[NA] Why not outsource (international) AI

Let'sgoflying!

Touchdown! Greaser!
Joined
Feb 23, 2005
Messages
21,872
Location
west Texas
Display Name
Dave Taylor
Home Airfield
KMRF
Teach the uninformed (me) about this.
I read every day about the huge pushback on data centers in communities (electricity & water use, visual affront & noise, possible light pollution).
So I asked AI (irony, right?) why we don't build them in India etc. I wasn't ready to buy the answers (security & latency, foreign control) so I l want to hear from POA, who knows so much about how the world works.
Security? Seems like either that can be dealt with - or frankly, US-based data centers will likewise have as much as hacking risk as offshore ones. Having physical access to a computer server, does that vastly increase the security risk (only of accessing data, not vandalizing)? We can't hardly keep anything secure right here (name any government agency or large corporation which has not been hacked), so I can't see this being any riskier.
Latency? Seem like we can fly bits around the planet with fiber and sats in milliseconds already?
If we are worried about India going rogue, distribute the assets among 50 countries?
Just seemed silly that we can easily move a s-load of jobs that were needed here, to other countries - but when it comes to something that is looking to be pretty harmful to our environment we can't seem to move them to places with an abundance of water & electricity.
Probably build these structures a lot cheaper too.
 
Here are two reasons to consider:

Firstly, data governance laws may have a lot to say about where you store and where you process data. European laws certainly do; as do US and Canadian laws… as do non-government industry standards (e.g. PCI). You can't simply put your US (or Australian or European or whatever) customer data in a data center in a random location if you want to stay on the right side of the law — or of your clients. This is a big selling point for, e.g. Google Cloud or AWS having regional data centers all over the place.

And, secondly, you discount it, but latency is a killer for data and processing that's not local. It takes roughly 25ms to ping my nearest Google datacenters in LA or Oregon; more than ten times that (325ms) to ping one in India from where I am (I just measured it). You can't do much about the fact that light only goes a foot in a nanosecond :) (or much reduce the time it takes to traverse all the switches and routers between here and there). All that adds up — a typical AI or big data transaction may go back and forth several times in that transaction (dozens of times, actually), and it all adds up. That LA data center will do the back and forth in maybe a tenth of a second; the Indian one? More than a second, which is easily noticeable to the end user. In real life it's worse than that…

Source: I teach this stuff (and yes, I simplified it a bit).
 
Last edited:
To add a point to the above, I currently spend my days trying to shave tenths of a millisecond off of vector search queries. The "ship it to india" solution is putting square wheels on the racecar, regardless of geopolitics.
 
As with clown computing, I'll be more interested in offshore AI when we have fully homomorphic encryption, ie never.

To add a point to the above, I currently spend my days trying to shave tenths of a millisecond off of vector search queries.
duckdb u akbar!
 
Back
Top