> Nations including India, Japan, France and Canada are talking about the importance of investing in “sovereign AI capabilities,” Huang said in an interview Thursday with Bloomberg Television. “Their natural resource – data – should be refined and produced for their country. The recognition of sovereign AI capabilities is global.”
Data as a natural resource is an interesting analogy. Especially if you frame it in the category of nuclear fissile material.
But why would we do that? Nuclear fissile material is noted for its ability to be formed into weapons that generate dangerous explosions and substances both acutely and chronically hazardous to human life, increasing cancer risk.
This is not a particularly obvious analogy to “data”.
The analogy is the costs and risks of storage. Data might not cause cancer if you stand next to it, but there’s plenty of risks involved in mass accumulation of data controlled by government entities. And anything that gives more incentive for that accumulation, like being able to run in house ML models, will surely lead to even more accumulation.
ai is a national defense industry and every major country would like their own if they can afford it. just like energy, defense, space, telecom, etc. ai might have a cheaper barrier to entry (for now) than the others
That’s only if there is an overlap between the customers who want to use cloud and the ones who want to run things on premise.
I would assume that those who want to run things on prem aren’t currently cloud customers for other services. And those who are already heavily invested in a cloud provider will look to the same provider for AI compute.
I would assume these are entirely net new use cases. I also don’t know why this wouldn’t be developed in localized cloud regions. Aws, for instance, running in India is locally compliant with regulations. It’s cheaper for aws to comply with Indian regulations than it is for a random company, or worse a government, to build aws from whole cloth in India (nothing here is specific to India, just used as an example).
Before anyone steps in in favor of on prem vs aws, I’ve worked at cloud providers, building megacorp on prem, and migrating said on prem to the cloud, and building native cloud companies. Execution risks and the enormous high energy state of building full service on prem solutions make it much more reasonable to use cloud providers for stuff like this, and the economics of capital investment, the accounting of things, etc, makes it unreasonable for governments to NOT use cloud providers.
Regardless of our particular feelings about working in the defense industry, it is well known that this is a topic that many people have strong feeling about (I intentionally didn’t state my feeling on the matter, because I know this is a basic right-and-wrong issue for people who’ve thought seriously about it, so I’m not going to try and make an argument based on my biases there).
Nvidia’s engineers are at the top of their field and have chosen not to work directly in that field at least, so I’d expect the company to have a higher than average density of folks who are opposed to it.
I think AI training presents quite a nice opportunity for states like Arizona, Nevada, New Mexico, Texas (plenty of available land for solar power plants, many sunny days, few rainy ones, nights not that long in the winter). Mexico and Australia are also very well positioned.
A hyperscale (cloud) datacenter uses about 2000 m3 of water per day [1]. That sounds like a lot, without any frame of reference. But if we compare with the average golf course in the South-West, that uses 459 acre-feet of water per year ([2], Table 2) which comes to 1550 m3 of water per day. There are more than 200 golf courses in the South West. You can draw your own conclusions.
28 comments
[ 3.6 ms ] story [ 78.2 ms ] threadData as a natural resource is an interesting analogy. Especially if you frame it in the category of nuclear fissile material.
This is not a particularly obvious analogy to “data”.
The TikTok fears are limited to capturing like/don't-like.
Generalize that to the massive amount of data streaming between cloud models and back?
Future ASI being dangerous is very feasible. For arguments, see LessWrong or Bostrom's work.
I would assume that those who want to run things on prem aren’t currently cloud customers for other services. And those who are already heavily invested in a cloud provider will look to the same provider for AI compute.
Before anyone steps in in favor of on prem vs aws, I’ve worked at cloud providers, building megacorp on prem, and migrating said on prem to the cloud, and building native cloud companies. Execution risks and the enormous high energy state of building full service on prem solutions make it much more reasonable to use cloud providers for stuff like this, and the economics of capital investment, the accounting of things, etc, makes it unreasonable for governments to NOT use cloud providers.
These seems less like information and more like news with a market agenda.
Nvidia’s engineers are at the top of their field and have chosen not to work directly in that field at least, so I’d expect the company to have a higher than average density of folks who are opposed to it.
I need to save a few thousand dollars for the US Libertarian Party and the EU Pirate Party.
[1] https://dgtlinfra.com/data-center-water-usage/
[2]https://usgatero.msu.edu/v11/216335.pdf