---
title: 'How Many Developers Actually Use That Million-Token Context Window?'
source: 'https://youtube.com/watch?v=pQMvO_ssdcA'
video_id: 'pQMvO_ssdcA'
date: 2026-08-10
duration_sec: 66
---

# How Many Developers Actually Use That Million-Token Context Window?

> Source: [How Many Developers Actually Use That Million-Token Context Window?](https://youtube.com/watch?v=pQMvO_ssdcA)

## Summary

The video discusses the practical use of million-token context windows in large language models, questioning why developers rarely utilize them despite their availability. It highlights issues like context rot, cost, and the mismatch between context window size and enterprise-scale data needs.

### Key Points

- **Million-token context windows are rarely used** [00:01] — Despite million-token context windows being available for two years since Gemini introduced them, most users stay under 200k tokens.
- **Context rot degrades output quality** [00:15] — The more context fed to an LLM, the lower the quality of output generally becomes, a phenomenon known as context rot.
- **Cost increases with context size** [00:29] — Users are charged for every context token, so voluntarily increasing context length makes the model more expensive.
- **Order of magnitude mismatch with enterprise data** [00:41] — Even a perfect million-token context window is insufficient for enterprise data, which can start at 8 trillion tokens, making 100x improvements irrelevant.

### Conclusion

The video concludes that million-token context windows are largely unused due to context rot, cost, and the vast scale of enterprise data, making them impractical for many real-world applications.

## Transcript

lengths of LLMs a lot further. Like and and my I obviously I might regret saying this, but we've had million token context windows for 2 years now since Gemini first introduced it, and no one uses it. Everyone stays under 200k. Is
of context rot and the fact that the more context you feed it generally the lower the quality output will become? So So yes, that that is that is up there uh also cost reasons, right? Like you get charged for every single context token
you voluntarily make it more expensive for yourself? Uh and then there's also just order of magnitude reasons, right? Like let's say context window today with zero context rot. It's perfect recall across all of
it, right? Absolutely solved. 1 million. Then I Then I come along and go, you million token context window. Uh so you just 100x and like also perfect context you know, uh amazing. Great. But I'm sitting over here. I'm in enterprise,
and my document uh database and storage and notes started at 8 trillion tokens. What is your 100x going to do for me? Nothing. [laughter] Absolutely nothing.