It doesn't feel like DeepSeek has a big enough breakthrough here. This is just one of many optimizations we're going to see over the next years. How close this brings us to "AGI" is a complete unknown.
The large investments were mainly for training larger foundation models, or at the very least hedging for that. It hasn't been that clear over the last 1+ years that simply increasing the number of parameters continues to lead to the same improvements we've seen before.
Markets do not necessarily have any prediction power here. People were spooked by DeepSeek getting ahead of the competition and by the costs they report. There is still a lot of work and some of it may still require brute force and more resources (this seems to be true for training the foundation models still).
The large investments were mainly for training larger foundation models, or at the very least hedging for that. It hasn't been that clear over the last 1+ years that simply increasing the number of parameters continues to lead to the same improvements we've seen before.
Markets do not necessarily have any prediction power here. People were spooked by DeepSeek getting ahead of the competition and by the costs they report. There is still a lot of work and some of it may still require brute force and more resources (this seems to be true for training the foundation models still).