A credible lab making a credible claim to massive efficiency improvements is a credible threat to Nvidia's future earnings. Hence the stock got sold. It's not more complicated than that.
Its obviously constrained by this hardware and this model size as it does some strange things sometimes and it is slow (30 secs to respond) but I've got it to do some impressive things that GPT4 struggles with or fails on.
Also of note I asked it about Taiwan and it parroted the official CCP line about Taiwan being part of China, without even the usual delay while it generated the result.
Inference cost - DeepSeek is charging less than OpenAI to use its public API, but that isn't an indicator of anything since it doesn't reflect the actual cost of operation. It's pretty much a guarantee that both companies are losing money. Looking at DeepSeek's published models the inference cost is in the same ballpark as Llama and the rest.
Which leaves training, and that's what all the speculation is about. The CEO said that the model cost $5.5M and that's what the entire world is clinging on. We have literally no other info and no way to verify it (for now, until efforts to replicate it start to show results).
Again, the weights are public. You can run the full-fat version of R1 on your own hardware, or a cloud provider of your choice. The inference costs match what DeepSeek are claiming, for reasons that are entirely obvious based on the architecture. Either the incumbents are secretly making enormous margins on inference, or they're vastly less efficient; in the first case they're in trouble, in the second case they're in real trouble.
Deepseek has distilled deepseek R1 into a couple of smaller open source models, but neither R1 or v3 are distilled themselves.