Hi! 👋 Welcome to The Big Y!
Every time you prompt an LLM, a data server is handling your request drawing on power and local resources to get you the answer you need. Going beyond what your ChatGPT subscription costs you, or what each of your requests costs the LLM provider in cloud consumption, we can look at what the request costs in terms of power (electricity) and surprisingly also water.
Requests for GPUs are sky high and data center providers are doing everything to get more data centers built, expanding into rural America where they can get cheaper land and easier access to power. This article in The Atlantic, dives into the impact of these data centers on energy consumption and specifically the use of water to cool these data centers.
Arizona is a data center hub hosting all the various cloud providers with each continuing to build additional capacity. There are some complicated economic incentives attached to data centers, as rural cities try to attract them with rebates and preferred treatments. Arizona is entering its 30th year of drought, so you may ask whether it makes sense to build data centers that rely on water for cooling in a very warm place that already is facing water shortages for its humans?
In many places in the US, the power grid is already struggling to keep up with demand and is running into further issues as extreme weather places additional strain on the aging systems. When rolling blackouts hit during a heatwave, does the data center or do the local residents get their electricity needs met first?
While there are rumors of a 1KW GPU from Nvidia we can only hope for more energy efficient chips and more efficient models. However, in the current landscape, the trend is likely to persist with a surge in GPU-centric data center expansion as the cloud providers capitalize on the AI boom. We don’t need to stop using GenAI or developing it, but I do think it is important to realize the bigger implications and discuss a more sustainable way to fuel the benefits of GenAI.
Researchers at Google DeepMind released a paper in which they describe the development of a foundational model that animates a still image or drawing into an interactive environment. This could be impactful in generating simulated environments for robots or AI agents before transferring them into the real world.
Know someone who might enjoy this newsletter? Share it with them and help spread the word!
Thanks for reading! Have a great week! 😁