If not friend, why friend shape?

Also no.

Have a great holiday

Why should we *not* build new data centers?

The idea that these data centers are metabolizing the planet is hyperbolic. Agree or disagree?

Are you intentionally misconstruing my point to create a strawman?

> "What is happening all around us is the creation of a technical infrastructure that is directly competing with humans for basic resources like water, energy, and land-pushing the limits of a planet already under strain."

I disagree with this quote re: energy and land. Water maybe.

The original supposition is that AI workloads are metabolizing the planet. I disagree that AI workloads are a significant driving factor of energy consumption globally. No one is claiming GPU hungry workloads are less energy intensive. The discussion is what portion of consumption does it represent?

When faced with evidence and hypothesis, if your general activity is to go ad hominem and attack the integrity of the person you disagree with, then you kill any form of discussion.

I'm open to being wrong, are you also open to being wrong? Do you hold your opinions or do they hold you?

I will be direct.

When the evidence doesn't support your priors, maybe you should reexamine your assumptions?

Across hyperscalers and smaller operators there is no one number that is correct. The error bars are massive. Anyone claiming to have a perfect idea of the numbers is likely incorrect.

Yes. I worked:
1) at AWS
2) on the Pleiades supercomputer
3) on infra for PyTorch and model training at Meta across on prem and AWS
4) on prem DCs for SpaceX
5) currently for several customers who run large ML workloads across AWS and on prem GPU farms

I am an expert in this field and I am explaining to you that it is difficult to quantify due to a lack of public data.

One can hypothesize though and the fact is that power hungry GPU based workloads are not as widely deployed as overall data center capacity. At least not yet.

You're right I shouldn't have used an overloaded (pun intended) term there.

Workload here refers to load in the electrical sense

The term workload here is not workload in a devops sense but workload in an electrical sense

I'm not saying it's zero or meaningless to think about these things... but it's not metabolizing the world. That's a statement that sounds good but has little to no backing in actual numbers. Would 2% of global power usage be the first target for efficiency gains?

Here's my hypothesis:

Total large GPU deployments in data centers
-----
Total # of data centers

My bet is the denominator is *much* bigger than the numerator. There are fixed power costs for each of the DCs in the denominator that don't go down.

I don't? What evidence would you use to support that conclusion?

1. Building data centers takes time.
2. Not all data centers have massive GPU deployments, most around the world basically function as CDNs.