It's been a wildly busy week, so I will touch quickly on a few things that keep coming up in my discussions.
Here is what's ahead:
- AI and Trade Policy
- Nvidia as Pop Mart
- GPU Spot Prices
- The Five Alarm Fire in Power
- AI and Trade Policy
I regularly make the point, one that seemingly confuses people, that AI spending—capex, in particular—is a significant fraction of U.S. GDP growth, perhaps more than half, and that it can thus confound policy. For example, if I want to know the effect of tariffs it can be harder than expected to untangle it, as well as their effects being muted, by the massive stimulus flowing from AI capex.
This seems incontrovertible. Trade policy is hard, and evolving confounds make it harder. This is not the same thing, however, as saying "tariffs bad", "tariffs good", etc. It is simply saying that when you have large, novel, and fast-moving rivers of money flowing around as you conduct trade policy it can make it difficult to know what is causing what, which is among the main goals of evaluating said policy.
As I like to point out, the money flowing into AI data centers is coming from somewhere. And it is, by definition, thus not going other places. To the extent that a goal of current U.S. trade policy is to onshore more manufacturing, that can make it more difficult for manufacturers, at least at the margin, to get all the capital they need.
- Nvidia as Pop Mart
Arguing that it is fairly obvious we are in a financial bubble is not the same thing as saying AI is the Labubu of technology. This seems an obvious point, but is a constant rejoinder to talk of bubbles. AI is not some whimsical thing being driven higher by dopey investors chasing a shiny object; Nvidia is not the Pop Mart of technology.
Clearly stated: AI is among the most important developments in technology, and current AI progress is the most significant in that domain in my lifetime. It has myriad uses; there will be more, and progress continues. Saying that movement of the stocks, or the related overspending, have all the hallmarks of a financial bubble is to state the obvious—and is not the same thing as saying AI is useless.
The global financial crisis was among the largest bubbles in modern economic history, and one driver was excess home construction. But homes were not a fad. It was the excessive spending and speculative nexus around the homes that led the eventual collapse of the related credit complex, and ensuing systemic crisis.
- GPU Spot Prices
Once an idea takes hold among people, it is hard to get them to drop it. A current example is the idea that Nvidia GPUs are in scarce supply. They are not, or least aren't in the usual narrative sense. A quick check of marketplace prices will show you that prices are down by almost half this year, depending on location and specific GPU. This is not the hallmark of something in short supply.
The usual response to this is that marketplaces aren't good indications of Real AI Work. The latter, it is argued, requires service level agreements, redundant power, and reservation systems, the sort of things for which major hyperscalers get 2-4x rental premiums over marketplaces.
Outside of training frontier models or massive-scale chatbots from the big three/four, I am less convinced of this than ever. An increasing fraction of startup, emergent model, and other work is done on marketplaces, given the huge price advantage and increasing supply of rental GPUs. This is putting growing pressure on hyperscalers, at least at the margin, given that the latter are trapped by their cost of capital, which prevents them from competing on price.
- The Five Alarm Fire in Power