Firstly, Happy Mother's Day weekend! I hope you're all spending time with loved ones.
In the past few weeks, I've gotten asked quite a bit about my stance on the construction of additional data centers all over the world. They're controversial for many reasons, and as a deeply technical person myself (most likely the reason I was asked), I figure it'd be good to explicitly talk about my thoughts on them.
The stakes themselves are close to home and not unfamiliar to me - I'm in the Great Lakes region, and there, we have one of the world's largest freshwater supplies (~20% of the world's fresh water, and ~85% of North America's fresh water supply).
Long story short, I'm not entirely disconnected from the problem. As I write this, I'm looking at the beautiful Lake Michigan in all its vastness.
Data centers cost a huge amount of power (electricity), resources (semiconductors, chips, labor) to construct. When they operate, they create demand on the electricity grid of the locale in which they're placed.
They also have varying cooling systems to keep everything performant. These cooling systems can either be closed-loop (no additional water needed), or open-loop (i.e. they constantly cycle new water in - power plants and other industrial processes can do this).
Naturally, two things happen: more electricity production from the locale is needed, and if the cooling system is open-loop, it needs to dump the heated water out. Not all cooling systems are open-loop; some are closed loop and ignore this particular cost to the environment.
This reveals cost-wise what the true strains are on a system: dirty electricity, and imperfect cooling mechanisms. Both have been solved in certain respects (e.g. nuclear plants, closed loop cooling systems), but not every locale has adopted this. As a result, we blame the downstream cause ("AI!") as opposed to the upstream circumstance (non-modernized energy grids and cooling designs, neither of which are specific to data centers).
I was talking with someone a few days ago who verbosely expressed how bad they felt that every time they talked with ChatGPT, a few cups of fresh water would evaporate. However, that very same person was a bit astonished to learn that a local AI can run on their phone with no data center needed - when we charge phones, we don't think about that analogy or that cost. In a more severe example, we've accepted the presence of cars in our everyday lives at the cost of their downsides (killing people every year). However, even that example is being solved with self-driving cars safer than real humans operating them.
What I'm getting at here: data centers are a convenient messenger to blame. Their necessity in people's eyes is directly correlated to people's views of AI and its role in society, and this amplifies the effect even further. What we need is transparency in how our systems are built and operated - from energy, to water infrastructure, to all downstream consumers of those services.
These are not new concepts: data centers, power plants, and water/cooling infrastructure existed before AI. They will exist in the future, and we should be finding ways to adapt the upstream providers of our society to our direct needs, versus blaming the need itself.
My take here? The benefit of the compute is here to stay - and we're just getting started. Don't blame the messenger, and go as far upstream as you can to find the cause of things.
Be well,
Michael Kirsanov
P.S. Part 1 can be found here.