I'm trying to follow Harari's advice and refrain from saying this article is exciting. Instead I'll say I will need to digest the data and arguments slowly. There is a lot here. Just posing the question ... Will we have enough compute ... is a significant contribution.
Great data though your thesis is almost certainly the opposite of the truth. Don't believe me? Redo your demand chart through 2030 or 2033 instead of truncating it at 2028. If we went from 5% to 13% to 34% share in 2026/2027/2028 respectively, what would a reasonable estimate be of 2030? Of 2033? Is it impossible to believe that we are at 50% or even 70%? That reverses the conclusion of the paper.
Thanks for your comment. We deliberately use bullish demand assumptions and moderate supply estimates to stress-test the idea that GenAI won't break the regime of affordable, widely available computing power in the near-term. Moreover, by then there is sufficient time for typical business planning cycles to adapt.
"Sufficient time" is a good argument I think, though not the stated thesis of the paper. I am just concerned that we are missing the exponential view in favor of the truncated view. I probably wouldn't comment on the data interpretation if I didn't have such high respect for this substack's understanding of the rate of change.
That is fair, we should have stated it more explicitly. While our model does 'break' in the 2030s, it's important to understand that we're not attempting to predict the most likely scenario. Instead, we're presenting a steelman case to demonstrate we won't face a compute shortage in the near future. Our model deliberately understates the potential growth in compute supply, especially after 2025 (in reality we think it will be much stronger growth). The truly exponential aspect of our view lies not in the model, but the implications:: what transformative possibilities emerge when we have access to such vast computational resources?
I'm trying to follow Harari's advice and refrain from saying this article is exciting. Instead I'll say I will need to digest the data and arguments slowly. There is a lot here. Just posing the question ... Will we have enough compute ... is a significant contribution.
Thankyou! It is certainly wise advice!
Great data though your thesis is almost certainly the opposite of the truth. Don't believe me? Redo your demand chart through 2030 or 2033 instead of truncating it at 2028. If we went from 5% to 13% to 34% share in 2026/2027/2028 respectively, what would a reasonable estimate be of 2030? Of 2033? Is it impossible to believe that we are at 50% or even 70%? That reverses the conclusion of the paper.
Thanks for your comment. We deliberately use bullish demand assumptions and moderate supply estimates to stress-test the idea that GenAI won't break the regime of affordable, widely available computing power in the near-term. Moreover, by then there is sufficient time for typical business planning cycles to adapt.
"Sufficient time" is a good argument I think, though not the stated thesis of the paper. I am just concerned that we are missing the exponential view in favor of the truncated view. I probably wouldn't comment on the data interpretation if I didn't have such high respect for this substack's understanding of the rate of change.
That is fair, we should have stated it more explicitly. While our model does 'break' in the 2030s, it's important to understand that we're not attempting to predict the most likely scenario. Instead, we're presenting a steelman case to demonstrate we won't face a compute shortage in the near future. Our model deliberately understates the potential growth in compute supply, especially after 2025 (in reality we think it will be much stronger growth). The truly exponential aspect of our view lies not in the model, but the implications:: what transformative possibilities emerge when we have access to such vast computational resources?