America's $1T AI Gamble

Possibly, but that assumes continuity. New math and algorithmic breakthroughs could make much of today’s AI stack legacy, reshuffling both costs and winners.

A significant part of the capex is just energy, so even if there is some sort of AI black swan event and the data centers become obsolete overnight (unlikely), energy is literally the root of all bounty so it is good that something is incentivizing increased resource allocation in that area.

Most of this land was low-utility anyway. You should realise it is good for the land owners to convert it to high yield output, which in turn the government can tax and return some of the gains to the people.

What's the alternative ?

This is a good analyst report - lots of data. Conclusion - firms are spending ahead of sustained revenues right now, and a lot of the money is going offshore to TSMC, basically.

I’m not certain of the conclusion - I think a lot depends on amortization schedules - if data centers are fully booked right now, then we don’t need very long amortization schedules at the reported 60+% margin on inference to see this capex fully paid off.

My prior is that we are seeing something like 1/10,000th or so of the reasonable inference demand the world has fulfilled. There’s a note in the analysis that might back this - it says that we are seeing one of the only times ever where hardware prices are rising over time. Combined with spot prices at lambda labs (still quite high I’d say), it doesn’t look like we’re seeing a drop in inference demand.

Under those circumstances, the first phases of this bet, cross-industry, look like they will pay off. If that’s true, as an investment strategy, I’d just buy the basket - oAI, Anthropic, GOOG, META, SpaceX, MSFT, probably even Oracle, and wait. We’ll either get the rotating state of the art frontier capacity we’ve gotten in the last 18 months, or one of those will have lift off.

Of those, I think MSFT is the value play - they’re down something like 20% in the last six months? Satya’s strategy seems very sensible to me - slow hyperscale buildouts in the US (lots of competition) and do them everywhere else in the world (still not much competition). For countries that can’t build their own frontier models, the next best thing is going to be running them in local datacenters; MSFT has long standing operational bases everywhere in the world, it’s arguably one of their differentiators compared to GOOG/META.

Yeah, this is something I am thinking a lot about. Companies won't be able to sustain this level of spending forever, and one of two things will need to happen:

1. Models become commodities and immensely cheaper to operate for inference as a result of some future innovation. This would presumably be very bad for the handful of companies who have invested that $1T and want to recoup that, but great for those of us who love cheap inference.

2. #1 doesn't happen and the model providers start begin to feel empowered to pass the true cost of training + inference down to the model consumer. We start paying thousands of dollars per month for model usage and the price gate blocks out most people from reaping the benefits of bleeding-edge AI, instead being locked into cheaper models that are just there to extract cash by selling them things.

Personally I'm leaning toward #1. Future models near as good as the absolute best will get far cheaper to train, and new techniques and specialized inference chips will make them much cheaper to use. It isn't hard for me to imagine another Deepseek moment in the not-so-distant future. Perhaps Anthropic is thinking the same thing given the rumors that they are rumored to be pushing toward an IPO as early as this year.