Summary
Chinese AI startup DeepSeek initially claimed to have trained its competitive R1 model with only $6 million and 2,048 GPUs.
However, a SemiAnalysis report reveals the company has actually invested $1.6 billion in hardware and owns 50,000 Nvidia GPUs.
DeepSeek has also spent well over $500 million on AI development since its inception.
DeepSeek operates its own data centers and exclusively hires from China, offering top salaries.
The report suggests its success stems from major investments rather than radical efficiency, countering initial claims of disruptive cost reductions.
DeepSeek was open about this from the very beginning. The $6 million figure was about comparing apples with apples. Just look at page 5 in the original paper. The SemiAnalysis report compares apples with oranges to save the ass of OpenAI.
I wish the US tech industry cared more about making cool tech than marketing
Stop using it because they are lying to you.
They may own 50,000 Nvidia GPUs, but if they only used 2,048 then it doesn’t really matter how many they own.
Also, the specific GPU is very important here. OpenAI is using top of the line GPUs, but due to trade restrictions Chinese companies can only get much less powerful chips.
The other 47,952 GPUs are mining bitcoin
More like: Mining the private data of Chinese citizens.
I think the fact that it’s open source and freely available for anyone to use is likely more disruptive than it being more efficient. OpenAI had a monopoly that just disappeared overnight.
But there are already several dozen (at least) open source models?
Oh you mean yet another revolutionary breakthrough from China turned out to be just smoke in mirrors? I am shocked! shocked I tell you!
And still, the reaction on the stock market showed how fragile and overblown yet not well understood the whole LLM technology is. Maybe at least some investors learned to tread more carefully and be more realistic about their expectations… but I’m not holding my breath.
smoke in mirrors
Bone apple tea
Bone apple tea
For all intensive purposes, yes.
A recent claim that DeepSeek trained its latest model for just $6 million has fueled much of the hype. However, this figure refers only to a portion of the total training cost— specifically, the GPU time required for pre-training.
How much have other companies spent just on pre-training? If that figure is just for pre-training, it would be useful to know what current industry leaders have spent, to make an apples to apples comparison.
It does not account for research, model refinement, data processing, or overall infrastructure expenses. In reality, DeepSeek has spent well over $500 million on AI development since its inception.
But later in the article they quote Elon Musk, saying “if you want to be competitive in AI, you have to spend billions per year,” but $500 million is significantly less than “billions.” And that’s since its inception, which was about 18 months ago. So, that’s less than half a billion dollars, per year. That’s much, much less than “billions per year.”
Also, the title says that DeepSeek spent “$1.6 billion,” but further on in the article they say “well over $500 million.” $1.6 billion is “well over $500 million,” but conventionally you wouldn’t phrase it like that if the amount was that much higher (over 3x) than $500 million. That leads me to believe the amount DeepSeek spent on AI development is much closer to $500 million than $1.6 billion. Apparently, that $1.6 billion figure includes costs not associated with AI development.
you mean ai peddlers are liars? unbelievable, i tell you
Keep in mind this is US AI peddlers trying to discredit Chinese AI peddlers. I would take both sides with a metric fuckton of salt.
In any case, even if it was true, the $6M would be misleading, as the distilled other models, so the total it was more like 100 billion and 6 million dollars total cost.
Still a cool result, but short of revolutionary.
You know that if you want to make cake from scratch you first need to invent the universe…
