NVIDIA opened down something like 11-12% today. Microsoft took a bath in early trading—to the point where Satya was tweeting about Jevons last night lol—the sellers don't understand how AI works
On Twitter, I wrote yesterday (riffing off of one of your TT videos), “I personally would say, “intelligence” is always in demand and a cheaper price for that intelligence will spur on greater demand, thus requiring more chips. It’s just that the supply curve for those chips will be less steep than it was before.” After reading this essay, I would amend that ending to say that the supply curve will only change temporarily, if that much. To be honest, it will probably be more cost-efficient for Nvidia and the other chip makers to just plow through and maintain production, as if nothing has really changed.
I think Dario Amodei put the ongoing rationale for buying chips really well today:
"There is an ongoing trend where companies spend more and more on training powerful AI models, even as the curve is periodically shifted and the cost of training a given level of model intelligence declines rapidly. It's just that the economic value of training more and more intelligent models is so great that any cost gains are more than eaten up almost immediately — they're poured back into making even smarter models for the same huge cost we were originally planning to spend."
1. Even if they aren't hiding a bag of chips, agree that these efficiency gains don't mean diminished chip demand. If anything, it seems the opposite - we can now do more with less, meaning more competition and accelerated pace of innovation
2. I do question how this impacts the economics of leading AI labs investing to push frontier models forward. If scorched earth by open source is the strategy, what's the incentive to continue spending big money to advance new models when it quickly becomes replicated and given away for free afterwards?
3. Would love your expanded thoughts on how the price of energy plays into this. China has a seemingly huge advantage unless there's a lot of deregulation in US around nuclear or rapid improvement in renewables. Beyond chip demand, energy cost to train and serve the models seems like an important input that becomes more pronounced over a long timeline
I think fundamentally I don't see a world where we need less chips. Every scenario I can think of ends up on an accelerating demand curve.
There are two plausible reasons to keep pushing for frontier quality at enormous expense in this world: 1) you believe that Reinforcement Learning (RL) will eventually hit takeoff velocity and you'll be able to outrun everyone, 2) you believe that the demand for intelligence will be so high that there will be multiple winners. Both of these can be true.
As to energy: the question is around proximate dependable energy availability for data centers. I think in that sense Stargate functions as a market demand signal / trigger for private grid acceleration in the US. The total availability of energy matters less than the 24 hour dependable nearby energy for data centers. I'll probably write more here, but I think that fundamentally if we frame the problem as US is short energy, China is short chips, I'd rather be the US in that game. Energy is a more easily solvable problem.
Nate, this is about open-source. And, this is about the possibility of Deepseek not using Nvidia chips but rather heaper alternatives (announced by deepseek somewhere). Also, this is about Elon Musk saying: "I always wanted AI to be open-source". If you know what I mean.
I think it's about the bet that more efficient RL leads to lower demand for chips, all in. Even if AGI is open-source (which I think is likely) I don't think that really changes the thesis that there will be skyrocketing demands for intelligence.
Open-source is a part of the tech ecosystem and has been forever, but I don't see the fact that it exists preventing commercialization here anymore than it did for Linux.
What we’re witnessing here is a much-needed spark of competition, which is an exceptionally positive development. The drop in share price is simply a great opportunity in disguise. This marks the beginning of an even faster-paced race toward AGI, as it becomes clear that the United States is no longer the sole contender in this space. :D
Joseph Schumpeter, an economist known for his theories on innovation and entrepreneurship, introduced the concept of "Creative Destruction" in his works, most notably in "Capitalism, Socialism and Democracy" (1942). Creative destruction refers to the process through which innovation leads to the dismantling of long-established economic practices, industries, firms, or products, thereby making way for new ones.
So lets see how Zuck, Sam and Jeff responde. This can only be helpfull to shake the up a bit.
This reminds me of the "paradox" of L.A. freeways: They build more lanes onto the 405, thinking it will ease congestion. But the effect is to encourage more drivers to use it. I think David Owen gets at this in his underrated book The Conundrum.
