this post was submitted on 17 Jan 2026
90 points (100.0% liked)
technology
24173 readers
123 users here now
On the road to fully automated luxury gay space communism.
Spreading Linux propaganda since 2020
- Ways to run Microsoft/Adobe and more on Linux
- The Ultimate FOSS Guide For Android
- Great libre software on Windows
- Hey you, the lib still using Chrome. Read this post!
Rules:
- 1. Obviously abide by the sitewide code of conduct. Bigotry will be met with an immediate ban
- 2. This community is about technology. Offtopic is permitted as long as it is kept in the comment sections
- 3. Although this is not /c/libre, FOSS related posting is tolerated, and even welcome in the case of effort posts
- 4. We believe technology should be liberating. As such, avoid promoting proprietary and/or bourgeois technology
- 5. Explanatory posts to correct the potential mistakes a comrade made in a post of their own are allowed, as long as they remain respectful
- 6. No crypto (Bitcoin, NFT, etc.) speculation, unless it is purely informative and not too cringe
- 7. Absolutely no tech bro shit. If you have a good opinion of Silicon Valley billionaires please manifest yourself so we can ban you.
founded 5 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
The expensive part of LLMs is the training though. Actual token output is rather efficent and quite cheap. For example, for Deepseek to generate a 200 token paragraph of text it costs about $0.000084. Image generation is also rather expensive, but most of the data centers and cost are around training models, not serving LLM output. It still might be more expensive than advertisers are willing to pay, but not crazy expensive.
You'd think, but efficiency gains are erased by the LLMs having bigger context windows and self-referencing "thinking" or "agent" modes that massively extend token burn. There's public data out there showing how training costs are an enormous fixed point, but then inference costs very quickly catch up and exceed the training cost.
A model that's token-efficient is a model that's pretty useless and a model that's useable for anything is so inefficient as to have massively negative profit margins. If there was even one model out there that was cost effective for the number of tokens burned, the provider would never shut up about it to buyers
Wow, really? I guess context windows have been going up but did not realise they were so ruinously expensive. Where can I read more about this?