About a quarter of South Koreans already pay for AI services, compared to just 2% of Americans
South Korea is preparing to offer free generative AI services to its entire population, using the technology to help people handle everyday tasks such as booking medical appointments, finding housing, and sorting out taxes.
The program is intended to expand access to homegrown AI systems while giving the government a larger role in how they are used for everyday services.
I’m no expert so take what I say with a grain of salt.
LLMs are prohibitively expensive to run.
If I wanted to make a basic rollerball pen, the cost would be high because I would have to gather the resources just to make a single unit. Mass manufacturing drops the price considerably to make it affordable for me to buy a pen.
But tokens aren’t the same as a factory line. If it’s $1 for me to use an LLM then it’s $100 for 100 people. It scales poorly.
LLM availability over cloud infrastructure requires resources that are expensive to build and depreciate quickly to the point that ROI is untenable (ie. Nvidia’s GPUs).
Now imagine a stock market that hinges around a handful of companies that are involved in that infrastructure component (eg. SK hynix). If many retail investors have dumped their savings into those stocks which can’t provide an expected return then the value of people’s savings will inevitably fluctuate on the whims of the AI industry.
You can imagine what will happen the next time a Chinese company releases a free and efficient LLM that can run on an 8GB RAM personal laptop. The panic selling will reverberate through the markets and a lot of people’s savings will evaporate overnight.
They still scale to some degree. There’s a certain amount of VRAM you need to even load a model (frontier models likely over a terabyte now), but once you’re there, adding extra users is less cost since you only need to accommodate their KV caches on top of the model of itself. You’ll need to add more GPUs once you’ve saturated what you have, but it’s nothing compared to the ~5 B300s for the first user (they actually come in servers containing 8 GPUs that cost about 500k EUR each, so realistically you’d buy one DGX B300 server to start with). Once you’ve loaded the model itself which has a fixed size, all the extra capacity you add is free to use for KV cache (which means more users).
It’s true, but the reason they depreciate so quickly is that there’s a new generation out every 2-3 years that does more using less energy. You could actually run old GPUs, it’s just cheaper to buy new ones if you’re a hyperscaler.
on the point of depreciation. would a hyperscaler be able to make a profitable return on a per GPU basis before they’ve maxed out its use?
genuine question not a rhetorical.
I have to admit that I do not understand most of what you said :(
But, I can see that (I think) you are saying that many people in South Korea invested in AI, and that is problematic. I do know that, as you said, LLMs are resource intensive and do not really work financially. If many people there invested in AI companies, I could see them also subscribing to them. I think.
My avoidance of AI has lead to me not knowing enough about it. I don’t know what you mean by token… This is a common theme for me today! I won’t be able to avoid this if I don’t educate myself.
LLMs ingest tokens as input. It is a core component of natural language processing. When entering a query into an LLM the text will be parsed as tokens.
I don’t know if the investing is inherently problematic. I do believe that the current situation is the culmination of many intersecting societal issues.
This helps a lot! Thanks! It fits in well with your ball point pen comparison. I see why it scales poorly.
As for your second part, I really have nothing to say. I do not know much about investing, but your earlier comment said a lot about the societal side of things. Thanks for helping through this!