On Twitter, I wrote yesterday (riffing off of one of your TT videos), “I personally would say, “intelligence” is always in demand and a cheaper price for that intelligence will spur on greater demand, thus requiring more chips. It’s just that the supply curve for those chips will be less steep than it was before.” After reading this essay, I would amend that ending to say that the supply curve will only change temporarily, if that much. To be honest, it will probably be more cost-efficient for Nvidia and the other chip makers to just plow through and maintain production, as if nothing has really changed.
I think Dario Amodei put the ongoing rationale for buying chips really well today:
"There is an ongoing trend where companies spend more and more on training powerful AI models, even as the curve is periodically shifted and the cost of training a given level of model intelligence declines rapidly. It's just that the economic value of training more and more intelligent models is so great that any cost gains are more than eaten up almost immediately — they're poured back into making even smarter models for the same huge cost we were originally planning to spend."
https://darioamodei.com/on-deepseek-and-export-controls
1. Even if they aren't hiding a bag of chips, agree that these efficiency gains don't mean diminished chip demand. If anything, it seems the opposite - we can now do more with less, meaning more competition and accelerated pace of innovation
2. I do question how this impacts the economics of leading AI labs investing to push frontier models forward. If scorched earth by open source is the strategy, what's the incentive to continue spending big money to advance new models when it quickly becomes replicated and given away for free afterwards?
3. Would love your expanded thoughts on how the price of energy plays into this. China has a seemingly huge advantage unless there's a lot of deregulation in US around nuclear or rapid improvement in renewables. Beyond chip demand, energy cost to train and serve the models seems like an important input that becomes more pronounced over a long timeline
I think fundamentally I don't see a world where we need less chips. Every scenario I can think of ends up on an accelerating demand curve.
There are two plausible reasons to keep pushing for frontier quality at enormous expense in this world: 1) you believe that Reinforcement Learning (RL) will eventually hit takeoff velocity and you'll be able to outrun everyone, 2) you believe that the demand for intelligence will be so high that there will be multiple winners. Both of these can be true.
As to energy: the question is around proximate dependable energy availability for data centers. I think in that sense Stargate functions as a market demand signal / trigger for private grid acceleration in the US. The total availability of energy matters less than the 24 hour dependable nearby energy for data centers. I'll probably write more here, but I think that fundamentally if we frame the problem as US is short energy, China is short chips, I'd rather be the US in that game. Energy is a more easily solvable problem.
Nate, this is about open-source. And, this is about the possibility of Deepseek not using Nvidia chips but rather heaper alternatives (announced by deepseek somewhere). Also, this is about Elon Musk saying: "I always wanted AI to be open-source". If you know what I mean.
I think it's about the bet that more efficient RL leads to lower demand for chips, all in. Even if AGI is open-source (which I think is likely) I don't think that really changes the thesis that there will be skyrocketing demands for intelligence.
Open-source is a part of the tech ecosystem and has been forever, but I don't see the fact that it exists preventing commercialization here anymore than it did for Linux.
thanks Nate
What we’re witnessing here is a much-needed spark of competition, which is an exceptionally positive development. The drop in share price is simply a great opportunity in disguise. This marks the beginning of an even faster-paced race toward AGI, as it becomes clear that the United States is no longer the sole contender in this space. :D
Joseph Schumpeter, an economist known for his theories on innovation and entrepreneurship, introduced the concept of "Creative Destruction" in his works, most notably in "Capitalism, Socialism and Democracy" (1942). Creative destruction refers to the process through which innovation leads to the dismantling of long-established economic practices, industries, firms, or products, thereby making way for new ones.
So lets see how Zuck, Sam and Jeff responde. This can only be helpfull to shake the up a bit.
This reminds me of the "paradox" of L.A. freeways: They build more lanes onto the 405, thinking it will ease congestion. But the effect is to encourage more drivers to use it. I think David Owen gets at this in his underrated book The Conundrum.