Great post! A few comments on points where I have a slightly different or otherwise complementary view.
In my view, the market has already been pricing in a turn in the cycle to some extent for a good while. Nvidia’s earnings have grown significantly, and valuation multiples have been on a downward trend throughout the investment cycle. There are also numerous other companies in the semiconductor sector whose price multiples are exceptionally low compared to their own history and the general market pricing. This is despite the fact that growth figures and margins will almost certainly remain unprecedented for at least the next few quarters.
Based on this, one could interpret that the market does indeed recognize the exceptional nature of this cycle and its likely end (the so-called Molodovsky effect). Of course, it is also a fact that when the cycle levels off or turns, the blow comes from three directions: 1) revenue growth stalls, 2) margins drop, 3) valuation multiples contract. So, if a dollar of Nvidia’s sales currently flows through at a 55% net margin and is valued at 40x, it is possible (and likely) that in the future, $0.8 in sales will result in only a 20% margin and be valued at 15x. And by no means is the market pricing in a turn around 2027–2028 yet. This suggests some ugly results are ahead if your guess is correct (in my numerical example, which I don’t think is utopian at all, that’s a -90% drop).
In my view, this cyclical bubble differs significantly because the rise in prices and margins is specifically caused by supply bottlenecks. Typically, in bubbles caused by fundamental technological shifts, (debt-funded) supply exceeds demand—whereas now, it’s the other way around, and largely funded by cash flow. As far as I understand, there is no oversupply yet in any part of the AI stack. My biggest concern is the data center buildout, where there is a risk of debt-funded oversupply, which could result in a “bullwhip effect” hitting harder than just the AI field. My own assessment is that the knot will start to unravel there—not from Nvidia. That’s why my “bubble radar” is aimed at these companies and their ability to sell/lease capacity (preferably long before the data centers are even finished).
This is true, but I would be cautious about thinking that humans would settle for the level of technology that was the best in year X. We each have a more powerful computing device in our pockets than the world’s most powerful supercomputer was in 2000 (ASCI Red). A lot of hardware and algorithmic development needs to happen before, for example, even current (primitive) video or image modality model computing requirements can be run locally. If/when energy constraints are somehow managed, my view is that demand for computing will continue to grow “limitlessly,” even though efficiency is constantly improving (in accordance with Jevons’ Paradox). This has happened with virtually every technology developed by humans.
I see a major threat here. However, not because World models and the pursuit of AGI would no longer interest the AI community, but because being a listed company brings financial realities into play. Google’s history is a good example; after its IPO, there was a long period where the company essentially stopped all innovation. Even Nvidia’s Huang was almost ousted back in the day when Cuda development was considered completely insane on Wall Street, and it was a close call that a corporate raider didn’t take over the company.
As a private company, experimental R&D is simply much easier. And when frontier companies in the future have to turn their operations profitable, it is a likely scenario that focus shifts to short-term ROI maximization. Especially since Anthropic and OAI don’t exactly have other cash flow sources to fund experimental work. The threat is that a couple of major AGI competitors might ease the “prisoner’s dilemma” problem by going public.
I don’t believe that hardware investments will stop, though. Simply because the performance-per-watt benefit, which converts directly into revenue for AI companies, is greater with each hardware generation than the dollars invested in the hardware.