Tricky, since as you say, the article has now vanished.
It may be of interest to photographers, that noise reduction software, [I have DxO Photolab], uses the graphics processor for the calculations, as it’s better geared to number crunching then the motherboard CPU.

I don’t know if this link addresses any of your queries, written by a tech journalist Karen Hao
It was originally a series of tweets on Twitter, but for those without an account, here it is helpfully rolled together via Threadreader.
After reading the article, continue to the foot of the page, where there are a few others by the same author.

https://threadreaderapp.com/thread/1883877986656825503.html

Wiki is interesting
https://en.wikipedia.org/wiki/DeepSeek

My first experience of reading about AI was this, a couple of years ago; quite remarkable fore the time

Bill

One view is that China/deepseek developers were acting under contraints (lack of billions of dollars of venture capital and restricted access to the latest GPUs). These constraints encouraged innovation to get more bang for your buck and more bang for your CPU power.

Deepseek is open source in the true meaning of the word. i think it the MIT license which is very permissive. You could run it at home if you got the computing power. You can modify it.

The paper on the development is supposed to be very detailed – I got the impression that they are not hiding anything.

It is the combination of efficiency and openess that scared the US stock market. The market was relying on massive barrier to entry and dependency on hundreds of thousands of the latest Nvidia cards. Deepseek was no doubt very expensive to create but some estimate it is two orders of magnitude cheaper than OpenAI (OpenAI is ironically no where near as ‘open’ as deepseek).

This needs checking as I heard something like it on a podcast in the middle of the night… it is estimated that the next gen AI training compute centres will consume 20% of current US electricity requirements. I probably got that wrong, but probably the right order of magnitude. The big bois (e.g. microsoft) are funding nuclear power plantd to run the datacentres.

If they train it on instruction manuals for Chinese made products………