BTW: Morgan Stanley pegs nVidia as 65% Datacenter, so more aggressive that your baseline.
When examining technology capex, there should be two questions in the investor's mind: Where are we on the S-Curve and what is the price elasticity of my company's product, which unfortunately seems to be missed as the central part of Jevons "Paradox."
The cloud is basically turning into a token utility. In 2022, the cost of 1M tokens was $90. Today with a Llama model, you are in the range of 1 cent. The investment is not only for scale, but it is also necessary to drive down the cost of the tokens. While nVidia says Blackwells is 12x more power efficient or 30x faster on some workloads, the at scale data centers will have hard data as they scale. While 30x is marketing fluff, I have no doubt the end TCO will make Blackwell a brain dead decision for upgrades and capex.
Jevons paradox is a poor marketing terms and gives the idea that somehow there is something magical happening with tech. Better termed is the Jevons Effect, and it simply states that the demand is elastic.
The only question that the investor should be making about these investments is "what is the slope of my elasticity curve?" If you use the tech, it will be apparent it is elastic, and a 20% reduction in cost is going to drive >20% increase in demand.
Just checking: did you disaggregate AMZN Capex between the retail and AWS businesses?
I am so curious if we added Stargate what that would look like.
How long will this go on for…?
BTW: Morgan Stanley pegs nVidia as 65% Datacenter, so more aggressive that your baseline.
When examining technology capex, there should be two questions in the investor's mind: Where are we on the S-Curve and what is the price elasticity of my company's product, which unfortunately seems to be missed as the central part of Jevons "Paradox."
The cloud is basically turning into a token utility. In 2022, the cost of 1M tokens was $90. Today with a Llama model, you are in the range of 1 cent. The investment is not only for scale, but it is also necessary to drive down the cost of the tokens. While nVidia says Blackwells is 12x more power efficient or 30x faster on some workloads, the at scale data centers will have hard data as they scale. While 30x is marketing fluff, I have no doubt the end TCO will make Blackwell a brain dead decision for upgrades and capex.
Jevons paradox is a poor marketing terms and gives the idea that somehow there is something magical happening with tech. Better termed is the Jevons Effect, and it simply states that the demand is elastic.
The only question that the investor should be making about these investments is "what is the slope of my elasticity curve?" If you use the tech, it will be apparent it is elastic, and a 20% reduction in cost is going to drive >20% increase in demand.