As of writing, NVIDIA stock still struggles below 3.0T, and its price change is again in negative territory.
I see many people below expressing concerns that the training cost reported by DeepSeek may turn out to be bogus, and that NVIDIA stock price may rebound in such a case.
But DeepSeek was not the only recent breakthrough regarding efficiency and economy. Just about a week before DeepSeek R1 was released, NovaSky, which is currently not a company but a collaborative initiative led by students and advisors at UC Berkeley’s Sky Computing Lab, released Sky-T1, a reasoning LLM on par with OpenAI's o1-preview (currently at #7 of the Chatbot Arena LB), with a total training cost of $450 (not a typo, $450, without being accompanied with "thousand" or anything else):
I can only guess that the main reason why Sky-T1 did not generate headlines (and financial panic) is that they did not release a web interface and/or app with a working model (hence the model did not enter any leaderboards); but I don't think that this will be the case for much longer.
Why do you think you're right?
As of writing, NVIDIA stock still struggles below 3.0T, and its price change is again in negative territory.
I see many people below expressing concerns that the training cost reported by DeepSeek may turn out to be bogus, and that NVIDIA stock price may rebound in such a case.
But DeepSeek was not the only recent breakthrough regarding efficiency and economy. Just about a week before DeepSeek R1 was released, NovaSky, which is currently not a company but a collaborative initiative led by students and advisors at UC Berkeley’s Sky Computing Lab, released Sky-T1, a reasoning LLM on par with OpenAI's o1-preview (currently at #7 of the Chatbot Arena LB), with a total training cost of $450 (not a typo, $450, without being accompanied with "thousand" or anything else):
https://novasky-ai.github.io/posts/sky-t1/
https://github.com/NovaSky-AI/SkyThought
https://techcrunch.com/2025/01/11/researchers-open-source-sky-t1-a-reasoning-ai-model-that-can-be-trained-for-less-than-450/
I can only guess that the main reason why Sky-T1 did not generate headlines (and financial panic) is that they did not release a web interface and/or app with a working model (hence the model did not enter any leaderboards); but I don't think that this will be the case for much longer.
Why might you be wrong?
Volatility