I'm not sure that needing less hardware for AI training is a negative for ASML. I think that the biggest threat to the AI boom is the difficulty to monetize it, at least to the extent needed to offset the immense spending. This can lead to the same situation as with the dot-com bubble, where investors got tired of the lack of ROI at one point, and the companies with poor business models got culled, leading to a major disinvestment.
Cheaper training makes more business models viable, and thus should lead to less disinvestment. At the very least, it should make demand for ASML machines less volatile, resulting in a substantial demand that persists (although not at an unsustainable Boom-level).
However...
Deepseek has obviously been lying about their training costs. The hardware they claim to have used is more expensive than they claim. They don't count any of the experimental training runs. Etc.
I also have my doubts whether they actually used as little hardware as they claimed, given that China seems to have been legally and illegally buying 4090's, to convert them into servers, using new PCBs. So if they used these GPUs, it's in their best interest to hide that fact, to prevent a further crackdown by the US on the routes they use to gather 4090's (and probably 5090's before long).
There is a project called Open R1 that attempts to replicate the software used for training Deepseek, and it might tell us a lot more about what the actual training costs are for a model like R1.
Excellent post and visuals!
I'm not sure that needing less hardware for AI training is a negative for ASML. I think that the biggest threat to the AI boom is the difficulty to monetize it, at least to the extent needed to offset the immense spending. This can lead to the same situation as with the dot-com bubble, where investors got tired of the lack of ROI at one point, and the companies with poor business models got culled, leading to a major disinvestment.
Cheaper training makes more business models viable, and thus should lead to less disinvestment. At the very least, it should make demand for ASML machines less volatile, resulting in a substantial demand that persists (although not at an unsustainable Boom-level).
However...
Deepseek has obviously been lying about their training costs. The hardware they claim to have used is more expensive than they claim. They don't count any of the experimental training runs. Etc.
I also have my doubts whether they actually used as little hardware as they claimed, given that China seems to have been legally and illegally buying 4090's, to convert them into servers, using new PCBs. So if they used these GPUs, it's in their best interest to hide that fact, to prevent a further crackdown by the US on the routes they use to gather 4090's (and probably 5090's before long).
There is a project called Open R1 that attempts to replicate the software used for training Deepseek, and it might tell us a lot more about what the actual training costs are for a model like R1.