DeepSeek had plenty of R&D expertise which were not included in the (declared) model training cost. Here we are talking about building something nearly from scratch, even if there is an open source starting point you still need the infrastructure, expertise and people to make it work, which with that budget are going to be hard to secure. Moreover these projects take months and months to get approved, meaning that this one was conceived long before DeepSeek, thus highlighting the original disalignment between the goal and the budget. DeepSeek might have changed the scenario (I hope so) but it would be just a lucky ex-post event… not a conscious choice behind that budget.
Aleph Alpha is a business that has been going for some time in this sector, at least a couple of years with commercial LLM products. It's likely they'll provide hardware and base models for this project.
DeepSeek probably spent closer to two billion on hardware. And then there’s the energy cost of numerous runs, staff costs, all of that. The 5.5m cost was basically misleading info, maybe used strategically to create doubt in the US tech industry or for DeepSeek’s parent hedge fund to make money off shorts.
DeepSeek had plenty of R&D expertise which were not included in the (declared) model training cost. Here we are talking about building something nearly from scratch, even if there is an open source starting point you still need the infrastructure, expertise and people to make it work, which with that budget are going to be hard to secure. Moreover these projects take months and months to get approved, meaning that this one was conceived long before DeepSeek, thus highlighting the original disalignment between the goal and the budget. DeepSeek might have changed the scenario (I hope so) but it would be just a lucky ex-post event… not a conscious choice behind that budget.
What do you mean, "nearly from scratch"?
Aleph Alpha is a business that has been going for some time in this sector, at least a couple of years with commercial LLM products. It's likely they'll provide hardware and base models for this project.
DeepSeek probably spent closer to two billion on hardware. And then there’s the energy cost of numerous runs, staff costs, all of that. The 5.5m cost was basically misleading info, maybe used strategically to create doubt in the US tech industry or for DeepSeek’s parent hedge fund to make money off shorts.
https://semianalysis.com/2025/01/31/deepseek-debates/