My guess is: The assumption has been that sufficient revenue from the SOTA GenAI models will only be generated by making them more capable. That requires training new base models, which in turn is quite expensive.
But if that can be done more efficiently, then what justifies the valuations...
Oh and I guess there's the openness of the methods (publishing what didn't work in addition to what did) by DeepSeek. Which is only tangential, but might have some effect on people's thinking.
Yes, I wonder if this means new entrants with alternative data sets and more experimentation, which would not lead to a net decrease in demand for hardware but potentially increase it?
I think this is a good bet. Seems like the Transformer-based chatbots alone will have some staying power, even if it's in low level forms (apps for homework help, image generation, etc). People use them. Revenue is a problem for the biggest firms though.
i think the fear is that if efficiency gains like deepseek's continue to happen then the demand curve is longer than previously anticipated. which could mean the growth priced into nvidia's valuation as of yesterday will take longer to achieve. i don't rlly think this tho.
The endgame for LLMs (most/all "AI" systems) is small purpose-specific models. Think of things on your phone that extract text, add alt text for people with disabilities, or language translators. These models would be composable agents.
Task oriented efficient LLMs are difficult/expsive to iterate.
The "big players" are in a race to iterate on the most well performing model. Creating efficiencies in the process allows more people to participate even with commodity hardware. Open source community and AI/LLM communities have been doing this for years with low traction because of cost/difficulty
If you were really inclined, you can now train something like Novasky or Deepseek with an AMD gaming PC instead of a whole array of specifically NVIDIA H100 specialized hardware. Your capital for an AI startup producing models is slashed to 1/3 of what it was before.
This gets to the gap in my knowledge. My initial reaction is that efficiency would lead to more people being able to train, which would maybe lead to more demand. But your point is that Nvidia competitors can now provide sufficient hardware? Is compute still going to be useful?
Not only Nvidia competitors. Competitors across the whole AI landscape from model trainers to hardware companies, cloud providers, etc, etc, can all purchase/provide fewer and cheaper resources to accomplish the same thing.
Why rent a truck from Peterbilt when you only need a bicycle for groceries?
Stock prices weren't based on real demand. They were based on a complete fantasy of large expensive-to-train and expensive-to-run models creating unprecedented value through AGI.
Making a mediocre LLM that's cheaper to train and much cheaper to run reduces the whole dream to just boring IT
it may not reduce demand directly, but it does decentralize it. certainly undermines the calls for 500billion dollar super computers built around power plants.
tbh i'm not sure how that shakes out. a lot of demand rn is from anticipating that next-gen models will be significantly larger and more compute-intensive. if that changes, a lot of organizations might reassess and conclude that they already have the infrastructure they need.
Comments
But if that can be done more efficiently, then what justifies the valuations...
This is just a guess tho. I have my opinions about all this, but the reaction has been kind of irrational.
https://en.wikipedia.org/wiki/Induced_demand
The same way, people think that increasing road capacity will reduce congestion (on the contrary).
Task oriented efficient LLMs are difficult/expsive to iterate.
Why rent a truck from Peterbilt when you only need a bicycle for groceries?
Making a mediocre LLM that's cheaper to train and much cheaper to run reduces the whole dream to just boring IT
I believe it’s an impulsive panic reaction because investors are afraid AF to lose money, and two few of them have a functioning brain.
I believe the markets will correct in the long run—reduced AI compute is a great opportunity for innovation.
https://bsky.app/profile/volkan.io/post/3lgtp55rpb22n