Investment bank Evercore today published a comprehensive model that attempts to estimate how many tokens AI systems will consume by the end of the decade, and what that means for the data center, server, network, and electricity industries.
To understand the magnitude of the event, one must first understand what a 'token' is. A token is a small unit of information that an AI model processes.
It can be a word, part of a word, a number, or a symbol.
Every query to ChatGPT, every autonomous agent, and every AI search engine constantly consume tokens.
The more users there are, more AI agents, and more complex uses, token consumption jumps exponentially.
Evercore's base scenario: The company estimates that annual token consumption will jump from approximately 100 quadrillion tokens in 2026 to approximately 4 quintillion tokens in 2030.
This represents an average annual growth rate of approximately 150%.
- According to the estimate:
- Consumer AI users will grow from approximately 1.5 billion to 3.7 billion users.
- Autonomous AI agents will grow from 50 million to 800 million.
Upside scenario: Evercore estimates the possibility of even more aggressive growth, 8.3 quintillion tokens in 2030, then the CAGR will reach approximately 170%.
- In this scenario:
- AI users reach 4.4 billion.
- The number of AI agents jumps to 1 billion.
Even the downside scenario remains very aggressive: 2 quintillion tokens in 2030, even in this case, the CAGR is still above 130%.
Meaning, even in a relatively weak scenario, the world still needs massive computing infrastructure.
The truly important point is, what does this mean for data centers?
Evercore attempted to translate token consumption into real infrastructure requirements.
According to their model: if the world reaches a consumption of 4 quintillion tokens in 2030,
then demand for data center capacity could reach approximately 250GW by the end of the decade!
A magnitude of 250GW is an enormous scale of infrastructure for electricity, cooling, servers, GPUs, communication networks, and fiber optics.
This essentially strengthens the thesis that the world is entering an era where AI becomes a foundational infrastructure, similar to electricity or the internet.
Who stands to benefit from this trend?
The value chain here is very broad: Chips and accelerators - $NVDA $AMD $AVGO HBM memory and DRAM memory - $MU $DRAM Network and communication equipment - $ANET $CSCO $NOK Optics and fibers - $LITE $AAOI $COHR Servers and cooling - $DELL $SMCI $VRT Electricity and energy infrastructure - $VRT $CEG $PWR $OKLO $BE %GEV Data centers and construction - $EQIX $DLR
One of the interesting points in the model is that growth does not come only from human users, but mainly from AI Agents.
An AI agent consumes many more tokens than a regular user because it:
Thinks in multiple steps Performs searches Communicates with other models Runs automations Operates continuously
Meaning, the transition from a world of 'Chatbots' to a world of 'Agents' may be the main driver of future demand.
The broader implication is that the market is beginning to understand that AI is not just a software story, but primarily a story of new infrastructure on a historical scale.