The prisoner's dilemma framing captures exactly why the capex cycle won't slow despite valuation concerns. When Zuck says he'd rather waste billions than lose the AI race, that's not hyperbole, its game theory driving a multi hundred billion dollar reallocation across the entire tech sector. The most compelling datapoint is the 10 point ROIC increase for hyperscalers since ramping GPU spend, which completely undercuts the bubble narrative when you compare it to the 97% dark fiber peak in 2000.
On the point about comparing the current AI buildout to the fiber buildout of the late 90s/200s: I get why people do this. It's a seductive and easy comparison to make. But dark fiber sat there unused, ready to be lit when demand arose. GPUs, on the other hand, become obsolete after 3 years. So if demand for GPUs decline, either due to algorithmic improvements or the rise of edge-based inference, or a combination of those two factors, then you will likely have a lot of stranded assets can't readily be repurposed.
Gavin Baker makes interesting points on the valuations & supply/demand. If humans are humans, perhaps this means that the chip cos will carry on going until there ARE a load of dark GPUs, and that will be the Bubble peak?
As an outsider, I still frame it simply as, 'AI is not in a bubble, IF it does what people believe it will'. But I am still not convinced that it will. Clearly, the MAG7 CEOs SHOULD have a much better view on this than myself. OR you could say that they have been blinded by science & the companies are wearing the 'emperor's new clothes'.
I keep coming back to the line, 'don't run before you can walk'. It's undeniable that the majority of people struggle with everyday processes due to 'enshittification'. Whether it's dealing with bureaucrats for basic services, or big companies that still have archaic tech infrastructure - this is the real world. To me, these are the shaky foundations that AI will effectively rest on, and so while there are many things AI WILL do (e.g. DeepMind health stuff) I still think there is much that is claimed that will not work.
Accordingly, I am still sceptical that AI will achieve as much as is suggested, though I'm not at Ed Zitron levels...
The nuclear power revival you document is maybe the most underappreciated aspect of AI infrastructure buildout. Mitsubishi Heavy Industries' involvement in the $100B Westinghouse AP1000 project alongside that NextEra restart shows how the energy bottleneck is forcing hyperscalers to think 10+ year timelines. What's striking is the profit sharing mechanisms that give the US government upside once thresholds are met, essentially making taxpayers equity holders in the nuclear builout. If Satya is right that power availability matters more than compute now, then companies like MHI that can deliver reliable gigawatt scale capacity become strategic kingmakers in determining who wins the AI arms race.
The prisoner's dilemma framing of big tech's AI arms race really captures why this spending trajectory is locked in regardless of near term ROIC concerns. Gavin's point about Broadcom and AMD effectively going to market together with custom ASICs plus Ethernet fabric is critical for understanding the competitive dynamics beyond just Nvidia versus Google TPUs. The hyperscaler capex projections through 2028 show sustained demand not just for compute but also for networking and custom silicon where Broadcom has estblished positions. The valuations relative to growth rates make it clear that the market is pricing in execution risk, but the fundamental drivers around AI infrastructure remain incredibly robust.
Very valuable! Thank you
The prisoner's dilemma framing captures exactly why the capex cycle won't slow despite valuation concerns. When Zuck says he'd rather waste billions than lose the AI race, that's not hyperbole, its game theory driving a multi hundred billion dollar reallocation across the entire tech sector. The most compelling datapoint is the 10 point ROIC increase for hyperscalers since ramping GPU spend, which completely undercuts the bubble narrative when you compare it to the 97% dark fiber peak in 2000.
Good overview.
On the point about comparing the current AI buildout to the fiber buildout of the late 90s/200s: I get why people do this. It's a seductive and easy comparison to make. But dark fiber sat there unused, ready to be lit when demand arose. GPUs, on the other hand, become obsolete after 3 years. So if demand for GPUs decline, either due to algorithmic improvements or the rise of edge-based inference, or a combination of those two factors, then you will likely have a lot of stranded assets can't readily be repurposed.
Great piece - thank you.
Gavin Baker makes interesting points on the valuations & supply/demand. If humans are humans, perhaps this means that the chip cos will carry on going until there ARE a load of dark GPUs, and that will be the Bubble peak?
As an outsider, I still frame it simply as, 'AI is not in a bubble, IF it does what people believe it will'. But I am still not convinced that it will. Clearly, the MAG7 CEOs SHOULD have a much better view on this than myself. OR you could say that they have been blinded by science & the companies are wearing the 'emperor's new clothes'.
I keep coming back to the line, 'don't run before you can walk'. It's undeniable that the majority of people struggle with everyday processes due to 'enshittification'. Whether it's dealing with bureaucrats for basic services, or big companies that still have archaic tech infrastructure - this is the real world. To me, these are the shaky foundations that AI will effectively rest on, and so while there are many things AI WILL do (e.g. DeepMind health stuff) I still think there is much that is claimed that will not work.
Accordingly, I am still sceptical that AI will achieve as much as is suggested, though I'm not at Ed Zitron levels...
Thanks again.
The nuclear power revival you document is maybe the most underappreciated aspect of AI infrastructure buildout. Mitsubishi Heavy Industries' involvement in the $100B Westinghouse AP1000 project alongside that NextEra restart shows how the energy bottleneck is forcing hyperscalers to think 10+ year timelines. What's striking is the profit sharing mechanisms that give the US government upside once thresholds are met, essentially making taxpayers equity holders in the nuclear builout. If Satya is right that power availability matters more than compute now, then companies like MHI that can deliver reliable gigawatt scale capacity become strategic kingmakers in determining who wins the AI arms race.
The prisoner's dilemma framing of big tech's AI arms race really captures why this spending trajectory is locked in regardless of near term ROIC concerns. Gavin's point about Broadcom and AMD effectively going to market together with custom ASICs plus Ethernet fabric is critical for understanding the competitive dynamics beyond just Nvidia versus Google TPUs. The hyperscaler capex projections through 2028 show sustained demand not just for compute but also for networking and custom silicon where Broadcom has estblished positions. The valuations relative to growth rates make it clear that the market is pricing in execution risk, but the fundamental drivers around AI infrastructure remain incredibly robust.