Market behaviour and pricing strategies for consumer GPUs
- Thread starter Seanspeed
- Start date
Well Nvidia' stock price rises based on low information emotion so it makes sense that it's falling for the same reason. I'm sure it'll go back up when people realize the Deepseek news is a nothing burger.
Is it really a nothingburger that a random Chinese firm can produce a working LLM model for $5MM instead of tens of billions?
Did they have the same starting point? Thought I read somewhere they used data that was already processed by another LLM.
arandomguy
Veteran
The bigger downstream risk from a market perspective I'd be worried about is if this increases the trade war surrounding AI.
But in terms of the AI bubble popping I think there needs to be some context when throwing that around in terms of what people actually mean. Nvidia's valuation, and this growth tech companies in general, are based essentially on future growth. Nvidia's actual AI sales/revenue can still continue to grow and their stock could plummet as long as that growth rate is just perceived to be lower. So the AI bubble popping from a company valuation standpoint is very different than the AI bubble popping from the product stand point.
But in terms of the AI bubble popping I think there needs to be some context when throwing that around in terms of what people actually mean. Nvidia's valuation, and this growth tech companies in general, are based essentially on future growth. Nvidia's actual AI sales/revenue can still continue to grow and their stock could plummet as long as that growth rate is just perceived to be lower. So the AI bubble popping from a company valuation standpoint is very different than the AI bubble popping from the product stand point.
Does it matter? All the LLMs are (at least partly) trained with stolen data, why not steal other LLMs data too?
Nvidia will tank hard if/when Google, Meta, Tesla, Microsoft, Oracle announce they’re no longer building massive AI data centers. That’s what’s fueling the Nvidia hype train right now. It’s slim odds that a small rando firm knows something that all those guys don’t.
Did they “solve AI” in general or did they share non-repeatable results from a niche experiment? It matters.
DegustatoR
Legend
It does matter as you need to get the dataset for "cheap" training from "expensive" training.
Also this dataset is still running on the same h/w so these GPUs will still sell. The fact that it can be done cheaper means that more people will be doing it and in the end the overall market volume may stay the same or even increase.
Stock prices are basically nothing when we're talking about a company's operations.
The market rarely cares if they came from the same starting point. If they can do the same thing but for 3% of the cost the market will correct, regardless of if they are riding on coattails.
In fact this happens again and again in tech, everyone thinks the moat is unbreakable until newcomers come along and use the work you've done to break your moat.
You can download DeepSeek today and see for yourself if it's a non-repeatable experiment or a functional LLM made with 3% of the budget.
"AI" in general is full of scam artists and basically speculative hype people, so reading the news today and trying to understand the implications are impossible. So far I've heard that deepseek was likely trained with embargoed hardware. Some people say they've really created some new way of training. Some people are saying the industry is a mess and deepseek basically just did good engineering and implemented the best research that's available, where the big companies are just throwing money at the problems and aren't carefully optimizing. No idea what the real story is.
arandomguy
Veteran
But based on the logic above is that using the same reasoning the market would've stalled already as companies like OpenAI would not have needed to iterate further as they already themselves had a "good enough" model without any future training. Which is why I'm not sure why this new development from DeepSeek seems to be driving the idea of a broader scale back of hardware aquisitions for AI.
At least personally, as I mentioned above, I feel a bigger concern is that because it is a chinese company this might lead to more trade hostility between the US and China in terms of the broader impact. With that said I hope the gaming/consumer crowd may want to take step back and not feel this the "win" they might think it is due to the broader impacts of escalation on that front in terms of how it would likely also affect gaming/electronic consumer hardware.
Also from the IHV angle I just want note that of the now 3 GPU vendors all 3 saw major drops today but only 1 is actually in the green in terms of valuation over 6 months or even a 1 year, and I'm referring to the change in valuation not the companies colors.
At least personally, as I mentioned above, I feel a bigger concern is that because it is a chinese company this might lead to more trade hostility between the US and China in terms of the broader impact. With that said I hope the gaming/consumer crowd may want to take step back and not feel this the "win" they might think it is due to the broader impacts of escalation on that front in terms of how it would likely also affect gaming/electronic consumer hardware.
Also from the IHV angle I just want note that of the now 3 GPU vendors all 3 saw major drops today but only 1 is actually in the green in terms of valuation over 6 months or even a 1 year, and I'm referring to the change in valuation not the companies colors.
Last edited:
Not sure what you mean. Do you believe that all workloads have already been encoded as neural networks and therefore we’re now in the second phase of the revolution that can run on a fraction of the hardware? Can Deepseek repeat the outcome with other workloads e.g. protein folding or weather simulation? Think I saw somewhere that it doesn’t have ChatGPT’s conversational chops. No idea what to believe as there’s an immense amount of FUD being tossed around right now.
I doubt downloading it will tell us anything about the repeatability of their approach to diverse ML workloads.
How on earth did you get that from what I said at all?
DeepSeek is an LLM, why would it be used for protein folding and weather simulation?
(as an aside I am skeptical American companies can successfully use AI for either of those in any meaningful way)
Ok I’m trying to follow your point. If deepseek is just an LLM how is it a harbinger of the end of the AI arms race? Aren’t we just skimming the surface of ML workloads?
If Americans or anybody else don’t figure out other use cases besides building LLMs on historical data this AI revolution will be short lived.
DavidGraham
Veteran
Sam Altman is doubling down on compute.
Perplexity too.
Perplexity too.
Last edited:
DavidGraham
Veteran
This is the dumbest number I have seen in this whole fiasco. 5 million is maybe maybe the cost of the final training run. It in no way encompasses the cost of the 50k hopper cluster they set up, or the cost of research, development, data collection, or the cost of the previous training runs.
Quote me where I called this the “harbinger of the end”.
I don’t think we’re skimming the surface of anything because I think most of this tech is nonsense and most of the leaders in this space are bullshit artists, like Saltman.
Sure, 5MM probably isn’t all of it, but I think it’s nowhere near the billions used by their American counterparts.
ChatGPT and other large LLMs are like Ray Tracing.
DeepSeek is baked lighting.
Deepseek could never be created, without the ground truth tool, which are the larger LLMs. That also means that while Deepseek can run much faster and cheaper than the larger LLMs, its likely unable to ever produce a better result than the LLMs.
For most use cases this might be sufficient, but for others like asking it to come up with solutions to PHD level problems, deepseek may not be useful at all.
DeepSeek is baked lighting.
Deepseek could never be created, without the ground truth tool, which are the larger LLMs. That also means that while Deepseek can run much faster and cheaper than the larger LLMs, its likely unable to ever produce a better result than the LLMs.
For most use cases this might be sufficient, but for others like asking it to come up with solutions to PHD level problems, deepseek may not be useful at all.
Something that is 95% as good for a fraction of the cost is basically catnip for investors.
Who is having LLMs solve PHD level problems? These models can’t think, they regurgitate information they’ve been trained on.