AI token demand could reach 4,000 quadrillion tokens annually by 2030.
That is Evercore’s base case—more than 20× today’s consumption. Goldman Sachs projects a similar 24× increase. The biggest driver? Agentic AI. AI agents could become the largest source of token demand by 2030. They operate continuously—and a single agentic task can consume 10–100× more tokens than a chatbot response. But the forecast excludes one potentially enormous category: physical AI—robots, autonomous machines and intelligent industrial systems. Every token ultimately requires physical infrastructure: 🔹 GPUs 🔹 Memory 🔹 Networking 🔹 Electricity 🔹 Data centres That infrastructure largely does not exist yet. AI-related stocks will experience corrections. But forecasts for underlying demand continue to be revised higher—not lower. The AI infrastructure cycle may have much further to run than markets currently assume. Source: Kyle Reidhead | Milk Road
AI Agents are using 1400% more tokens today than February, 500% more than human.
Remember, this is inference token traffic. Every token lands in the KV cache which is a memory problem. Will supply be able to catch up? Source: Trade Whisperer
