AI companies are maxing out tokens just to squeeze more out of models

An Elastic research paper shows how much extra tokens models are gulping down as the industry chases better answers. The pattern is real and the cost is real: companies are sending 3–4× the tokens a model needs for the actual work. That bloat isn't free. At scale it can be the difference between a model fitting in memory and needing a second GPU.

The trick is tokenmaxxing — the kind where you pad the prompt with extra context, chain multiple reasoning steps, or send the same question twice to different models and compare. It often improves accuracy, but it also means you're paying for words that don't need to be there. The paper tracks this across several models and finds the overhead is nowhere near uniform. Some models eat tokens like they're free. Others choke on it.

Why this matters for us: if you're running models on anything but a data center — on a VPS, in a small team, or for a local business — token waste is a real line item. The fix is usually trimming prompts, not buying more GPU.