The new Chinese LLM, DeepSeek, has caused a US AI stock market crash. The best analogy I read, elsewhere on Lemmy that icba to dig up right now, is that while Nvidia has been insistent that you need their Rolls Royce or you have to walk, a Chinese company has released an affordable family car.
That analogy is flawed to the point I’d say it’s completely incorrect. To try and save it, I would go with …
A Chinese company has released a free car into a market full of free cars, but their car is the 2025 model so everyone wants it as its new. It won’t be new for long, and everyone will want a different model soon.
Nvidia targets businesses with their products, consumers having free cars isn’t a big issue for them as companies will still need their trucks.
Nvidia stockholders think the sky is falling and are pulling out, causing them to think the sky is falling, causing them to pull out. The real threat here isn’t DeepSeek, it’s that stockholders start to see AI doesn’t actually offer all the benefits that have been promised to companies looking to cut cost.
Edit: Oh and nobody is running the actual real 720GB, Deepseek R 671b model that can beat GPT, without using very high end expensive Nvidia cards. The stuff people are running on their machines at home is like a go-kart compared to the car.
Eh, analogy will be imperfect due to nuance, but I’d say it is close.
The big deals are:
DeepSeek isn’t one of the “presumed winners” that investors had been betting on, and they caught up very quickly
DeepSeek let people download the model, meaning others can host it free and clear. The investors largely assumed at least folks would all abide by the ‘keep our model private if it is competitive and only allow access as a service offering’, and this really fouls up assumptions that an AI provider would hold lock-in
DeepSeek is pricing way way lower than OpenAI.
Purportedly they didn’t need to push their luck with just tons of H100 to get where they are. You are right that you still need pretty beefy to run it, but nVidia’s stock was predicated on even bigger stakes. Reportedly an attempt to train a model by OpenAI involved $500 million, and a claim to train a “good enough” for less than $10 million dramatically reduces the value of nVidia. Note that why they are “way down” they still have almost a 3 trillion dollar market cap. That’s still over 30 Intels or 12 AMDs. There’s just some pessimism because OpenAI and Anthropic either directly or indirectly drove potentially a majority of nVidia revenue, and there’s a lot more uncertainty about those companies now.
I also think this is on the back of a fairly long relatively stagnant run. After the folks saw the leap from GPT2 to ChatGPT they assumed a future of similar dramatic leaps, but have instead gotten increasingly modest refinements. So against a backdrop of a more “meh” sentiment over where they are going you have this thing to disturb some presumed fundamentals in the popular opinion.
The new Chinese LLM, DeepSeek, has caused a US AI stock market crash. The best analogy I read, elsewhere on Lemmy that icba to dig up right now, is that while Nvidia has been insistent that you need their Rolls Royce or you have to walk, a Chinese company has released an affordable family car.
That analogy is flawed to the point I’d say it’s completely incorrect. To try and save it, I would go with …
A Chinese company has released a free car into a market full of free cars, but their car is the 2025 model so everyone wants it as its new. It won’t be new for long, and everyone will want a different model soon.
Nvidia targets businesses with their products, consumers having free cars isn’t a big issue for them as companies will still need their trucks.
Nvidia stockholders think the sky is falling and are pulling out, causing them to think the sky is falling, causing them to pull out. The real threat here isn’t DeepSeek, it’s that stockholders start to see AI doesn’t actually offer all the benefits that have been promised to companies looking to cut cost.
Edit: Oh and nobody is running the actual real 720GB, Deepseek R 671b model that can beat GPT, without using very high end expensive Nvidia cards. The stuff people are running on their machines at home is like a go-kart compared to the car.
Eh, analogy will be imperfect due to nuance, but I’d say it is close.
The big deals are:
I also think this is on the back of a fairly long relatively stagnant run. After the folks saw the leap from GPT2 to ChatGPT they assumed a future of similar dramatic leaps, but have instead gotten increasingly modest refinements. So against a backdrop of a more “meh” sentiment over where they are going you have this thing to disturb some presumed fundamentals in the popular opinion.
DeepSeek runs so much faster than gpt4
No idea when its feed data cutoff is, it like all others isn’t allowed to look at the live internet
Does is use up fresh water faster than GPT4 too? Because no one seems to be talking about that. They barely talk about it with American LLMs.
It seems to need an order of magnitude less compute, so probably. It’s also open source