I think I disagree strongly with this essay, but I struggled to follow the argument structure. What's wrong with the plan "We think renting out compute will be lucrative, so we're investing in data center buildout to enable that in a couple year"? Maybe it doesn't count as national strategy, but it's definitely business strategy. You wrote,
> Simply building datacenters and renting out relatively small batches of compute ad-hoc with no long-term plannability is not a very attractive business model for most countries
But, like, this strategy seems to be doing great currently and in the last couple years. I would imagine a lot of it will come down to how much demand there is for compute in a couple years, and I think there's a strong case for "tons".
I think the argument structure takes for granted the premise that there is some deeper strategic aim around 'sovereignty' in these buildouts. My strong sense is that this is the political motivation behind most of these projects, so I went through different concepts of how that would be achieved.
I get to the question of lucrativity only later because I think of that as more of a 'backstop' for if your strategic ambition didn't really work out. High-level, there just seems to be a disconnect between the extraordinary scale of industrial and infrastructural policy invested and the notion that this is primarily about running a lucrative datacenter business.
That said, I think there is a strong version of the business-focused case for buildouts, which is that it at least gives you some participation in AI-driven growth, and obviously that has strategic merit if you think most other growth avenues will close. I think that's not super realistic and trades off directly against participating at other points of the value chain, but let's assume it's true - I still think buildouts are unattractive business for most middle powers.
That's mostly for the reasons listed in the post: I think as a middle power, you'll be competing with countries that are a lot better at deploying this compute. Either great powers that have a lot of political will and crazy scaling effects, or the AI oil powers that I describe, which have a lot of favourable regulatory & energy conditions as well. Against that competition, I think you'll find it very difficult to create a competitive and profitable datacenter product. Except, of course, if there is so little compute that even inefficiently deployed compute can be offered at decent margins - so you'll probably be outcompeted for GPUs by more efficient products, or at least have to spend so much that the datacenter business becomes even less profitable.
To add, I think I disagree with your notion that this strategy seems to be doing great currently. Running datacenters is kind of lucrative - in favourable conditions, with close and lasting commitments with compute-intensive labs/developers, etc. But I don't think the middle power buildouts are headed for these conditions. I don't know that currently, government-driven datacenter projects in sometimes high-energy-price countries with large overheads that rely on ad-hoc renting-out are very profitable, and that's the scenario I'm trying to get at with the quoted line!
I think I disagree strongly with this essay, but I struggled to follow the argument structure. What's wrong with the plan "We think renting out compute will be lucrative, so we're investing in data center buildout to enable that in a couple year"? Maybe it doesn't count as national strategy, but it's definitely business strategy. You wrote,
> Simply building datacenters and renting out relatively small batches of compute ad-hoc with no long-term plannability is not a very attractive business model for most countries
But, like, this strategy seems to be doing great currently and in the last couple years. I would imagine a lot of it will come down to how much demand there is for compute in a couple years, and I think there's a strong case for "tons".
Thanks for the comment!
I think the argument structure takes for granted the premise that there is some deeper strategic aim around 'sovereignty' in these buildouts. My strong sense is that this is the political motivation behind most of these projects, so I went through different concepts of how that would be achieved.
I get to the question of lucrativity only later because I think of that as more of a 'backstop' for if your strategic ambition didn't really work out. High-level, there just seems to be a disconnect between the extraordinary scale of industrial and infrastructural policy invested and the notion that this is primarily about running a lucrative datacenter business.
That said, I think there is a strong version of the business-focused case for buildouts, which is that it at least gives you some participation in AI-driven growth, and obviously that has strategic merit if you think most other growth avenues will close. I think that's not super realistic and trades off directly against participating at other points of the value chain, but let's assume it's true - I still think buildouts are unattractive business for most middle powers.
That's mostly for the reasons listed in the post: I think as a middle power, you'll be competing with countries that are a lot better at deploying this compute. Either great powers that have a lot of political will and crazy scaling effects, or the AI oil powers that I describe, which have a lot of favourable regulatory & energy conditions as well. Against that competition, I think you'll find it very difficult to create a competitive and profitable datacenter product. Except, of course, if there is so little compute that even inefficiently deployed compute can be offered at decent margins - so you'll probably be outcompeted for GPUs by more efficient products, or at least have to spend so much that the datacenter business becomes even less profitable.
To add, I think I disagree with your notion that this strategy seems to be doing great currently. Running datacenters is kind of lucrative - in favourable conditions, with close and lasting commitments with compute-intensive labs/developers, etc. But I don't think the middle power buildouts are headed for these conditions. I don't know that currently, government-driven datacenter projects in sometimes high-energy-price countries with large overheads that rely on ad-hoc renting-out are very profitable, and that's the scenario I'm trying to get at with the quoted line!
Thanks for explaining, I feel like I understand your position better now.
This is remarkably insightful and provides helpful context I haven’t had available before. Thank you.