3 ms·

I really don't understand the market thinks Nvidia is losing its value. If DeepSeek reduce the required computational resources, we can pour more computational

by ezoe 2y ago

I really don't understand the market thinks Nvidia is losing its value. If DeepSeek reduce the required computational resources, we can pour more computational resources to improve it further. There's nothing bad about more resources.

tonyhart7 2y ago
- "I really don't understand the market thinks Nvidia is losing its value." because the less GPU need to train, the less money to be made - "If DeepSeek reduce the required computational resources, we can pour more computational resources to improve it further. There's nothing bad about more resources." thats why you are not hedgefund manager, these guys job is to ensure that the HYPETRAIN for company to buy as many nvidia gpu to sell no matter what, if we can produce comparable model without using B (as it stands billions of dollar), it means there are less billions of dollar to be made and the HYPETRAIN is near the end
skellington 2y ago
Markets aren't rational. NVidia is currently a hype stock which means LOTS of speculation, probably with lots of leverage. So, the people who have made large gains and/or are leveraged are highly incentivized to panic sell on any PERCEIVED bad news. It doesn't even matter if the bad news will materially impact sales. What matters is how the other gamblers will react to the news and getting in front of them. :)
- aunty_helen 2y ago
  Another thing, the markets had priced in X demand scaling @ ~145 and suddenly it's X/30 demand scaling and therefore the price should drop.
  - skellington 2y ago
    True, but current price isn't based on fundamentals, it's based on hype-value. nVidia is going to be a very volatile stock for years to come. I don't see deepseek changing nvidia's short term growth potential though. Efficiencies in training were always inevitable, but more GPU still equals smarter AI....probably.
  - Jlagreen 2y ago
    That's wrong. DeepSeek is a problem for Big Tech, not for Nvidia. Why? Imagine a small startup can do something better than Gemini or ChatGPT or Claude. So it can be disruptive. What can Big Tech do to avoid disruption? Buying every SINGLE GPU Nvidia produces! They have the money and they can use the GPUs in research. The worst nightmare of any Tech CEO is a startup which disrupts you so you have to either be faster or you kill access to needed infrastructure for the startup. Or even better, the startup has to rent your cloud infrastructure, this way you earn money and you have an eye on what's going on. Additionally, Hyperscalers only get 50-60% of Nvidia's supply. They all complain of being undersupplied yet they get only 60% and not 99% of Nvidia's supply. How come? Because Nvidia has a lot of other customers they like to supply to. That alone tells you how huge the demand is that Nvidia even has to delay Big Tech deliveries. Also the demand for Nvidia didn't drop. DeepSeek isn't a frontier model. It's a distilled model therefore the moment OpenAI, Meta or the others release a new frontier model, DeepSeek will become obsolete and will have to start again to optimize.
    - Tostino 2y ago
      You're wrong on deepseek. It's not distilled (actual r1 model is based off v3). So, so many misinformed takes in this thread (not just you...)
BoredomHeights 2y ago
The market might be right that Nvidia is overvalued, but if so I think only accidentally and not because of this news. Like you said, at least for now I think it's fairly clear that if a company has X resources and finds a way to do the same thing with half, instead of using less they'll just try to do twice as much. This could eventually changed but I don't think AI is anywhere near that point yet.
luxuryballs 2y ago
yeah I think people just got spooked and sold to take some profits they were going to take soon anyways
qqtt 2y ago
Well you have to keep in mind that Nvidia has a 3 trillion dollar valuation. That kind of heavy valuation comes with heavy expectations about future growth. Some of those assumptions about future Nvidia growth are their ability to maintain their heavy growth rates, for very far into the future. Training is a huge component of Nvidia's projected growth. Inference is actually much more competitive, but training is almost exclusively Nvidia's domain. If Deepseek's claims are true, that would represent a 10x reduction in cost for training for similar models (6 million for r1 vs 60 million for something like o1). It is absolutely not the case in ML that "there is nothing bad about more resources". There is something very bad - cost. And another bad thing - depreciation. And finally, another bad thing - the fact that new chips and approaches are coming out all the time, so if you are on older hardware you might be missing out. Training complex models for cheaper will allow companies to potentially re-allocate away from hardware into software (ie, hiring more engineering to build more models, instead of less engineers and more hardware to build less models). Finally, there is a giant elephant in the room that it is very unclear if throwing more resources at LLM training will net better results. There are diminishing returns in terms of return on investment in training, especially with LLM-style use cases. It is actually very non-obvious right now how pouring more compute specifically at training will result in better LLMs.
- msoad 2y ago
  My layman view is that more compute (more reasoning) will not solve harder problems. I'm using those models every day and when problem hits a certain complexity it will fail, no matter how much it "reasons"
  - johnfn 2y ago
    I think this is fairly easily debunked by o1, which is basically just 4o in a thinking for loop, and performs better on difficult tasks. Not a LOT better, mind you, but better enough to be measurable.
  - jes5199 2y ago
    I had a similar intuition for a long time, but I’ve watched the threshold of “certain complexity” move, and I’m no longer convinced that I know when it’s going to stop

