At scale, this happens worldwide in shopping malls already and the loadshedding demand management in HVAC and chiller room is immensely valuable.
Sure. It's only 2%. Two percent of Musks net worth would fix homelessness in San Francisco applied annually. [edit: Actually, it's 2x too much]
2% of cost savings on inputs to a product could translate to a 50% increase in net profit. "oh well. its only 2% so why bother..."
I know this wasn't the point of your analogy. But the ill-conceived idea that money can fix homelessness in San Francisco is exactly the reason why we got here in the first place.
Richest city in the world needs more money to fix simple problems, says blameless local govt.I used the homelessness problem because it has a $cost at scale which people see as intractable. The wisdom of fixing it by throwing money at it, rather than throwing money at lawyers to overturn the stupid planning issues is a different question.
Actually, throwing $1.5b to the homeless would probably work pretty well.
Cost of a new affordable housing unit in SF ~$1.17M. [1] Number of homeless people $1.5B could house ~1,282. Number of homeless people in SF 7,754 people in 2022 [2].
So $1.5B could only solve ~16.5% of SF's housing problem or leave 84% unsolved. Which shows homelessness really is a "housing cost" problem in SF :)
[1] https://www.sfchronicle.com/sf/article/It-now-costs-more-tha... [2]https://sfgov.org/scorecards/safety-net/homeless-population
plus, solving 16% of the shortage of housing supply would be bigger than any other intervention at scale: it would probably force downward pricing in rent and sales overall.
https://sfstandard.com/public-health/homelessness/ending-hom...
...Ending unsheltered homelessness would require vast increases in housing, shelter and prevention services, according to the report from the city’s Department of Homelessness and Supportive Housing and homelessness consultancy firm Focus Strategies.
That would cost the city another $1.45 billion over three years. And it would take over $410 million each year after that to keep the programs running. ...
I don't remember the exact number but raising the temperature of an olympic sized swimming pool by a degree celsius takes megawatt hours of energy. that's a lot and a datacenter is perfect for this, even in hot climates.
Just a note: we are probably right now at peak inefficiency for LLMs. There are obvious improvements available but they will only be explored by organizations which have different time vs resources trade-off calculations compared to OpenAI and Google.
Right now OpenAi and Google are racing to ship products. A working training framework is more important than saving $X (where X is a very large number) because capturing marketshare is seen as vital and they have no shortage of capital.
But the massive and rapid progress seen by projects like llama-cpp show how much performance they are leaving on the table. (And yes this also applies at training time).
others are going to go onprem
I think this might already be near peak data center LLM, the use case for the most genius cloud heavy LLMs is going to be small once it gets baked into OS layer
However, with LLMs coming, I could see data center energy usage jump 10x over the next decade. At that point, re-using data center heat will be be a real industry with real profits. Imagine data center heated apartment heating, vertical farms.
I don't know how hotter regions are going to cope with the heat though. These AI datacenters could end up being entirely concentrated in the coldest regions.