I find it hard to see an upper bound to demand for compute/power.
If models/hardware become 1000X more efficient/cheaper/smaller, people will be running 1,000,000X more jobs with even bigger models on even more intensive tasks, running them for even longer, etc.
We haven't even scratched the surface for video, world models and robotics.
When the markets panicked during the first DeepSeek episode, I just bought more Nvidia for this very reason. NVidia may very well tank sooner or later, but I'm sure it won't be because demand for compute tanked due to more efficient models. A more plausible scenario would be a competitor dethroning NVidia, kinda like Google's recent surge indicated.
Nature Is Laughing at the AI Build Out · HackerTrans
If models/hardware become 1000X more efficient/cheaper/smaller, people will be running 1,000,000X more jobs with even bigger models on even more intensive tasks, running them for even longer, etc.
We haven't even scratched the surface for video, world models and robotics.
When the markets panicked during the first DeepSeek episode, I just bought more Nvidia for this very reason. NVidia may very well tank sooner or later, but I'm sure it won't be because demand for compute tanked due to more efficient models. A more plausible scenario would be a competitor dethroning NVidia, kinda like Google's recent surge indicated.