by ezoe 2y ago

tonyhart7 2y ago

- "I really don't understand the market thinks Nvidia is losing its value." because the less GPU need to train, the less money to be made - "If DeepSeek reduce the required computational resources, we can pour more computational resources to improve it further. There's nothing bad about more resources." thats why you are not hedgefund manager, these guys job is to ensure that the HYPETRAIN for company to buy as many nvidia gpu to sell no matter what, if we can produce comparable model without using B (as it stands billions of dollar), it means there are less billions of dollar to be made and the HYPETRAIN is near the end

skellington 2y ago

Markets aren't rational. NVidia is currently a hype stock which means LOTS of speculation, probably with lots of leverage. So, the people who have made large gains and/or are leveraged are highly incentivized to panic sell on any PERCEIVED bad news. It doesn't even matter if the bad news will materially impact sales. What matters is how the other gamblers will react to the news and getting in front of them. :)

aunty_helen 2y ago
Another thing, the markets had priced in X demand scaling @ ~145 and suddenly it's X/30 demand scaling and therefore the price should drop.
- skellington 2y ago
  True, but current price isn't based on fundamentals, it's based on hype-value. nVidia is going to be a very volatile stock for years to come. I don't see deepseek changing nvidia's short term growth potential though. Efficiencies in training were always inevitable, but more GPU still equals smarter AI....probably.
- Jlagreen 2y ago
  That's wrong. DeepSeek is a problem for Big Tech, not for Nvidia. Why? Imagine a small startup can do something better than Gemini or ChatGPT or Claude. So it can be disruptive. What can Big Tech do to avoid disruption? Buying every SINGLE GPU Nvidia produces! They have the money and they can use the GPUs in research. The worst nightmare of any Tech CEO is a startup which disrupts you so you have to either be faster or you kill access to needed infrastructure for the startup. Or even better, the startup has to rent your cloud infrastructure, this way you earn money and you have an eye on what's going on. Additionally, Hyperscalers only get 50-60% of Nvidia's supply. They all complain of being undersupplied yet they get only 60% and not 99% of Nvidia's supply. How come? Because Nvidia has a lot of other customers they like to supply to. That alone tells you how huge the demand is that Nvidia even has to delay Big Tech deliveries. Also the demand for Nvidia didn't drop. DeepSeek isn't a frontier model. It's a distilled model therefore the moment OpenAI, Meta or the others release a new frontier model, DeepSeek will become obsolete and will have to start again to optimize.
  - Tostino 2y ago
    You're wrong on deepseek. It's not distilled (actual r1 model is based off v3). So, so many misinformed takes in this thread (not just you...)

BoredomHeights 2y ago

The market might be right that Nvidia is overvalued, but if so I think only accidentally and not because of this news. Like you said, at least for now I think it's fairly clear that if a company has X resources and finds a way to do the same thing with half, instead of using less they'll just try to do twice as much. This could eventually changed but I don't think AI is anywhere near that point yet.

luxuryballs 2y ago

yeah I think people just got spooked and sold to take some profits they were going to take soon anyways

qqtt 2y ago

Well you have to keep in mind that Nvidia has a 3 trillion dollar valuation. That kind of heavy valuation comes with heavy expectations about future growth. Some of those assumptions about future Nvidia growth are their ability to maintain their heavy growth rates, for very far into the future. Training is a huge component of Nvidia's projected growth. Inference is actually much more competitive, but training is almost exclusively Nvidia's domain. If Deepseek's claims are true, that would represent a 10x reduction in cost for training for similar models (6 million for r1 vs 60 million for something like o1). It is absolutely not the case in ML that "there is nothing bad about more resources". There is something very bad - cost. And another bad thing - depreciation. And finally, another bad thing - the fact that new chips and approaches are coming out all the time, so if you are on older hardware you might be missing out. Training complex models for cheaper will allow companies to potentially re-allocate away from hardware into software (ie, hiring more engineering to build more models, instead of less engineers and more hardware to build less models). Finally, there is a giant elephant in the room that it is very unclear if throwing more resources at LLM training will net better results. There are diminishing returns in terms of return on investment in training, especially with LLM-style use cases. It is actually very non-obvious right now how pouring more compute specifically at training will result in better LLMs.

msoad 2y ago
My layman view is that more compute (more reasoning) will not solve harder problems. I'm using those models every day and when problem hits a certain complexity it will fail, no matter how much it "reasons"
- johnfn 2y ago
  I think this is fairly easily debunked by o1, which is basically just 4o in a thinking for loop, and performs better on difficult tasks. Not a LOT better, mind you, but better enough to be measurable.
- jes5199 2y ago
  I had a similar intuition for a long time, but I’ve watched the threshold of “certain complexity” move, and I’m no longer convinced that I know when it’s going to stop