Post #3351505

2026-06-08 19:16 UTC

The worst (best) thing I can imagine is that they have some sort of mathematical proof that inference costs won't drop or that training costs will. Their only path to profitability was a moat in the form of high cost to train a model and dirt cheap inference they can resell at a high price. If they finally have some idea which way it will fall then there's nothing they can do.

Replies (2)

@tinsuke@lemmy.world 2026-06-08 20:30

I think that this would be too rational and scientific to cause an irrational market to correct. I'm leaning more towards a money related confirmation. Maybe an OpenAI or Anthropic internal shared numbers. Or, one of [the four horsemen of the AIpocalypse](https://www.wheresyoured.at/four-horsemen-of-the-aipocalypse/)?

Open ##3351613
@NotMyOldRedditName@lemmy.world 2026-06-08 20:25

There's one company trying to etch models into the silicon, they got a working prototype in a 8b model and it does 17k t/s. They suggest at data center scales with bigger models it'd pay for itself in a year in saved costs in electricity, cooling, space required etc. You can apply lora's to it, but it will fall behind as it ages, im not quite sure how long a static model would last to make up the money after that 1st year, but for something like creative writing where it doesnt need current data (e.g programming but the languages are always changing and evolving is a problem) it might work?

Open ##3351614