During the Industrial Revolution, the cheaper the coal, the more everyone burned. Now, in the AI era, tokens are the same way.

When tokens become cheaper, AI products can burn even more.
In the past, you asked a question, the model answered, and that was the end.
Now, you click once, and an agent breaks down tasks, searches for information, adjusts tools, writes code, fixes bugs, summarizes, running a whole set of processes.
So, individual tokens are cheaper, but the tokens consumed per task are much higher.
That’s why the bills keep rising.
After steam engines became more fuel-efficient, the UK didn’t burn less coal. Because coal became more cost-effective, more machines, factories, and railways started burning coal.
Tokens follow this same logic.
After becoming cheaper, agents, deep research, AI programming, long contexts, and enterprise automation really took off.
So, the cost center of AI is shifting from training to inference.
Training burns for a while, but inference burns continuously.
When users are online, it’s burning.
When agents are running, it’s burning.
The longer the context, the larger the cache, and the more memory, bandwidth, electricity, and heat are consumed.
That’s also why the AI supply chain can’t just focus on GPUs.
HBM, DRAM, SSDs, advanced packaging, optical modules, switching chips, CPUs, inference chips—all will be repriced due to this wave of inference demand.
AI application companies will also be forced to layer their services.
Companies that only provide a UI layer and rely entirely on closed-source APIs—
the more users they have, the bigger the bills, and the thinner the profit margins.
True barriers will go deeper: routing, quantization, caching, batch processing, context trimming, small models replacing large models.
Yeah, tokens are like coal.